Upload 29 files

.gitignore

@@ -0,0 +1,162 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/latest/usage/project/#working-with-version-control
+.pdm.toml
+.pdm-python
+.pdm-build/
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/

requirements.txt

@@ -0,0 +1,12 @@
+gradio==4.29.0
+tensorboardX
+einops
+local-attention
+pedalboard==0.7.7
+
+fairseq==0.12.2
+faiss-cpu==1.7.3
+ffmpeg-python>=0.2.0
+praat-parselmouth>=0.4.2
+pyworld==0.3.4
+torchcrepe==0.0.20

rvc_models/MODELS.txt

@@ -0,0 +1,2 @@
+RVC Models can be added as a folder here. Each folder should contain the model file (.pth extension), and an index file (.index extension).
+For example, a folder called Maya, containing 2 files, Maya.pth and added_IVF1905_Flat_nprobe_Maya_v2.index.

song_output/OUTPUT.txt

@@ -0,0 +1 @@
+Output is stored in this folder, where directory names represent the YouTube IDs from the original song.

src/CoverGenLite.py

@@ -0,0 +1,195 @@
+import os
+import shutil
+import urllib.request
+import zipfile
+import gdown
+import gradio as gr
+
+from main import song_cover_pipeline
+from audio_effects import add_audio_effects
+from modules.model_management import ignore_files, update_models_list, extract_zip, download_from_url, upload_zip_model
+from modules.ui_updates import show_hop_slider, update_f0_method, update_button_text, update_button_text_voc, update_button_text_inst
+from modules.file_processing import process_file_upload
+
+BASE_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+rvc_models_dir = os.path.join(BASE_DIR, 'rvc_models')
+output_dir = os.path.join(BASE_DIR, 'song_output')
+
+
+if __name__ == '__main__':
+    voice_models = ignore_files(rvc_models_dir)
+
+    with gr.Blocks(title='CoverGen Lite - Politrees (v0.2)', theme=gr.themes.Soft(primary_hue="green", secondary_hue="green", neutral_hue="neutral", spacing_size="sm", radius_size="lg")) as app:
+        with gr.Tab("Велком/Контакты"):
+            gr.HTML("<center><h1>Добро пожаловать в CoverGen Lite - Politrees (v0.2)</h1></center>")
+            with gr.Row():
+                with gr.Column(variant='panel'):
+                    gr.HTML("<center><h2><a href='https://www.youtube.com/channel/UCHb3fZEVxUisnqLqCrEM8ZA'>YouTube: Politrees</a></h2></center>")
+                    gr.HTML("<center><h2><a href='https://vk.com/artem__bebroy'>ВКонтакте (страница)</a></h2></center>")
+                with gr.Column(variant='panel'):
+                    gr.HTML("<center><h2><a href='https://t.me/pol1trees'>Telegram Канал</a></h2></center>")
+                    gr.HTML("<center><h2><a href='https://t.me/+GMTP7hZqY0E4OGRi'>Telegram Чат</a></h2></center>")
+            with gr.Column(variant='panel'):
+                gr.HTML("<center><h2><a href='https://github.com/Bebra777228/Pol-Litres-RVC'>GitHub проекта</a></h2></center>")
+
+        with gr.Tab("Преобразование голоса"):
+            with gr.Row(equal_height=False):
+                with gr.Column(scale=1, variant='panel'):
+                    with gr.Group():
+                        rvc_model = gr.Dropdown(voice_models, label='Модели голоса')
+                        ref_btn = gr.Button('Обновить список моделей', variant='primary')
+                    with gr.Group():
+                        pitch = gr.Slider(-24, 24, value=0, step=0.5, label='Изменение тона голоса', info='-24 - мужской голос || 24 - женский голос')
+
+                with gr.Column(scale=2, variant='panel'):
+                    with gr.Group():
+                        local_file = gr.Audio(label='Аудио-файл', interactive=False, show_download_button=False)
+                        uploaded_file = gr.UploadButton(label='Загрузить аудио-файл', file_types=['audio'], variant='primary')
+                        uploaded_file.upload(process_file_upload, inputs=[uploaded_file], outputs=[local_file])
+                        uploaded_file.upload(update_button_text, outputs=[uploaded_file])
+
+            with gr.Group():
+                with gr.Row(variant='panel'):
+                    generate_btn = gr.Button("Генерировать", variant='primary', scale=1)
+                    converted_voice = gr.Audio(label='Преобразованный голос', scale=5)
+                    output_format = gr.Dropdown(['mp3', 'flac', 'wav'], value='mp3', label='Формат файла', scale=0.1, allow_custom_value=False, filterable=False)
+
+            with gr.Accordion('Настройки преобразования голоса', open=False):
+                with gr.Group():
+                    with gr.Column(variant='panel'):
+                        use_hybrid_methods = gr.Checkbox(label="Использовать гибридные методы", value=False)
+                        f0_method = gr.Dropdown(['rmvpe+', 'fcpe', 'rmvpe', 'mangio-crepe', 'crepe'], value='rmvpe+', label='Метод выделения тона', allow_custom_value=False, filterable=False)
+                        use_hybrid_methods.change(update_f0_method, inputs=use_hybrid_methods, outputs=f0_method)
+                        crepe_hop_length = gr.Slider(8, 512, value=128, step=8, visible=False, label='Длина шага Crepe')
+                        f0_method.change(show_hop_slider, inputs=f0_method, outputs=crepe_hop_length)
+                    with gr.Column(variant='panel'):
+                        index_rate = gr.Slider(0, 1, value=0, label='Влияние индекса', info='Контролирует степень влияния индексного файла на результат анализа. Более высокое значение увеличивает влияние индексного файла, но может усилить артефакты в аудио. Выбор более низкого значения может помочь сниз��ть артефакты.')
+                        filter_radius = gr.Slider(0, 7, value=3, step=1, label='Радиус фильтра', info='Управляет радиусом фильтрации результатов анализа тона. Если значение фильтрации равняется или превышает три, применяется медианная фильтрация для уменьшения шума дыхания в аудиозаписи.')
+                        rms_mix_rate = gr.Slider(0, 1, value=0.25, step=0.01, label='Скорость смешивания RMS', info='Контролирует степень смешивания выходного сигнала с его оболочкой громкости. Значение близкое к 1 увеличивает использование оболочки громкости выходного сигнала, что может улучшить качество звука.')
+                        protect = gr.Slider(0, 0.5, value=0.33, step=0.01, label='Защита согласных', info='Контролирует степень защиты отдельных согласных и звуков дыхания от электроакустических разрывов и других артефактов. Максимальное значение 0,5 обеспечивает наибольшую защиту, но может увеличить эффект индексирования, который может негативно влиять на качество звука. Уменьшение значения может уменьшить степень защиты, но снизить эффект индексирования.')
+
+            ref_btn.click(update_models_list, None, outputs=rvc_model)
+            generate_btn.click(song_cover_pipeline,
+                              inputs=[uploaded_file, rvc_model, pitch, index_rate, filter_radius, rms_mix_rate, f0_method, crepe_hop_length, protect, output_format],
+                              outputs=[converted_voice])
+
+        with gr.Tab('Объединение/Обработка'):
+            with gr.Row(equal_height=False):
+                with gr.Column(variant='panel'):
+                    with gr.Group():
+                        vocal_audio = gr.Audio(label='Вокал', interactive=False, show_download_button=False)
+                        upload_vocal_audio = gr.UploadButton(label='Загрузить вокал', file_types=['audio'], variant='primary')
+                        upload_vocal_audio.upload(process_file_upload, inputs=[upload_vocal_audio], outputs=[vocal_audio])
+                        upload_vocal_audio.upload(update_button_text_voc, outputs=[upload_vocal_audio])
+        
+                with gr.Column(variant='panel'):
+                    with gr.Group():
+                        instrumental_audio = gr.Audio(label='Инструментал', interactive=False, show_download_button=False)
+                        upload_instrumental_audio = gr.UploadButton(label='Загрузить инструментал', file_types=['audio'], variant='primary')
+                        upload_instrumental_audio.upload(process_file_upload, inputs=[upload_instrumental_audio], outputs=[instrumental_audio])
+                        upload_instrumental_audio.upload(update_button_text_inst, outputs=[upload_instrumental_audio])
+        
+            with gr.Group():
+                with gr.Row(variant='panel'):
+                    process_btn = gr.Button("Обработать", variant='primary', scale=1)
+                    ai_cover = gr.Audio(label='Ai-Cover', scale=5)
+                    output_format = gr.Dropdown(['mp3', 'flac', 'wav'], value='mp3', label='Формат файла', scale=0.1, allow_custom_value=False, filterable=False)
+
+            with gr.Accordion('Настройки сведения аудио', open=False):
+                gr.HTML('<center><h2>Изменение громкости</h2></center>')
+                with gr.Row(variant='panel'):
+                    vocal_gain = gr.Slider(-10, 10, value=0, step=1, label='Вокал', scale=1)
+                    instrumental_gain = gr.Slider(-10, 10, value=0, step=1, label='Инструментал', scale=1)
+                    clear_btn = gr.Button("Сбросить все эффекты", scale=0.1)
+
+                with gr.Accordion('Эффекты', open=False):
+                    with gr.Accordion('Реверберация', open=False):
+                        with gr.Group():
+                            with gr.Column(variant='panel'):
+                                with gr.Row():
+                                    reverb_rm_size = gr.Slider(0, 1, value=0.15, label='Размер комнаты', info='Этот параметр отвечает за размер виртуального помещения, в котором будет звучать реверберация. Большее значение означает больш��й размер комнаты и более длительное звучание реверберации.')
+                                    reverb_width = gr.Slider(0, 1, value=1.0, label='Ширина реверберации', info='Этот параметр отвечает за ширину звучания реверберации. Чем выше значение, тем шире будет звучание реверберации.')
+                                with gr.Row():
+                                    reverb_wet = gr.Slider(0, 1, value=0.1, label='Уровень влажности', info='Этот параметр отвечает за уровень реверберации. Чем выше значение, тем сильнее будет слышен эффект реверберации и тем дольше будет звучать «хвост».')
+                                    reverb_dry = gr.Slider(0, 1, value=0.8, label='Уровень сухости', info='Этот параметр отвечает за уровень исходного звука без реверберации. Чем меньше значение, тем тише звук ai вокала. Если значение будет на 0, то исходный звук полностью исчезнет.')
+                                with gr.Row():
+                                    reverb_damping = gr.Slider(0, 1, value=0.7, label='Уровень демпфирования', info='Этот параметр отвечает за поглощение высоких частот в реверберации. Чем выше его значение, тем сильнее будет поглощение частот и тем менее будет «яркий» звук реверберации.')
+    
+                    with gr.Accordion('Хорус', open=False):
+                        with gr.Group():
+                            with gr.Column(variant='panel'):
+                                with gr.Row():
+                                    chorus_rate_hz = gr.Slider(0.1, 10, value=0, label='Скорость хоруса', info='Этот параметр отвечает за скорость колебаний эффекта хоруса в герцах. Чем выше значение, тем быстрее будут колебаться звуки.')
+                                    chorus_depth = gr.Slider(0, 1, value=0, label='Глубина хоруса', info='Этот параметр отвечает за глубину эффекта хоруса. Чем выше значение, тем сильнее будет эффект хоруса.')
+                                with gr.Row():
+                                    chorus_centre_delay_ms = gr.Slider(0, 50, value=0, label='Задержка центра (мс)', info='Этот параметр отвечает за задержку центрального сигнала эффекта хоруса в миллисекундах. Чем выше значение, тем дольше будет задержка.')
+                                    chorus_feedback = gr.Slider(0, 1, value=0, label='Обратная связь', info='Этот параметр отвечает за уровень обратной связи эффекта хоруса. Чем выше значение, тем сильнее будет эффект обратной связи.')
+                                with gr.Row():
+                                    chorus_mix = gr.Slider(0, 1, value=0, label='Смешение', info='Этот параметр отвечает за уровень смешивания оригинального сигнала и эффекта хоруса. Чем выше значение, тем сильнее будет эффект хоруса.')
+
+                with gr.Accordion('Обработка', open=False):
+                    with gr.Accordion('Компрессор', open=False):
+                        with gr.Row(variant='panel'):
+                            compressor_ratio = gr.Slider(1, 20, value=4, label='Соотношение', info='Этот параметр контролирует количество применяемого сжатия аудио. Большее значение означает большее сжатие, которое уменьшает динамический диапазон аудио, делая громкие части более тихими и тихие части более громкими.')
+                            compressor_threshold = gr.Slider(-60, 0, value=-16, label='Порог', info='Этот параметр устанавливает порог, при превышении которого начинает действовать компрессор. Компрессор сжимает громкие звуки, чтобы сделать звук более ровным. Чем ниже порог, тем большее коли��ество звуков будет подвергнуто компрессии.')
+    
+                    with gr.Accordion('Фильтры', open=False):
+                        with gr.Row(variant='panel'):
+                            low_shelf_gain = gr.Slider(-20, 20, value=0, label='Фильтр нижних частот', info='Этот параметр контролирует усиление (громкость) низких частот. Положительное значение усиливает низкие частоты, делая звук более басским. Отрицательное значение ослабляет низкие частоты, делая звук более тонким.')
+                            high_shelf_gain = gr.Slider(-20, 20, value=0, label='Фильтр высоких частот', info='Этот параметр контролирует усиление высоких частот. Положительное значение усиливает высокие частоты, делая звук более ярким. Отрицательное значение ослабляет высокие частоты, делая звук более тусклым.')
+    
+                    with gr.Accordion('Подавление шума', open=False):
+                        with gr.Group():
+                            with gr.Column(variant='panel'):
+                                with gr.Row():
+                                    noise_gate_threshold = gr.Slider(-60, 0, value=-30, label='Порог', info='Этот параметр устанавливает пороговое значение в децибелах, ниже которого сигнал считается шумом. Когда сигнал опускается ниже этого порога, шумовой шлюз активируется и уменьшает громкость сигнала.')
+                                    noise_gate_ratio = gr.Slider(1, 20, value=6, label='Соотношение', info='Этот параметр устанавливает уровень подавления шума. Большее значение означает более сильное подавление шума.')
+                                with gr.Row():
+                                    noise_gate_attack = gr.Slider(0, 100, value=10, label='Время атаки (мс)', info='Этот параметр контролирует скорость, с которой шумовой шлюз открывается, когда звук становится достаточно громким. Большее значение означает, что шлюз открывается медленнее.')
+                                    noise_gate_release = gr.Slider(0, 1000, value=100, label='Время спада (мс)', info='Этот параметр контролирует скорость, с которой шумовой шлюз закрывается, когда звук становится достаточно тихим. Большее значение означает, что шлюз закрывается медленнее.')
+
+            process_btn.click(add_audio_effects,
+                             inputs=[upload_vocal_audio, upload_instrumental_audio, reverb_rm_size, reverb_wet, reverb_dry, reverb_damping,
+                             reverb_width, low_shelf_gain, high_shelf_gain, compressor_ratio, compressor_threshold,
+                             noise_gate_threshold, noise_gate_ratio, noise_gate_attack, noise_gate_release,
+                             chorus_rate_hz, chorus_depth, chorus_centre_delay_ms, chorus_feedback, chorus_mix,
+                             output_format, vocal_gain, instrumental_gain],
+                             outputs=[ai_cover])
+
+            default_values = [0, 0, 0.15, 1.0, 0.1, 0.8, 0.7, 0, 0, 0, 0, 0, 4, -16, 0, 0, -30, 6, 10, 100]
+            clear_btn.click(lambda: default_values,
+                            outputs=[vocal_gain, instrumental_gain, reverb_rm_size, reverb_width, reverb_wet, reverb_dry, reverb_damping,
+                            chorus_rate_hz, chorus_depth, chorus_centre_delay_ms, chorus_feedback, chorus_mix,
+                            compressor_ratio, compressor_threshold, low_shelf_gain, high_shelf_gain, noise_gate_threshold,
+                            noise_gate_ratio, noise_gate_attack, noise_gate_release])
+
+        with gr.Tab('Загрузка модели'):
+            with gr.Tab('Загрузить по ссылке'):
+                with gr.Row():
+                    with gr.Column(variant='panel'):
+                        gr.HTML("<center><h3>Вставьте в поле ниже ссылку от <a href='https://huggingface.co/' target='_blank'>HuggingFace</a>, <a href='https://pixeldrain.com/' target='_blank'>Pixeldrain</a> или <a href='https://drive.google.com/' target='_blank'>Google Drive</a></h3></center>")
+                        model_zip_link = gr.Text(label='Ссылка на загрузку модели')
+                    with gr.Column(variant='panel'):
+                        with gr.Group():
+                            model_name = gr.Text(label='Имя модели', info='Дайте вашей загружаемой модели уникальное имя, отличное от других голосовых моделей.')
+                            download_btn = gr.Button('Загрузить модель', variant='primary')
+
+                dl_output_message = gr.Text(label='Сообщение вывода', interactive=False)
+                download_btn.click(download_from_url, inputs=[model_zip_link, model_name], outputs=dl_output_message)
+
+            with gr.Tab('Загрузить локально'):
+                with gr.Row():
+                    with gr.Column(variant='panel'):
+                        zip_file = gr.File(label='Zip-файл', file_types=['.zip'], file_count='single')
+                    with gr.Column(variant='panel'):
+                        gr.HTML("<h3>1. Найдите и скачайте файлы: .pth и необязательный файл .index</h3>")
+                        gr.HTML("<h3>2. Закиньте файл(-ы) в ZIP-архив и поместите его в область загрузки</h3>")
+                        gr.HTML('<h3>3. Дождитесь полной загрузки ZIP-архива в интерфейс</h3>')
+                        with gr.Group():
+                            local_model_name = gr.Text(label='Имя модели', info='Дайте вашей загружаемой модели уникальное имя, отличное от других голосовых моделей.')
+                            model_upload_button = gr.Button('Загрузить модель', variant='primary')
+
+                local_upload_output_message = gr.Text(label='Сообщение вывода', interactive=False)
+                model_upload_button.click(upload_zip_model, inputs=[zip_file, local_model_name], outputs=local_upload_output_message)
+
+    app.launch(share=True, quiet=True)

src/audio_effects.py

@@ -0,0 +1,74 @@
+import os
+import librosa
+import numpy as np
+import gradio as gr
+import soundfile as sf
+from pedalboard import (
+    Pedalboard, Reverb, Compressor, HighpassFilter,
+    LowShelfFilter, HighShelfFilter, NoiseGate, Chorus
+)
+from pedalboard.io import AudioFile
+from pydub import AudioSegment
+
+BASE_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+
+def display_progress(percent, message, progress=gr.Progress()):
+    progress(percent, desc=message)
+
+def combine_audio(vocal_path, instrumental_path, output_path, vocal_gain, instrumental_gain, output_format):
+    vocal_format = vocal_path.split('.')[-1]
+    instrumental_format = instrumental_path.split('.')[-1]
+
+    vocal = AudioSegment.from_file(vocal_path, format=vocal_format)
+    instrumental = AudioSegment.from_file(instrumental_path, format=instrumental_format)
+
+    vocal += vocal_gain
+    instrumental += instrumental_gain
+
+    combined = vocal.overlay(instrumental)
+    combined.export(output_path, format=output_format)
+
+def add_audio_effects(vocal_audio_path, instrumental_audio_path, reverb_rm_size, reverb_wet, reverb_dry, reverb_damping, reverb_width,
+                      low_shelf_gain, high_shelf_gain, compressor_ratio, compressor_threshold, noise_gate_threshold, noise_gate_ratio,
+                      noise_gate_attack, noise_gate_release, chorus_rate_hz, chorus_depth, chorus_centre_delay_ms, chorus_feedback,
+                      chorus_mix, output_format, vocal_gain, instrumental_gain, progress=gr.Progress()):
+
+    if not vocal_audio_path or not instrumental_audio_path:
+        raise ValueError("Оба пути к аудиофайлам должны быть заполнены.")
+
+    display_progress(0.2, "Применение аудиоэффектов к вокалу...", progress)
+    board = Pedalboard(
+        [
+            HighpassFilter(),
+            Compressor(ratio=compressor_ratio, threshold_db=compressor_threshold),
+            NoiseGate(threshold_db=noise_gate_threshold, ratio=noise_gate_ratio, attack_ms=noise_gate_attack, release_ms=noise_gate_release),
+            Reverb(room_size=reverb_rm_size, dry_level=reverb_dry, wet_level=reverb_wet, damping=reverb_damping, width=reverb_width),
+            LowShelfFilter(gain_db=low_shelf_gain),
+            HighShelfFilter(gain_db=high_shelf_gain),
+            Chorus(rate_hz=chorus_rate_hz, depth=chorus_depth, centre_delay_ms=chorus_centre_delay_ms, feedback=chorus_feedback, mix=chorus_mix),
+         ]
+    )
+
+    vocal_output_path = f'Vocal_Effects.wav'
+    with AudioFile(vocal_audio_path) as f:
+        with AudioFile(vocal_output_path, 'w', f.samplerate, 2) as o:
+            while f.tell() < f.frames:
+                chunk = f.read(int(f.samplerate))
+                chunk = np.tile(chunk, (2, 1)).T
+                effected = board(chunk, f.samplerate, reset=False)
+                o.write(effected)
+
+    display_progress(0.5, "Объединение вокала и инструментальной части...", progress)
+    output_dir = os.path.join(BASE_DIR, 'processed_output')
+    if not os.path.exists(output_dir):
+        os.makedirs(output_dir)
+    combined_output_path = os.path.join(output_dir, f'AiCover_combined.{output_format}')
+
+    if os.path.exists(combined_output_path):
+        os.remove(combined_output_path)
+
+    combine_audio(vocal_output_path, instrumental_audio_path, combined_output_path, vocal_gain, instrumental_gain, output_format)
+
+    display_progress(1.0, "Готово!", progress)
+
+    return combined_output_path

src/configs/32k.json

@@ -0,0 +1,46 @@
+{
+  "train": {
+    "log_interval": 200,
+    "seed": 1234,
+    "epochs": 20000,
+    "learning_rate": 1e-4,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 4,
+    "fp16_run": false,
+    "lr_decay": 0.999875,
+    "segment_size": 12800,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0
+  },
+  "data": {
+    "max_wav_value": 32768.0,
+    "sampling_rate": 32000,
+    "filter_length": 1024,
+    "hop_length": 320,
+    "win_length": 1024,
+    "n_mel_channels": 80,
+    "mel_fmin": 0.0,
+    "mel_fmax": null
+  },
+  "model": {
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+    "upsample_rates": [10,4,2,2,2],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [16,16,4,4,4],
+    "use_spectral_norm": false,
+    "gin_channels": 256,
+    "spk_embed_dim": 109
+  }
+}

src/configs/32k_v2.json

@@ -0,0 +1,46 @@
+{
+  "train": {
+    "log_interval": 200,
+    "seed": 1234,
+    "epochs": 20000,
+    "learning_rate": 1e-4,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 4,
+    "fp16_run": true,
+    "lr_decay": 0.999875,
+    "segment_size": 12800,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0
+  },
+  "data": {
+    "max_wav_value": 32768.0,
+    "sampling_rate": 32000,
+    "filter_length": 1024,
+    "hop_length": 320,
+    "win_length": 1024,
+    "n_mel_channels": 80,
+    "mel_fmin": 0.0,
+    "mel_fmax": null
+  },
+  "model": {
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+    "upsample_rates": [10,8,2,2],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [20,16,4,4],
+    "use_spectral_norm": false,
+    "gin_channels": 256,
+    "spk_embed_dim": 109
+  }
+}

src/configs/40k.json

@@ -0,0 +1,46 @@
+{
+  "train": {
+    "log_interval": 200,
+    "seed": 1234,
+    "epochs": 20000,
+    "learning_rate": 1e-4,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 4,
+    "fp16_run": false,
+    "lr_decay": 0.999875,
+    "segment_size": 12800,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0
+  },
+  "data": {
+    "max_wav_value": 32768.0,
+    "sampling_rate": 40000,
+    "filter_length": 2048,
+    "hop_length": 400,
+    "win_length": 2048,
+    "n_mel_channels": 125,
+    "mel_fmin": 0.0,
+    "mel_fmax": null
+  },
+  "model": {
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+    "upsample_rates": [10,10,2,2],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [16,16,4,4],
+    "use_spectral_norm": false,
+    "gin_channels": 256,
+    "spk_embed_dim": 109
+  }
+}

src/configs/48k.json

@@ -0,0 +1,46 @@
+{
+  "train": {
+    "log_interval": 200,
+    "seed": 1234,
+    "epochs": 20000,
+    "learning_rate": 1e-4,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 4,
+    "fp16_run": false,
+    "lr_decay": 0.999875,
+    "segment_size": 11520,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0
+  },
+  "data": {
+    "max_wav_value": 32768.0,
+    "sampling_rate": 48000,
+    "filter_length": 2048,
+    "hop_length": 480,
+    "win_length": 2048,
+    "n_mel_channels": 128,
+    "mel_fmin": 0.0,
+    "mel_fmax": null
+  },
+  "model": {
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+    "upsample_rates": [10,6,2,2,2],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [16,16,4,4,4],
+    "use_spectral_norm": false,
+    "gin_channels": 256,
+    "spk_embed_dim": 109
+  }
+}

src/configs/48k_v2.json

@@ -0,0 +1,46 @@
+{
+  "train": {
+    "log_interval": 200,
+    "seed": 1234,
+    "epochs": 20000,
+    "learning_rate": 1e-4,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 4,
+    "fp16_run": true,
+    "lr_decay": 0.999875,
+    "segment_size": 17280,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0
+  },
+  "data": {
+    "max_wav_value": 32768.0,
+    "sampling_rate": 48000,
+    "filter_length": 2048,
+    "hop_length": 480,
+    "win_length": 2048,
+    "n_mel_channels": 128,
+    "mel_fmin": 0.0,
+    "mel_fmax": null
+  },
+  "model": {
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+    "upsample_rates": [12,10,2,2],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [24,20,4,4],
+    "use_spectral_norm": false,
+    "gin_channels": 256,
+    "spk_embed_dim": 109
+  }
+}

src/download_models.py

@@ -0,0 +1,29 @@
+from pathlib import Path
+import requests
+
+RVC_other_DOWNLOAD_LINK = 'https://huggingface.co/Politrees/all_RVC-pretrained_and_other/resolve/main/other/'
+RVC_hubert_DOWNLOAD_LINK = 'https://huggingface.co/Politrees/all_RVC-pretrained_and_other/resolve/main/HuBERTs/'
+
+BASE_DIR = Path(__file__).resolve().parent.parent
+rvc_models_dir = BASE_DIR / 'rvc_models'
+
+
+def dl_model(link, model_name, dir_name):
+    with requests.get(f'{link}{model_name}') as r:
+        r.raise_for_status()
+        with open(dir_name / model_name, 'wb') as f:
+            for chunk in r.iter_content(chunk_size=8192):
+                f.write(chunk)
+
+if __name__ == '__main__':
+    rvc_other_names = ['rmvpe.pt', 'fcpe.pt']
+    for model in rvc_other_names:
+        print(f'Downloading {model}...')
+        dl_model(RVC_other_DOWNLOAD_LINK, model, rvc_models_dir)
+
+    rvc_hubert_names = ['hubert_base.pt']
+    for model in rvc_hubert_names:
+        print(f'Downloading {model}...')
+        dl_model(RVC_hubert_DOWNLOAD_LINK, model, rvc_models_dir)
+
+    print('All models downloaded!')

src/infer_pack/attentions.py

@@ -0,0 +1,417 @@
+import copy
+import math
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from infer_pack import commons
+from infer_pack import modules
+from infer_pack.modules import LayerNorm
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        window_size=10,
+        **kwargs
+    ):
+        super().__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+
+        self.drop = nn.Dropout(p_dropout)
+        self.attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+        for i in range(self.n_layers):
+            self.attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    p_dropout=p_dropout,
+                    window_size=window_size,
+                )
+            )
+            self.norm_layers_1.append(LayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FFN(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                )
+            )
+            self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask):
+        attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+        for i in range(self.n_layers):
+            y = self.attn_layers[i](x, x, attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_1[i](x + y)
+
+            y = self.ffn_layers[i](x, x_mask)
+            y = self.drop(y)
+            x = self.norm_layers_2[i](x + y)
+        x = x * x_mask
+        return x
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        proximal_bias=False,
+        proximal_init=True,
+        **kwargs
+    ):
+        super().__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+
+        self.drop = nn.Dropout(p_dropout)
+        self.self_attn_layers = nn.ModuleList()
+        self.norm_layers_0 = nn.ModuleList()
+        self.encdec_attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+        for i in range(self.n_layers):
+            self.self_attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    p_dropout=p_dropout,
+                    proximal_bias=proximal_bias,
+                    proximal_init=proximal_init,
+                )
+            )
+            self.norm_layers_0.append(LayerNorm(hidden_channels))
+            self.encdec_attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout
+                )
+            )
+            self.norm_layers_1.append(LayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FFN(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                    causal=True,
+                )
+            )
+            self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask, h, h_mask):
+        """
+        x: decoder input
+        h: encoder output
+        """
+        self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(
+            device=x.device, dtype=x.dtype
+        )
+        encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+        for i in range(self.n_layers):
+            y = self.self_attn_layers[i](x, x, self_attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_0[i](x + y)
+
+            y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_1[i](x + y)
+
+            y = self.ffn_layers[i](x, x_mask)
+            y = self.drop(y)
+            x = self.norm_layers_2[i](x + y)
+        x = x * x_mask
+        return x
+
+
+class MultiHeadAttention(nn.Module):
+    def __init__(
+        self,
+        channels,
+        out_channels,
+        n_heads,
+        p_dropout=0.0,
+        window_size=None,
+        heads_share=True,
+        block_length=None,
+        proximal_bias=False,
+        proximal_init=False,
+    ):
+        super().__init__()
+        assert channels % n_heads == 0
+
+        self.channels = channels
+        self.out_channels = out_channels
+        self.n_heads = n_heads
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+        self.heads_share = heads_share
+        self.block_length = block_length
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+        self.attn = None
+
+        self.k_channels = channels // n_heads
+        self.conv_q = nn.Conv1d(channels, channels, 1)
+        self.conv_k = nn.Conv1d(channels, channels, 1)
+        self.conv_v = nn.Conv1d(channels, channels, 1)
+        self.conv_o = nn.Conv1d(channels, out_channels, 1)
+        self.drop = nn.Dropout(p_dropout)
+
+        if window_size is not None:
+            n_heads_rel = 1 if heads_share else n_heads
+            rel_stddev = self.k_channels**-0.5
+            self.emb_rel_k = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+            self.emb_rel_v = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+
+        nn.init.xavier_uniform_(self.conv_q.weight)
+        nn.init.xavier_uniform_(self.conv_k.weight)
+        nn.init.xavier_uniform_(self.conv_v.weight)
+        if proximal_init:
+            with torch.no_grad():
+                self.conv_k.weight.copy_(self.conv_q.weight)
+                self.conv_k.bias.copy_(self.conv_q.bias)
+
+    def forward(self, x, c, attn_mask=None):
+        q = self.conv_q(x)
+        k = self.conv_k(c)
+        v = self.conv_v(c)
+
+        x, self.attn = self.attention(q, k, v, mask=attn_mask)
+
+        x = self.conv_o(x)
+        return x
+
+    def attention(self, query, key, value, mask=None):
+        # reshape [b, d, t] -> [b, n_h, t, d_k]
+        b, d, t_s, t_t = (*key.size(), query.size(2))
+        query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
+        key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+        value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+
+        scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
+        if self.window_size is not None:
+            assert (
+                t_s == t_t
+            ), "Relative attention is only available for self-attention."
+            key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+            rel_logits = self._matmul_with_relative_keys(
+                query / math.sqrt(self.k_channels), key_relative_embeddings
+            )
+            scores_local = self._relative_position_to_absolute_position(rel_logits)
+            scores = scores + scores_local
+        if self.proximal_bias:
+            assert t_s == t_t, "Proximal bias is only available for self-attention."
+            scores = scores + self._attention_bias_proximal(t_s).to(
+                device=scores.device, dtype=scores.dtype
+            )
+        if mask is not None:
+            scores = scores.masked_fill(mask == 0, -1e4)
+            if self.block_length is not None:
+                assert (
+                    t_s == t_t
+                ), "Local attention is only available for self-attention."
+                block_mask = (
+                    torch.ones_like(scores)
+                    .triu(-self.block_length)
+                    .tril(self.block_length)
+                )
+                scores = scores.masked_fill(block_mask == 0, -1e4)
+        p_attn = F.softmax(scores, dim=-1)  # [b, n_h, t_t, t_s]
+        p_attn = self.drop(p_attn)
+        output = torch.matmul(p_attn, value)
+        if self.window_size is not None:
+            relative_weights = self._absolute_position_to_relative_position(p_attn)
+            value_relative_embeddings = self._get_relative_embeddings(
+                self.emb_rel_v, t_s
+            )
+            output = output + self._matmul_with_relative_values(
+                relative_weights, value_relative_embeddings
+            )
+        output = (
+            output.transpose(2, 3).contiguous().view(b, d, t_t)
+        )  # [b, n_h, t_t, d_k] -> [b, d, t_t]
+        return output, p_attn
+
+    def _matmul_with_relative_values(self, x, y):
+        """
+        x: [b, h, l, m]
+        y: [h or 1, m, d]
+        ret: [b, h, l, d]
+        """
+        ret = torch.matmul(x, y.unsqueeze(0))
+        return ret
+
+    def _matmul_with_relative_keys(self, x, y):
+        """
+        x: [b, h, l, d]
+        y: [h or 1, m, d]
+        ret: [b, h, l, m]
+        """
+        ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
+        return ret
+
+    def _get_relative_embeddings(self, relative_embeddings, length):
+        max_relative_position = 2 * self.window_size + 1
+        # Pad first before slice to avoid using cond ops.
+        pad_length = max(length - (self.window_size + 1), 0)
+        slice_start_position = max((self.window_size + 1) - length, 0)
+        slice_end_position = slice_start_position + 2 * length - 1
+        if pad_length > 0:
+            padded_relative_embeddings = F.pad(
+                relative_embeddings,
+                commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]),
+            )
+        else:
+            padded_relative_embeddings = relative_embeddings
+        used_relative_embeddings = padded_relative_embeddings[
+            :, slice_start_position:slice_end_position
+        ]
+        return used_relative_embeddings
+
+    def _relative_position_to_absolute_position(self, x):
+        """
+        x: [b, h, l, 2*l-1]
+        ret: [b, h, l, l]
+        """
+        batch, heads, length, _ = x.size()
+        # Concat columns of pad to shift from relative to absolute indexing.
+        x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]]))
+
+        # Concat extra elements so to add up to shape (len+1, 2*len-1).
+        x_flat = x.view([batch, heads, length * 2 * length])
+        x_flat = F.pad(
+            x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [0, length - 1]])
+        )
+
+        # Reshape and slice out the padded elements.
+        x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[
+            :, :, :length, length - 1 :
+        ]
+        return x_final
+
+    def _absolute_position_to_relative_position(self, x):
+        """
+        x: [b, h, l, l]
+        ret: [b, h, l, 2*l-1]
+        """
+        batch, heads, length, _ = x.size()
+        # padd along column
+        x = F.pad(
+            x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]])
+        )
+        x_flat = x.view([batch, heads, length**2 + length * (length - 1)])
+        # add 0's in the beginning that will skew the elements after reshape
+        x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
+        x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:]
+        return x_final
+
+    def _attention_bias_proximal(self, length):
+        """Bias for self-attention to encourage attention to close positions.
+        Args:
+          length: an integer scalar.
+        Returns:
+          a Tensor with shape [1, 1, length, length]
+        """
+        r = torch.arange(length, dtype=torch.float32)
+        diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
+        return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
+
+
+class FFN(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        filter_channels,
+        kernel_size,
+        p_dropout=0.0,
+        activation=None,
+        causal=False,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.activation = activation
+        self.causal = causal
+
+        if causal:
+            self.padding = self._causal_padding
+        else:
+            self.padding = self._same_padding
+
+        self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
+        self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
+        self.drop = nn.Dropout(p_dropout)
+
+    def forward(self, x, x_mask):
+        x = self.conv_1(self.padding(x * x_mask))
+        if self.activation == "gelu":
+            x = x * torch.sigmoid(1.702 * x)
+        else:
+            x = torch.relu(x)
+        x = self.drop(x)
+        x = self.conv_2(self.padding(x * x_mask))
+        return x * x_mask
+
+    def _causal_padding(self, x):
+        if self.kernel_size == 1:
+            return x
+        pad_l = self.kernel_size - 1
+        pad_r = 0
+        padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = F.pad(x, commons.convert_pad_shape(padding))
+        return x
+
+    def _same_padding(self, x):
+        if self.kernel_size == 1:
+            return x
+        pad_l = (self.kernel_size - 1) // 2
+        pad_r = self.kernel_size // 2
+        padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = F.pad(x, commons.convert_pad_shape(padding))
+        return x

src/infer_pack/commons.py

@@ -0,0 +1,166 @@
+import math
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+def convert_pad_shape(pad_shape):
+    l = pad_shape[::-1]
+    pad_shape = [item for sublist in l for item in sublist]
+    return pad_shape
+
+
+def kl_divergence(m_p, logs_p, m_q, logs_q):
+    """KL(P||Q)"""
+    kl = (logs_q - logs_p) - 0.5
+    kl += (
+        0.5 * (torch.exp(2.0 * logs_p) + ((m_p - m_q) ** 2)) * torch.exp(-2.0 * logs_q)
+    )
+    return kl
+
+
+def rand_gumbel(shape):
+    """Sample from the Gumbel distribution, protect from overflows."""
+    uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
+    return -torch.log(-torch.log(uniform_samples))
+
+
+def rand_gumbel_like(x):
+    g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
+    return g
+
+
+def slice_segments(x, ids_str, segment_size=4):
+    ret = torch.zeros_like(x[:, :, :segment_size])
+    for i in range(x.size(0)):
+        idx_str = ids_str[i]
+        idx_end = idx_str + segment_size
+        ret[i] = x[i, :, idx_str:idx_end]
+    return ret
+
+
+def slice_segments2(x, ids_str, segment_size=4):
+    ret = torch.zeros_like(x[:, :segment_size])
+    for i in range(x.size(0)):
+        idx_str = ids_str[i]
+        idx_end = idx_str + segment_size
+        ret[i] = x[i, idx_str:idx_end]
+    return ret
+
+
+def rand_slice_segments(x, x_lengths=None, segment_size=4):
+    b, d, t = x.size()
+    if x_lengths is None:
+        x_lengths = t
+    ids_str_max = x_lengths - segment_size + 1
+    ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
+    ret = slice_segments(x, ids_str, segment_size)
+    return ret, ids_str
+
+
+def get_timing_signal_1d(length, channels, min_timescale=1.0, max_timescale=1.0e4):
+    position = torch.arange(length, dtype=torch.float)
+    num_timescales = channels // 2
+    log_timescale_increment = math.log(float(max_timescale) / float(min_timescale)) / (
+        num_timescales - 1
+    )
+    inv_timescales = min_timescale * torch.exp(
+        torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment
+    )
+    scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
+    signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
+    signal = F.pad(signal, [0, 0, 0, channels % 2])
+    signal = signal.view(1, channels, length)
+    return signal
+
+
+def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
+    b, channels, length = x.size()
+    signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+    return x + signal.to(dtype=x.dtype, device=x.device)
+
+
+def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
+    b, channels, length = x.size()
+    signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+    return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
+
+
+def subsequent_mask(length):
+    mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
+    return mask
+
+
+@torch.jit.script
+def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
+    n_channels_int = n_channels[0]
+    in_act = input_a + input_b
+    t_act = torch.tanh(in_act[:, :n_channels_int, :])
+    s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
+    acts = t_act * s_act
+    return acts
+
+
+def convert_pad_shape(pad_shape):
+    l = pad_shape[::-1]
+    pad_shape = [item for sublist in l for item in sublist]
+    return pad_shape
+
+
+def shift_1d(x):
+    x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
+    return x
+
+
+def sequence_mask(length, max_length=None):
+    if max_length is None:
+        max_length = length.max()
+    x = torch.arange(max_length, dtype=length.dtype, device=length.device)
+    return x.unsqueeze(0) < length.unsqueeze(1)
+
+
+def generate_path(duration, mask):
+    """
+    duration: [b, 1, t_x]
+    mask: [b, 1, t_y, t_x]
+    """
+    device = duration.device
+
+    b, _, t_y, t_x = mask.shape
+    cum_duration = torch.cumsum(duration, -1)
+
+    cum_duration_flat = cum_duration.view(b * t_x)
+    path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
+    path = path.view(b, t_x, t_y)
+    path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
+    path = path.unsqueeze(1).transpose(2, 3) * mask
+    return path
+
+
+def clip_grad_value_(parameters, clip_value, norm_type=2):
+    if isinstance(parameters, torch.Tensor):
+        parameters = [parameters]
+    parameters = list(filter(lambda p: p.grad is not None, parameters))
+    norm_type = float(norm_type)
+    if clip_value is not None:
+        clip_value = float(clip_value)
+
+    total_norm = 0
+    for p in parameters:
+        param_norm = p.grad.data.norm(norm_type)
+        total_norm += param_norm.item() ** norm_type
+        if clip_value is not None:
+            p.grad.data.clamp_(min=-clip_value, max=clip_value)
+    total_norm = total_norm ** (1.0 / norm_type)
+    return total_norm

src/infer_pack/models.py

@@ -0,0 +1,1124 @@
+import math, pdb, os
+from time import time as ttime
+import torch
+from torch import nn
+from torch.nn import functional as F
+from infer_pack import modules
+from infer_pack import attentions
+from infer_pack import commons
+from infer_pack.commons import init_weights, get_padding
+from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
+from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
+from infer_pack.commons import init_weights
+import numpy as np
+from infer_pack import commons
+
+
+class TextEncoder256(nn.Module):
+    def __init__(
+        self,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        f0=True,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.emb_phone = nn.Linear(256, hidden_channels)
+        self.lrelu = nn.LeakyReLU(0.1, inplace=True)
+        if f0 == True:
+            self.emb_pitch = nn.Embedding(256, hidden_channels)  # pitch 256
+        self.encoder = attentions.Encoder(
+            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, phone, pitch, lengths):
+        if pitch == None:
+            x = self.emb_phone(phone)
+        else:
+            x = self.emb_phone(phone) + self.emb_pitch(pitch)
+        x = x * math.sqrt(self.hidden_channels)  # [b, t, h]
+        x = self.lrelu(x)
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.encoder(x * x_mask, x_mask)
+        stats = self.proj(x) * x_mask
+
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return m, logs, x_mask
+
+
+class TextEncoder768(nn.Module):
+    def __init__(
+        self,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        f0=True,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.emb_phone = nn.Linear(768, hidden_channels)
+        self.lrelu = nn.LeakyReLU(0.1, inplace=True)
+        if f0 == True:
+            self.emb_pitch = nn.Embedding(256, hidden_channels)  # pitch 256
+        self.encoder = attentions.Encoder(
+            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, phone, pitch, lengths):
+        if pitch == None:
+            x = self.emb_phone(phone)
+        else:
+            x = self.emb_phone(phone) + self.emb_pitch(pitch)
+        x = x * math.sqrt(self.hidden_channels)  # [b, t, h]
+        x = self.lrelu(x)
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.encoder(x * x_mask, x_mask)
+        stats = self.proj(x) * x_mask
+
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return m, logs, x_mask
+
+
+class ResidualCouplingBlock(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        n_flows=4,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+        for i in range(n_flows):
+            self.flows.append(
+                modules.ResidualCouplingLayer(
+                    channels,
+                    hidden_channels,
+                    kernel_size,
+                    dilation_rate,
+                    n_layers,
+                    gin_channels=gin_channels,
+                    mean_only=True,
+                )
+            )
+            self.flows.append(modules.Flip())
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x = flow(x, x_mask, g=g, reverse=reverse)
+        return x
+
+    def remove_weight_norm(self):
+        for i in range(self.n_flows):
+            self.flows[i * 2].remove_weight_norm()
+
+
+class PosteriorEncoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+
+        self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
+        self.enc = modules.WN(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            gin_channels=gin_channels,
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x, x_lengths, g=None):
+        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.pre(x) * x_mask
+        x = self.enc(x, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
+        return z, m, logs, x_mask
+
+    def remove_weight_norm(self):
+        self.enc.remove_weight_norm()
+
+
+class Generator(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels=0,
+    ):
+        super(Generator, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+    def forward(self, x, g=None):
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+
+class SineGen(torch.nn.Module):
+    """Definition of sine generator
+    SineGen(samp_rate, harmonic_num = 0,
+            sine_amp = 0.1, noise_std = 0.003,
+            voiced_threshold = 0,
+            flag_for_pulse=False)
+    samp_rate: sampling rate in Hz
+    harmonic_num: number of harmonic overtones (default 0)
+    sine_amp: amplitude of sine-wavefrom (default 0.1)
+    noise_std: std of Gaussian noise (default 0.003)
+    voiced_thoreshold: F0 threshold for U/V classification (default 0)
+    flag_for_pulse: this SinGen is used inside PulseGen (default False)
+    Note: when flag_for_pulse is True, the first time step of a voiced
+        segment is always sin(np.pi) or cos(0)
+    """
+
+    def __init__(
+        self,
+        samp_rate,
+        harmonic_num=0,
+        sine_amp=0.1,
+        noise_std=0.003,
+        voiced_threshold=0,
+        flag_for_pulse=False,
+    ):
+        super(SineGen, self).__init__()
+        self.sine_amp = sine_amp
+        self.noise_std = noise_std
+        self.harmonic_num = harmonic_num
+        self.dim = self.harmonic_num + 1
+        self.sampling_rate = samp_rate
+        self.voiced_threshold = voiced_threshold
+
+    def _f02uv(self, f0):
+        # generate uv signal
+        uv = torch.ones_like(f0)
+        uv = uv * (f0 > self.voiced_threshold)
+        return uv
+
+    def forward(self, f0, upp):
+        """sine_tensor, uv = forward(f0)
+        input F0: tensor(batchsize=1, length, dim=1)
+                  f0 for unvoiced steps should be 0
+        output sine_tensor: tensor(batchsize=1, length, dim)
+        output uv: tensor(batchsize=1, length, 1)
+        """
+        with torch.no_grad():
+            f0 = f0[:, None].transpose(1, 2)
+            f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, device=f0.device)
+            # fundamental component
+            f0_buf[:, :, 0] = f0[:, :, 0]
+            for idx in np.arange(self.harmonic_num):
+                f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (
+                    idx + 2
+                )  # idx + 2: the (idx+1)-th overtone, (idx+2)-th harmonic
+            rad_values = (f0_buf / self.sampling_rate) % 1  ###%1意味着n_har的乘积无法后处理优化
+            rand_ini = torch.rand(
+                f0_buf.shape[0], f0_buf.shape[2], device=f0_buf.device
+            )
+            rand_ini[:, 0] = 0
+            rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
+            tmp_over_one = torch.cumsum(rad_values, 1)  # % 1  #####%1意味着后面的cumsum无法再优化
+            tmp_over_one *= upp
+            tmp_over_one = F.interpolate(
+                tmp_over_one.transpose(2, 1),
+                scale_factor=upp,
+                mode="linear",
+                align_corners=True,
+            ).transpose(2, 1)
+            rad_values = F.interpolate(
+                rad_values.transpose(2, 1), scale_factor=upp, mode="nearest"
+            ).transpose(
+                2, 1
+            )  #######
+            tmp_over_one %= 1
+            tmp_over_one_idx = (tmp_over_one[:, 1:, :] - tmp_over_one[:, :-1, :]) < 0
+            cumsum_shift = torch.zeros_like(rad_values)
+            cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
+            sine_waves = torch.sin(
+                torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi
+            )
+            sine_waves = sine_waves * self.sine_amp
+            uv = self._f02uv(f0)
+            uv = F.interpolate(
+                uv.transpose(2, 1), scale_factor=upp, mode="nearest"
+            ).transpose(2, 1)
+            noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+            noise = noise_amp * torch.randn_like(sine_waves)
+            sine_waves = sine_waves * uv + noise
+        return sine_waves, uv, noise
+
+
+class SourceModuleHnNSF(torch.nn.Module):
+    """SourceModule for hn-nsf
+    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0)
+    sampling_rate: sampling_rate in Hz
+    harmonic_num: number of harmonic above F0 (default: 0)
+    sine_amp: amplitude of sine source signal (default: 0.1)
+    add_noise_std: std of additive Gaussian noise (default: 0.003)
+        note that amplitude of noise in unvoiced is decided
+        by sine_amp
+    voiced_threshold: threhold to set U/V given F0 (default: 0)
+    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+    F0_sampled (batchsize, length, 1)
+    Sine_source (batchsize, length, 1)
+    noise_source (batchsize, length 1)
+    uv (batchsize, length, 1)
+    """
+
+    def __init__(
+        self,
+        sampling_rate,
+        harmonic_num=0,
+        sine_amp=0.1,
+        add_noise_std=0.003,
+        voiced_threshod=0,
+        is_half=True,
+    ):
+        super(SourceModuleHnNSF, self).__init__()
+
+        self.sine_amp = sine_amp
+        self.noise_std = add_noise_std
+        self.is_half = is_half
+        # to produce sine waveforms
+        self.l_sin_gen = SineGen(
+            sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod
+        )
+
+        # to merge source harmonics into a single excitation
+        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
+        self.l_tanh = torch.nn.Tanh()
+
+    def forward(self, x, upp=None):
+        sine_wavs, uv, _ = self.l_sin_gen(x, upp)
+        if self.is_half:
+            sine_wavs = sine_wavs.half()
+        sine_merge = self.l_tanh(self.l_linear(sine_wavs))
+        return sine_merge, None, None  # noise, uv
+
+
+class GeneratorNSF(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels,
+        sr,
+        is_half=False,
+    ):
+        super(GeneratorNSF, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+
+        self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates))
+        self.m_source = SourceModuleHnNSF(
+            sampling_rate=sr, harmonic_num=0, is_half=is_half
+        )
+        self.noise_convs = nn.ModuleList()
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            c_cur = upsample_initial_channel // (2 ** (i + 1))
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+            if i + 1 < len(upsample_rates):
+                stride_f0 = np.prod(upsample_rates[i + 1 :])
+                self.noise_convs.append(
+                    Conv1d(
+                        1,
+                        c_cur,
+                        kernel_size=stride_f0 * 2,
+                        stride=stride_f0,
+                        padding=stride_f0 // 2,
+                    )
+                )
+            else:
+                self.noise_convs.append(Conv1d(1, c_cur, kernel_size=1))
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+        self.upp = np.prod(upsample_rates)
+
+    def forward(self, x, f0, g=None):
+        har_source, noi_source, uv = self.m_source(f0, self.upp)
+        har_source = har_source.transpose(1, 2)
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+            x_source = self.noise_convs[i](har_source)
+            x = x + x_source
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+
+sr2sr = {
+    "32k": 32000,
+    "40k": 40000,
+    "48k": 48000,
+}
+
+
+class SynthesizerTrnMs256NSFsid(nn.Module):
+    def __init__(
+        self,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        spk_embed_dim,
+        gin_channels,
+        sr,
+        **kwargs
+    ):
+        super().__init__()
+        if type(sr) == type("strr"):
+            sr = sr2sr[sr]
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.gin_channels = gin_channels
+        # self.hop_length = hop_length#
+        self.spk_embed_dim = spk_embed_dim
+        self.enc_p = TextEncoder256(
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+        )
+        self.dec = GeneratorNSF(
+            inter_channels,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            gin_channels=gin_channels,
+            sr=sr,
+            is_half=kwargs["is_half"],
+        )
+        self.enc_q = PosteriorEncoder(
+            spec_channels,
+            inter_channels,
+            hidden_channels,
+            5,
+            1,
+            16,
+            gin_channels=gin_channels,
+        )
+        self.flow = ResidualCouplingBlock(
+            inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels
+        )
+        self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
+        print("gin_channels:", gin_channels, "self.spk_embed_dim:", self.spk_embed_dim)
+
+    def remove_weight_norm(self):
+        self.dec.remove_weight_norm()
+        self.flow.remove_weight_norm()
+        self.enc_q.remove_weight_norm()
+
+    def forward(
+        self, phone, phone_lengths, pitch, pitchf, y, y_lengths, ds
+    ):  # 这里ds是id，[bs,1]
+        # print(1,pitch.shape)#[bs,t]
+        g = self.emb_g(ds).unsqueeze(-1)  # [b, 256, 1]##1是t，广播的
+        m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+        z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
+        z_p = self.flow(z, y_mask, g=g)
+        z_slice, ids_slice = commons.rand_slice_segments(
+            z, y_lengths, self.segment_size
+        )
+        # print(-1,pitchf.shape,ids_slice,self.segment_size,self.hop_length,self.segment_size//self.hop_length)
+        pitchf = commons.slice_segments2(pitchf, ids_slice, self.segment_size)
+        # print(-2,pitchf.shape,z_slice.shape)
+        o = self.dec(z_slice, pitchf, g=g)
+        return o, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
+
+    def infer(self, phone, phone_lengths, pitch, nsff0, sid, max_len=None):
+        g = self.emb_g(sid).unsqueeze(-1)
+        m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+        z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
+        z = self.flow(z_p, x_mask, g=g, reverse=True)
+        o = self.dec((z * x_mask)[:, :, :max_len], nsff0, g=g)
+        return o, x_mask, (z, z_p, m_p, logs_p)
+
+
+class SynthesizerTrnMs768NSFsid(nn.Module):
+    def __init__(
+        self,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        spk_embed_dim,
+        gin_channels,
+        sr,
+        **kwargs
+    ):
+        super().__init__()
+        if type(sr) == type("strr"):
+            sr = sr2sr[sr]
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.gin_channels = gin_channels
+        # self.hop_length = hop_length#
+        self.spk_embed_dim = spk_embed_dim
+        self.enc_p = TextEncoder768(
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+        )
+        self.dec = GeneratorNSF(
+            inter_channels,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            gin_channels=gin_channels,
+            sr=sr,
+            is_half=kwargs["is_half"],
+        )
+        self.enc_q = PosteriorEncoder(
+            spec_channels,
+            inter_channels,
+            hidden_channels,
+            5,
+            1,
+            16,
+            gin_channels=gin_channels,
+        )
+        self.flow = ResidualCouplingBlock(
+            inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels
+        )
+        self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
+        print("gin_channels:", gin_channels, "self.spk_embed_dim:", self.spk_embed_dim)
+
+    def remove_weight_norm(self):
+        self.dec.remove_weight_norm()
+        self.flow.remove_weight_norm()
+        self.enc_q.remove_weight_norm()
+
+    def forward(
+        self, phone, phone_lengths, pitch, pitchf, y, y_lengths, ds
+    ):  # 这里ds是id，[bs,1]
+        # print(1,pitch.shape)#[bs,t]
+        g = self.emb_g(ds).unsqueeze(-1)  # [b, 256, 1]##1是t，广播的
+        m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+        z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
+        z_p = self.flow(z, y_mask, g=g)
+        z_slice, ids_slice = commons.rand_slice_segments(
+            z, y_lengths, self.segment_size
+        )
+        # print(-1,pitchf.shape,ids_slice,self.segment_size,self.hop_length,self.segment_size//self.hop_length)
+        pitchf = commons.slice_segments2(pitchf, ids_slice, self.segment_size)
+        # print(-2,pitchf.shape,z_slice.shape)
+        o = self.dec(z_slice, pitchf, g=g)
+        return o, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
+
+    def infer(self, phone, phone_lengths, pitch, nsff0, sid, max_len=None):
+        g = self.emb_g(sid).unsqueeze(-1)
+        m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+        z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
+        z = self.flow(z_p, x_mask, g=g, reverse=True)
+        o = self.dec((z * x_mask)[:, :, :max_len], nsff0, g=g)
+        return o, x_mask, (z, z_p, m_p, logs_p)
+
+
+class SynthesizerTrnMs256NSFsid_nono(nn.Module):
+    def __init__(
+        self,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        spk_embed_dim,
+        gin_channels,
+        sr=None,
+        **kwargs
+    ):
+        super().__init__()
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.gin_channels = gin_channels
+        # self.hop_length = hop_length#
+        self.spk_embed_dim = spk_embed_dim
+        self.enc_p = TextEncoder256(
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+            f0=False,
+        )
+        self.dec = Generator(
+            inter_channels,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            gin_channels=gin_channels,
+        )
+        self.enc_q = PosteriorEncoder(
+            spec_channels,
+            inter_channels,
+            hidden_channels,
+            5,
+            1,
+            16,
+            gin_channels=gin_channels,
+        )
+        self.flow = ResidualCouplingBlock(
+            inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels
+        )
+        self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
+        print("gin_channels:", gin_channels, "self.spk_embed_dim:", self.spk_embed_dim)
+
+    def remove_weight_norm(self):
+        self.dec.remove_weight_norm()
+        self.flow.remove_weight_norm()
+        self.enc_q.remove_weight_norm()
+
+    def forward(self, phone, phone_lengths, y, y_lengths, ds):  # 这里ds是id，[bs,1]
+        g = self.emb_g(ds).unsqueeze(-1)  # [b, 256, 1]##1是t，广播的
+        m_p, logs_p, x_mask = self.enc_p(phone, None, phone_lengths)
+        z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
+        z_p = self.flow(z, y_mask, g=g)
+        z_slice, ids_slice = commons.rand_slice_segments(
+            z, y_lengths, self.segment_size
+        )
+        o = self.dec(z_slice, g=g)
+        return o, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
+
+    def infer(self, phone, phone_lengths, sid, max_len=None):
+        g = self.emb_g(sid).unsqueeze(-1)
+        m_p, logs_p, x_mask = self.enc_p(phone, None, phone_lengths)
+        z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
+        z = self.flow(z_p, x_mask, g=g, reverse=True)
+        o = self.dec((z * x_mask)[:, :, :max_len], g=g)
+        return o, x_mask, (z, z_p, m_p, logs_p)
+
+
+class SynthesizerTrnMs768NSFsid_nono(nn.Module):
+    def __init__(
+        self,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        spk_embed_dim,
+        gin_channels,
+        sr=None,
+        **kwargs
+    ):
+        super().__init__()
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.gin_channels = gin_channels
+        # self.hop_length = hop_length#
+        self.spk_embed_dim = spk_embed_dim
+        self.enc_p = TextEncoder768(
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+            f0=False,
+        )
+        self.dec = Generator(
+            inter_channels,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            gin_channels=gin_channels,
+        )
+        self.enc_q = PosteriorEncoder(
+            spec_channels,
+            inter_channels,
+            hidden_channels,
+            5,
+            1,
+            16,
+            gin_channels=gin_channels,
+        )
+        self.flow = ResidualCouplingBlock(
+            inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels
+        )
+        self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
+        print("gin_channels:", gin_channels, "self.spk_embed_dim:", self.spk_embed_dim)
+
+    def remove_weight_norm(self):
+        self.dec.remove_weight_norm()
+        self.flow.remove_weight_norm()
+        self.enc_q.remove_weight_norm()
+
+    def forward(self, phone, phone_lengths, y, y_lengths, ds):  # 这里ds是id，[bs,1]
+        g = self.emb_g(ds).unsqueeze(-1)  # [b, 256, 1]##1是t，广播的
+        m_p, logs_p, x_mask = self.enc_p(phone, None, phone_lengths)
+        z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
+        z_p = self.flow(z, y_mask, g=g)
+        z_slice, ids_slice = commons.rand_slice_segments(
+            z, y_lengths, self.segment_size
+        )
+        o = self.dec(z_slice, g=g)
+        return o, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
+
+    def infer(self, phone, phone_lengths, sid, max_len=None):
+        g = self.emb_g(sid).unsqueeze(-1)
+        m_p, logs_p, x_mask = self.enc_p(phone, None, phone_lengths)
+        z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
+        z = self.flow(z_p, x_mask, g=g, reverse=True)
+        o = self.dec((z * x_mask)[:, :, :max_len], g=g)
+        return o, x_mask, (z, z_p, m_p, logs_p)
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminator, self).__init__()
+        periods = [2, 3, 5, 7, 11, 17]
+        # periods = [3, 5, 7, 11, 17, 23, 37]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [
+            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
+        ]
+        self.discriminators = nn.ModuleList(discs)
+
+    def forward(self, y, y_hat):
+        y_d_rs = []  #
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            # for j in range(len(fmap_r)):
+            #     print(i,j,y.shape,y_hat.shape,fmap_r[j].shape,fmap_g[j].shape)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class MultiPeriodDiscriminatorV2(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminatorV2, self).__init__()
+        # periods = [2, 3, 5, 7, 11, 17]
+        periods = [2, 3, 5, 7, 11, 17, 23, 37]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [
+            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
+        ]
+        self.discriminators = nn.ModuleList(discs)
+
+    def forward(self, y, y_hat):
+        y_d_rs = []  #
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            # for j in range(len(fmap_r)):
+            #     print(i,j,y.shape,y_hat.shape,fmap_r[j].shape,fmap_g[j].shape)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class DiscriminatorS(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(DiscriminatorS, self).__init__()
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(Conv1d(1, 16, 15, 1, padding=7)),
+                norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
+                norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
+                norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
+                norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
+                norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
+            ]
+        )
+        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
+
+    def forward(self, x):
+        fmap = []
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class DiscriminatorP(torch.nn.Module):
+    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
+        super(DiscriminatorP, self).__init__()
+        self.period = period
+        self.use_spectral_norm = use_spectral_norm
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(
+                    Conv2d(
+                        1,
+                        32,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        32,
+                        128,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        128,
+                        512,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        512,
+                        1024,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        1024,
+                        1024,
+                        (kernel_size, 1),
+                        1,
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+            ]
+        )
+        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+
+    def forward(self, x):
+        fmap = []
+
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0:  # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), "reflect")
+            t = t + n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap

src/infer_pack/models_onnx.py

@@ -0,0 +1,818 @@
+import math, pdb, os
+from time import time as ttime
+import torch
+from torch import nn
+from torch.nn import functional as F
+from infer_pack import modules
+from infer_pack import attentions
+from infer_pack import commons
+from infer_pack.commons import init_weights, get_padding
+from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
+from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
+from infer_pack.commons import init_weights
+import numpy as np
+from infer_pack import commons
+
+
+class TextEncoder256(nn.Module):
+    def __init__(
+        self,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        f0=True,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.emb_phone = nn.Linear(256, hidden_channels)
+        self.lrelu = nn.LeakyReLU(0.1, inplace=True)
+        if f0 == True:
+            self.emb_pitch = nn.Embedding(256, hidden_channels)  # pitch 256
+        self.encoder = attentions.Encoder(
+            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, phone, pitch, lengths):
+        if pitch == None:
+            x = self.emb_phone(phone)
+        else:
+            x = self.emb_phone(phone) + self.emb_pitch(pitch)
+        x = x * math.sqrt(self.hidden_channels)  # [b, t, h]
+        x = self.lrelu(x)
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.encoder(x * x_mask, x_mask)
+        stats = self.proj(x) * x_mask
+
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return m, logs, x_mask
+
+
+class TextEncoder768(nn.Module):
+    def __init__(
+        self,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        f0=True,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.emb_phone = nn.Linear(768, hidden_channels)
+        self.lrelu = nn.LeakyReLU(0.1, inplace=True)
+        if f0 == True:
+            self.emb_pitch = nn.Embedding(256, hidden_channels)  # pitch 256
+        self.encoder = attentions.Encoder(
+            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, phone, pitch, lengths):
+        if pitch == None:
+            x = self.emb_phone(phone)
+        else:
+            x = self.emb_phone(phone) + self.emb_pitch(pitch)
+        x = x * math.sqrt(self.hidden_channels)  # [b, t, h]
+        x = self.lrelu(x)
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.encoder(x * x_mask, x_mask)
+        stats = self.proj(x) * x_mask
+
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return m, logs, x_mask
+
+
+class ResidualCouplingBlock(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        n_flows=4,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+        for i in range(n_flows):
+            self.flows.append(
+                modules.ResidualCouplingLayer(
+                    channels,
+                    hidden_channels,
+                    kernel_size,
+                    dilation_rate,
+                    n_layers,
+                    gin_channels=gin_channels,
+                    mean_only=True,
+                )
+            )
+            self.flows.append(modules.Flip())
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x = flow(x, x_mask, g=g, reverse=reverse)
+        return x
+
+    def remove_weight_norm(self):
+        for i in range(self.n_flows):
+            self.flows[i * 2].remove_weight_norm()
+
+
+class PosteriorEncoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+
+        self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
+        self.enc = modules.WN(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            gin_channels=gin_channels,
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x, x_lengths, g=None):
+        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.pre(x) * x_mask
+        x = self.enc(x, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
+        return z, m, logs, x_mask
+
+    def remove_weight_norm(self):
+        self.enc.remove_weight_norm()
+
+
+class Generator(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels=0,
+    ):
+        super(Generator, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+    def forward(self, x, g=None):
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+
+class SineGen(torch.nn.Module):
+    """Definition of sine generator
+    SineGen(samp_rate, harmonic_num = 0,
+            sine_amp = 0.1, noise_std = 0.003,
+            voiced_threshold = 0,
+            flag_for_pulse=False)
+    samp_rate: sampling rate in Hz
+    harmonic_num: number of harmonic overtones (default 0)
+    sine_amp: amplitude of sine-wavefrom (default 0.1)
+    noise_std: std of Gaussian noise (default 0.003)
+    voiced_thoreshold: F0 threshold for U/V classification (default 0)
+    flag_for_pulse: this SinGen is used inside PulseGen (default False)
+    Note: when flag_for_pulse is True, the first time step of a voiced
+        segment is always sin(np.pi) or cos(0)
+    """
+
+    def __init__(
+        self,
+        samp_rate,
+        harmonic_num=0,
+        sine_amp=0.1,
+        noise_std=0.003,
+        voiced_threshold=0,
+        flag_for_pulse=False,
+    ):
+        super(SineGen, self).__init__()
+        self.sine_amp = sine_amp
+        self.noise_std = noise_std
+        self.harmonic_num = harmonic_num
+        self.dim = self.harmonic_num + 1
+        self.sampling_rate = samp_rate
+        self.voiced_threshold = voiced_threshold
+
+    def _f02uv(self, f0):
+        # generate uv signal
+        uv = torch.ones_like(f0)
+        uv = uv * (f0 > self.voiced_threshold)
+        return uv
+
+    def forward(self, f0, upp):
+        """sine_tensor, uv = forward(f0)
+        input F0: tensor(batchsize=1, length, dim=1)
+                  f0 for unvoiced steps should be 0
+        output sine_tensor: tensor(batchsize=1, length, dim)
+        output uv: tensor(batchsize=1, length, 1)
+        """
+        with torch.no_grad():
+            f0 = f0[:, None].transpose(1, 2)
+            f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, device=f0.device)
+            # fundamental component
+            f0_buf[:, :, 0] = f0[:, :, 0]
+            for idx in np.arange(self.harmonic_num):
+                f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (
+                    idx + 2
+                )  # idx + 2: the (idx+1)-th overtone, (idx+2)-th harmonic
+            rad_values = (f0_buf / self.sampling_rate) % 1  ###%1意味着n_har的乘积无法后处理优化
+            rand_ini = torch.rand(
+                f0_buf.shape[0], f0_buf.shape[2], device=f0_buf.device
+            )
+            rand_ini[:, 0] = 0
+            rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
+            tmp_over_one = torch.cumsum(rad_values, 1)  # % 1  #####%1意味着后面的cumsum无法再优化
+            tmp_over_one *= upp
+            tmp_over_one = F.interpolate(
+                tmp_over_one.transpose(2, 1),
+                scale_factor=upp,
+                mode="linear",
+                align_corners=True,
+            ).transpose(2, 1)
+            rad_values = F.interpolate(
+                rad_values.transpose(2, 1), scale_factor=upp, mode="nearest"
+            ).transpose(
+                2, 1
+            )  #######
+            tmp_over_one %= 1
+            tmp_over_one_idx = (tmp_over_one[:, 1:, :] - tmp_over_one[:, :-1, :]) < 0
+            cumsum_shift = torch.zeros_like(rad_values)
+            cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
+            sine_waves = torch.sin(
+                torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi
+            )
+            sine_waves = sine_waves * self.sine_amp
+            uv = self._f02uv(f0)
+            uv = F.interpolate(
+                uv.transpose(2, 1), scale_factor=upp, mode="nearest"
+            ).transpose(2, 1)
+            noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+            noise = noise_amp * torch.randn_like(sine_waves)
+            sine_waves = sine_waves * uv + noise
+        return sine_waves, uv, noise
+
+
+class SourceModuleHnNSF(torch.nn.Module):
+    """SourceModule for hn-nsf
+    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0)
+    sampling_rate: sampling_rate in Hz
+    harmonic_num: number of harmonic above F0 (default: 0)
+    sine_amp: amplitude of sine source signal (default: 0.1)
+    add_noise_std: std of additive Gaussian noise (default: 0.003)
+        note that amplitude of noise in unvoiced is decided
+        by sine_amp
+    voiced_threshold: threhold to set U/V given F0 (default: 0)
+    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+    F0_sampled (batchsize, length, 1)
+    Sine_source (batchsize, length, 1)
+    noise_source (batchsize, length 1)
+    uv (batchsize, length, 1)
+    """
+
+    def __init__(
+        self,
+        sampling_rate,
+        harmonic_num=0,
+        sine_amp=0.1,
+        add_noise_std=0.003,
+        voiced_threshod=0,
+        is_half=True,
+    ):
+        super(SourceModuleHnNSF, self).__init__()
+
+        self.sine_amp = sine_amp
+        self.noise_std = add_noise_std
+        self.is_half = is_half
+        # to produce sine waveforms
+        self.l_sin_gen = SineGen(
+            sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod
+        )
+
+        # to merge source harmonics into a single excitation
+        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
+        self.l_tanh = torch.nn.Tanh()
+
+    def forward(self, x, upp=None):
+        sine_wavs, uv, _ = self.l_sin_gen(x, upp)
+        if self.is_half:
+            sine_wavs = sine_wavs.half()
+        sine_merge = self.l_tanh(self.l_linear(sine_wavs))
+        return sine_merge, None, None  # noise, uv
+
+
+class GeneratorNSF(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels,
+        sr,
+        is_half=False,
+    ):
+        super(GeneratorNSF, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+
+        self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates))
+        self.m_source = SourceModuleHnNSF(
+            sampling_rate=sr, harmonic_num=0, is_half=is_half
+        )
+        self.noise_convs = nn.ModuleList()
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            c_cur = upsample_initial_channel // (2 ** (i + 1))
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+            if i + 1 < len(upsample_rates):
+                stride_f0 = np.prod(upsample_rates[i + 1 :])
+                self.noise_convs.append(
+                    Conv1d(
+                        1,
+                        c_cur,
+                        kernel_size=stride_f0 * 2,
+                        stride=stride_f0,
+                        padding=stride_f0 // 2,
+                    )
+                )
+            else:
+                self.noise_convs.append(Conv1d(1, c_cur, kernel_size=1))
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+        self.upp = np.prod(upsample_rates)
+
+    def forward(self, x, f0, g=None):
+        har_source, noi_source, uv = self.m_source(f0, self.upp)
+        har_source = har_source.transpose(1, 2)
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+            x_source = self.noise_convs[i](har_source)
+            x = x + x_source
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+
+sr2sr = {
+    "32k": 32000,
+    "40k": 40000,
+    "48k": 48000,
+}
+
+
+class SynthesizerTrnMsNSFsidM(nn.Module):
+    def __init__(
+        self,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        spk_embed_dim,
+        gin_channels,
+        sr,
+        **kwargs
+    ):
+        super().__init__()
+        if type(sr) == type("strr"):
+            sr = sr2sr[sr]
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.gin_channels = gin_channels
+        # self.hop_length = hop_length#
+        self.spk_embed_dim = spk_embed_dim
+        if self.gin_channels == 256:
+            self.enc_p = TextEncoder256(
+                inter_channels,
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers,
+                kernel_size,
+                p_dropout,
+            )
+        else:
+            self.enc_p = TextEncoder768(
+                inter_channels,
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers,
+                kernel_size,
+                p_dropout,
+            )
+        self.dec = GeneratorNSF(
+            inter_channels,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            gin_channels=gin_channels,
+            sr=sr,
+            is_half=kwargs["is_half"],
+        )
+        self.enc_q = PosteriorEncoder(
+            spec_channels,
+            inter_channels,
+            hidden_channels,
+            5,
+            1,
+            16,
+            gin_channels=gin_channels,
+        )
+        self.flow = ResidualCouplingBlock(
+            inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels
+        )
+        self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
+        self.speaker_map = None
+        print("gin_channels:", gin_channels, "self.spk_embed_dim:", self.spk_embed_dim)
+
+    def remove_weight_norm(self):
+        self.dec.remove_weight_norm()
+        self.flow.remove_weight_norm()
+        self.enc_q.remove_weight_norm()
+
+    def construct_spkmixmap(self, n_speaker):
+        self.speaker_map = torch.zeros((n_speaker, 1, 1, self.gin_channels))
+        for i in range(n_speaker):
+            self.speaker_map[i] = self.emb_g(torch.LongTensor([[i]]))
+        self.speaker_map = self.speaker_map.unsqueeze(0)
+
+    def forward(self, phone, phone_lengths, pitch, nsff0, g, rnd, max_len=None):
+        if self.speaker_map is not None:  # [N, S]  *  [S, B, 1, H]
+            g = g.reshape((g.shape[0], g.shape[1], 1, 1, 1))  # [N, S, B, 1, 1]
+            g = g * self.speaker_map  # [N, S, B, 1, H]
+            g = torch.sum(g, dim=1)  # [N, 1, B, 1, H]
+            g = g.transpose(0, -1).transpose(0, -2).squeeze(0)  # [B, H, N]
+        else:
+            g = g.unsqueeze(0)
+            g = self.emb_g(g).transpose(1, 2)
+
+        m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+        z_p = (m_p + torch.exp(logs_p) * rnd) * x_mask
+        z = self.flow(z_p, x_mask, g=g, reverse=True)
+        o = self.dec((z * x_mask)[:, :, :max_len], nsff0, g=g)
+        return o
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminator, self).__init__()
+        periods = [2, 3, 5, 7, 11, 17]
+        # periods = [3, 5, 7, 11, 17, 23, 37]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [
+            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
+        ]
+        self.discriminators = nn.ModuleList(discs)
+
+    def forward(self, y, y_hat):
+        y_d_rs = []  #
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            # for j in range(len(fmap_r)):
+            #     print(i,j,y.shape,y_hat.shape,fmap_r[j].shape,fmap_g[j].shape)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class MultiPeriodDiscriminatorV2(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminatorV2, self).__init__()
+        # periods = [2, 3, 5, 7, 11, 17]
+        periods = [2, 3, 5, 7, 11, 17, 23, 37]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [
+            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
+        ]
+        self.discriminators = nn.ModuleList(discs)
+
+    def forward(self, y, y_hat):
+        y_d_rs = []  #
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            # for j in range(len(fmap_r)):
+            #     print(i,j,y.shape,y_hat.shape,fmap_r[j].shape,fmap_g[j].shape)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class DiscriminatorS(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(DiscriminatorS, self).__init__()
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(Conv1d(1, 16, 15, 1, padding=7)),
+                norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
+                norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
+                norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
+                norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
+                norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
+            ]
+        )
+        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
+
+    def forward(self, x):
+        fmap = []
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class DiscriminatorP(torch.nn.Module):
+    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
+        super(DiscriminatorP, self).__init__()
+        self.period = period
+        self.use_spectral_norm = use_spectral_norm
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(
+                    Conv2d(
+                        1,
+                        32,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        32,
+                        128,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        128,
+                        512,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        512,
+                        1024,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        1024,
+                        1024,
+                        (kernel_size, 1),
+                        1,
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+            ]
+        )
+        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+
+    def forward(self, x):
+        fmap = []
+
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0:  # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), "reflect")
+            t = t + n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap

src/infer_pack/models_onnx_moess.py

@@ -0,0 +1,849 @@
+import math, pdb, os
+from time import time as ttime
+import torch
+from torch import nn
+from torch.nn import functional as F
+from infer_pack import modules
+from infer_pack import attentions
+from infer_pack import commons
+from infer_pack.commons import init_weights, get_padding
+from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
+from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
+from infer_pack.commons import init_weights
+import numpy as np
+from infer_pack import commons
+
+
+class TextEncoder256(nn.Module):
+    def __init__(
+        self,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        f0=True,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.emb_phone = nn.Linear(256, hidden_channels)
+        self.lrelu = nn.LeakyReLU(0.1, inplace=True)
+        if f0 == True:
+            self.emb_pitch = nn.Embedding(256, hidden_channels)  # pitch 256
+        self.encoder = attentions.Encoder(
+            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, phone, pitch, lengths):
+        if pitch == None:
+            x = self.emb_phone(phone)
+        else:
+            x = self.emb_phone(phone) + self.emb_pitch(pitch)
+        x = x * math.sqrt(self.hidden_channels)  # [b, t, h]
+        x = self.lrelu(x)
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.encoder(x * x_mask, x_mask)
+        stats = self.proj(x) * x_mask
+
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return m, logs, x_mask
+
+
+class TextEncoder256Sim(nn.Module):
+    def __init__(
+        self,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        f0=True,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.emb_phone = nn.Linear(256, hidden_channels)
+        self.lrelu = nn.LeakyReLU(0.1, inplace=True)
+        if f0 == True:
+            self.emb_pitch = nn.Embedding(256, hidden_channels)  # pitch 256
+        self.encoder = attentions.Encoder(
+            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
+
+    def forward(self, phone, pitch, lengths):
+        if pitch == None:
+            x = self.emb_phone(phone)
+        else:
+            x = self.emb_phone(phone) + self.emb_pitch(pitch)
+        x = x * math.sqrt(self.hidden_channels)  # [b, t, h]
+        x = self.lrelu(x)
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.encoder(x * x_mask, x_mask)
+        x = self.proj(x) * x_mask
+        return x, x_mask
+
+
+class ResidualCouplingBlock(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        n_flows=4,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+        for i in range(n_flows):
+            self.flows.append(
+                modules.ResidualCouplingLayer(
+                    channels,
+                    hidden_channels,
+                    kernel_size,
+                    dilation_rate,
+                    n_layers,
+                    gin_channels=gin_channels,
+                    mean_only=True,
+                )
+            )
+            self.flows.append(modules.Flip())
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x = flow(x, x_mask, g=g, reverse=reverse)
+        return x
+
+    def remove_weight_norm(self):
+        for i in range(self.n_flows):
+            self.flows[i * 2].remove_weight_norm()
+
+
+class PosteriorEncoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+
+        self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
+        self.enc = modules.WN(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            gin_channels=gin_channels,
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x, x_lengths, g=None):
+        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.pre(x) * x_mask
+        x = self.enc(x, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
+        return z, m, logs, x_mask
+
+    def remove_weight_norm(self):
+        self.enc.remove_weight_norm()
+
+
+class Generator(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels=0,
+    ):
+        super(Generator, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+    def forward(self, x, g=None):
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+
+class SineGen(torch.nn.Module):
+    """Definition of sine generator
+    SineGen(samp_rate, harmonic_num = 0,
+            sine_amp = 0.1, noise_std = 0.003,
+            voiced_threshold = 0,
+            flag_for_pulse=False)
+    samp_rate: sampling rate in Hz
+    harmonic_num: number of harmonic overtones (default 0)
+    sine_amp: amplitude of sine-wavefrom (default 0.1)
+    noise_std: std of Gaussian noise (default 0.003)
+    voiced_thoreshold: F0 threshold for U/V classification (default 0)
+    flag_for_pulse: this SinGen is used inside PulseGen (default False)
+    Note: when flag_for_pulse is True, the first time step of a voiced
+        segment is always sin(np.pi) or cos(0)
+    """
+
+    def __init__(
+        self,
+        samp_rate,
+        harmonic_num=0,
+        sine_amp=0.1,
+        noise_std=0.003,
+        voiced_threshold=0,
+        flag_for_pulse=False,
+    ):
+        super(SineGen, self).__init__()
+        self.sine_amp = sine_amp
+        self.noise_std = noise_std
+        self.harmonic_num = harmonic_num
+        self.dim = self.harmonic_num + 1
+        self.sampling_rate = samp_rate
+        self.voiced_threshold = voiced_threshold
+
+    def _f02uv(self, f0):
+        # generate uv signal
+        uv = torch.ones_like(f0)
+        uv = uv * (f0 > self.voiced_threshold)
+        return uv
+
+    def forward(self, f0, upp):
+        """sine_tensor, uv = forward(f0)
+        input F0: tensor(batchsize=1, length, dim=1)
+                  f0 for unvoiced steps should be 0
+        output sine_tensor: tensor(batchsize=1, length, dim)
+        output uv: tensor(batchsize=1, length, 1)
+        """
+        with torch.no_grad():
+            f0 = f0[:, None].transpose(1, 2)
+            f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, device=f0.device)
+            # fundamental component
+            f0_buf[:, :, 0] = f0[:, :, 0]
+            for idx in np.arange(self.harmonic_num):
+                f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (
+                    idx + 2
+                )  # idx + 2: the (idx+1)-th overtone, (idx+2)-th harmonic
+            rad_values = (f0_buf / self.sampling_rate) % 1  ###%1意味着n_har的乘积无法后处理优化
+            rand_ini = torch.rand(
+                f0_buf.shape[0], f0_buf.shape[2], device=f0_buf.device
+            )
+            rand_ini[:, 0] = 0
+            rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
+            tmp_over_one = torch.cumsum(rad_values, 1)  # % 1  #####%1意味着后面的cumsum无法再优化
+            tmp_over_one *= upp
+            tmp_over_one = F.interpolate(
+                tmp_over_one.transpose(2, 1),
+                scale_factor=upp,
+                mode="linear",
+                align_corners=True,
+            ).transpose(2, 1)
+            rad_values = F.interpolate(
+                rad_values.transpose(2, 1), scale_factor=upp, mode="nearest"
+            ).transpose(
+                2, 1
+            )  #######
+            tmp_over_one %= 1
+            tmp_over_one_idx = (tmp_over_one[:, 1:, :] - tmp_over_one[:, :-1, :]) < 0
+            cumsum_shift = torch.zeros_like(rad_values)
+            cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
+            sine_waves = torch.sin(
+                torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi
+            )
+            sine_waves = sine_waves * self.sine_amp
+            uv = self._f02uv(f0)
+            uv = F.interpolate(
+                uv.transpose(2, 1), scale_factor=upp, mode="nearest"
+            ).transpose(2, 1)
+            noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+            noise = noise_amp * torch.randn_like(sine_waves)
+            sine_waves = sine_waves * uv + noise
+        return sine_waves, uv, noise
+
+
+class SourceModuleHnNSF(torch.nn.Module):
+    """SourceModule for hn-nsf
+    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0)
+    sampling_rate: sampling_rate in Hz
+    harmonic_num: number of harmonic above F0 (default: 0)
+    sine_amp: amplitude of sine source signal (default: 0.1)
+    add_noise_std: std of additive Gaussian noise (default: 0.003)
+        note that amplitude of noise in unvoiced is decided
+        by sine_amp
+    voiced_threshold: threhold to set U/V given F0 (default: 0)
+    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+    F0_sampled (batchsize, length, 1)
+    Sine_source (batchsize, length, 1)
+    noise_source (batchsize, length 1)
+    uv (batchsize, length, 1)
+    """
+
+    def __init__(
+        self,
+        sampling_rate,
+        harmonic_num=0,
+        sine_amp=0.1,
+        add_noise_std=0.003,
+        voiced_threshod=0,
+        is_half=True,
+    ):
+        super(SourceModuleHnNSF, self).__init__()
+
+        self.sine_amp = sine_amp
+        self.noise_std = add_noise_std
+        self.is_half = is_half
+        # to produce sine waveforms
+        self.l_sin_gen = SineGen(
+            sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod
+        )
+
+        # to merge source harmonics into a single excitation
+        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
+        self.l_tanh = torch.nn.Tanh()
+
+    def forward(self, x, upp=None):
+        sine_wavs, uv, _ = self.l_sin_gen(x, upp)
+        if self.is_half:
+            sine_wavs = sine_wavs.half()
+        sine_merge = self.l_tanh(self.l_linear(sine_wavs))
+        return sine_merge, None, None  # noise, uv
+
+
+class GeneratorNSF(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels,
+        sr,
+        is_half=False,
+    ):
+        super(GeneratorNSF, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+
+        self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates))
+        self.m_source = SourceModuleHnNSF(
+            sampling_rate=sr, harmonic_num=0, is_half=is_half
+        )
+        self.noise_convs = nn.ModuleList()
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            c_cur = upsample_initial_channel // (2 ** (i + 1))
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+            if i + 1 < len(upsample_rates):
+                stride_f0 = np.prod(upsample_rates[i + 1 :])
+                self.noise_convs.append(
+                    Conv1d(
+                        1,
+                        c_cur,
+                        kernel_size=stride_f0 * 2,
+                        stride=stride_f0,
+                        padding=stride_f0 // 2,
+                    )
+                )
+            else:
+                self.noise_convs.append(Conv1d(1, c_cur, kernel_size=1))
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+        self.upp = np.prod(upsample_rates)
+
+    def forward(self, x, f0, g=None):
+        har_source, noi_source, uv = self.m_source(f0, self.upp)
+        har_source = har_source.transpose(1, 2)
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+            x_source = self.noise_convs[i](har_source)
+            x = x + x_source
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+
+sr2sr = {
+    "32k": 32000,
+    "40k": 40000,
+    "48k": 48000,
+}
+
+
+class SynthesizerTrnMs256NSFsidM(nn.Module):
+    def __init__(
+        self,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        spk_embed_dim,
+        gin_channels,
+        sr,
+        **kwargs
+    ):
+        super().__init__()
+        if type(sr) == type("strr"):
+            sr = sr2sr[sr]
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.gin_channels = gin_channels
+        # self.hop_length = hop_length#
+        self.spk_embed_dim = spk_embed_dim
+        self.enc_p = TextEncoder256(
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+        )
+        self.dec = GeneratorNSF(
+            inter_channels,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            gin_channels=gin_channels,
+            sr=sr,
+            is_half=kwargs["is_half"],
+        )
+        self.enc_q = PosteriorEncoder(
+            spec_channels,
+            inter_channels,
+            hidden_channels,
+            5,
+            1,
+            16,
+            gin_channels=gin_channels,
+        )
+        self.flow = ResidualCouplingBlock(
+            inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels
+        )
+        self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
+        print("gin_channels:", gin_channels, "self.spk_embed_dim:", self.spk_embed_dim)
+
+    def remove_weight_norm(self):
+        self.dec.remove_weight_norm()
+        self.flow.remove_weight_norm()
+        self.enc_q.remove_weight_norm()
+
+    def forward(self, phone, phone_lengths, pitch, nsff0, sid, rnd, max_len=None):
+        g = self.emb_g(sid).unsqueeze(-1)
+        m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+        z_p = (m_p + torch.exp(logs_p) * rnd) * x_mask
+        z = self.flow(z_p, x_mask, g=g, reverse=True)
+        o = self.dec((z * x_mask)[:, :, :max_len], nsff0, g=g)
+        return o
+
+
+class SynthesizerTrnMs256NSFsid_sim(nn.Module):
+    """
+    Synthesizer for Training
+    """
+
+    def __init__(
+        self,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        spk_embed_dim,
+        # hop_length,
+        gin_channels=0,
+        use_sdp=True,
+        **kwargs
+    ):
+        super().__init__()
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.gin_channels = gin_channels
+        # self.hop_length = hop_length#
+        self.spk_embed_dim = spk_embed_dim
+        self.enc_p = TextEncoder256Sim(
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+        )
+        self.dec = GeneratorNSF(
+            inter_channels,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            gin_channels=gin_channels,
+            is_half=kwargs["is_half"],
+        )
+
+        self.flow = ResidualCouplingBlock(
+            inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels
+        )
+        self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
+        print("gin_channels:", gin_channels, "self.spk_embed_dim:", self.spk_embed_dim)
+
+    def remove_weight_norm(self):
+        self.dec.remove_weight_norm()
+        self.flow.remove_weight_norm()
+        self.enc_q.remove_weight_norm()
+
+    def forward(
+        self, phone, phone_lengths, pitch, pitchf, ds, max_len=None
+    ):  # y是spec不需要了现在
+        g = self.emb_g(ds.unsqueeze(0)).unsqueeze(-1)  # [b, 256, 1]##1是t，广播的
+        x, x_mask = self.enc_p(phone, pitch, phone_lengths)
+        x = self.flow(x, x_mask, g=g, reverse=True)
+        o = self.dec((x * x_mask)[:, :, :max_len], pitchf, g=g)
+        return o
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminator, self).__init__()
+        periods = [2, 3, 5, 7, 11, 17]
+        # periods = [3, 5, 7, 11, 17, 23, 37]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [
+            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
+        ]
+        self.discriminators = nn.ModuleList(discs)
+
+    def forward(self, y, y_hat):
+        y_d_rs = []  #
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            # for j in range(len(fmap_r)):
+            #     print(i,j,y.shape,y_hat.shape,fmap_r[j].shape,fmap_g[j].shape)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class DiscriminatorS(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(DiscriminatorS, self).__init__()
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(Conv1d(1, 16, 15, 1, padding=7)),
+                norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
+                norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
+                norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
+                norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
+                norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
+            ]
+        )
+        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
+
+    def forward(self, x):
+        fmap = []
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class DiscriminatorP(torch.nn.Module):
+    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
+        super(DiscriminatorP, self).__init__()
+        self.period = period
+        self.use_spectral_norm = use_spectral_norm
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(
+                    Conv2d(
+                        1,
+                        32,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        32,
+                        128,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        128,
+                        512,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        512,
+                        1024,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        1024,
+                        1024,
+                        (kernel_size, 1),
+                        1,
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+            ]
+        )
+        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+
+    def forward(self, x):
+        fmap = []
+
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0:  # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), "reflect")
+            t = t + n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap

src/infer_pack/modules.py

@@ -0,0 +1,522 @@
+import copy
+import math
+import numpy as np
+import scipy
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
+from torch.nn.utils import weight_norm, remove_weight_norm
+
+from infer_pack import commons
+from infer_pack.commons import init_weights, get_padding
+from infer_pack.transforms import piecewise_rational_quadratic_transform
+
+
+LRELU_SLOPE = 0.1
+
+
+class LayerNorm(nn.Module):
+    def __init__(self, channels, eps=1e-5):
+        super().__init__()
+        self.channels = channels
+        self.eps = eps
+
+        self.gamma = nn.Parameter(torch.ones(channels))
+        self.beta = nn.Parameter(torch.zeros(channels))
+
+    def forward(self, x):
+        x = x.transpose(1, -1)
+        x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+        return x.transpose(1, -1)
+
+
+class ConvReluNorm(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        hidden_channels,
+        out_channels,
+        kernel_size,
+        n_layers,
+        p_dropout,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.hidden_channels = hidden_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.p_dropout = p_dropout
+        assert n_layers > 1, "Number of layers should be larger than 0."
+
+        self.conv_layers = nn.ModuleList()
+        self.norm_layers = nn.ModuleList()
+        self.conv_layers.append(
+            nn.Conv1d(
+                in_channels, hidden_channels, kernel_size, padding=kernel_size // 2
+            )
+        )
+        self.norm_layers.append(LayerNorm(hidden_channels))
+        self.relu_drop = nn.Sequential(nn.ReLU(), nn.Dropout(p_dropout))
+        for _ in range(n_layers - 1):
+            self.conv_layers.append(
+                nn.Conv1d(
+                    hidden_channels,
+                    hidden_channels,
+                    kernel_size,
+                    padding=kernel_size // 2,
+                )
+            )
+            self.norm_layers.append(LayerNorm(hidden_channels))
+        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(self, x, x_mask):
+        x_org = x
+        for i in range(self.n_layers):
+            x = self.conv_layers[i](x * x_mask)
+            x = self.norm_layers[i](x)
+            x = self.relu_drop(x)
+        x = x_org + self.proj(x)
+        return x * x_mask
+
+
+class DDSConv(nn.Module):
+    """
+    Dialted and Depth-Separable Convolution
+    """
+
+    def __init__(self, channels, kernel_size, n_layers, p_dropout=0.0):
+        super().__init__()
+        self.channels = channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.p_dropout = p_dropout
+
+        self.drop = nn.Dropout(p_dropout)
+        self.convs_sep = nn.ModuleList()
+        self.convs_1x1 = nn.ModuleList()
+        self.norms_1 = nn.ModuleList()
+        self.norms_2 = nn.ModuleList()
+        for i in range(n_layers):
+            dilation = kernel_size**i
+            padding = (kernel_size * dilation - dilation) // 2
+            self.convs_sep.append(
+                nn.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    groups=channels,
+                    dilation=dilation,
+                    padding=padding,
+                )
+            )
+            self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
+            self.norms_1.append(LayerNorm(channels))
+            self.norms_2.append(LayerNorm(channels))
+
+    def forward(self, x, x_mask, g=None):
+        if g is not None:
+            x = x + g
+        for i in range(self.n_layers):
+            y = self.convs_sep[i](x * x_mask)
+            y = self.norms_1[i](y)
+            y = F.gelu(y)
+            y = self.convs_1x1[i](y)
+            y = self.norms_2[i](y)
+            y = F.gelu(y)
+            y = self.drop(y)
+            x = x + y
+        return x * x_mask
+
+
+class WN(torch.nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+        p_dropout=0,
+    ):
+        super(WN, self).__init__()
+        assert kernel_size % 2 == 1
+        self.hidden_channels = hidden_channels
+        self.kernel_size = (kernel_size,)
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+        self.p_dropout = p_dropout
+
+        self.in_layers = torch.nn.ModuleList()
+        self.res_skip_layers = torch.nn.ModuleList()
+        self.drop = nn.Dropout(p_dropout)
+
+        if gin_channels != 0:
+            cond_layer = torch.nn.Conv1d(
+                gin_channels, 2 * hidden_channels * n_layers, 1
+            )
+            self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name="weight")
+
+        for i in range(n_layers):
+            dilation = dilation_rate**i
+            padding = int((kernel_size * dilation - dilation) / 2)
+            in_layer = torch.nn.Conv1d(
+                hidden_channels,
+                2 * hidden_channels,
+                kernel_size,
+                dilation=dilation,
+                padding=padding,
+            )
+            in_layer = torch.nn.utils.weight_norm(in_layer, name="weight")
+            self.in_layers.append(in_layer)
+
+            # last one is not necessary
+            if i < n_layers - 1:
+                res_skip_channels = 2 * hidden_channels
+            else:
+                res_skip_channels = hidden_channels
+
+            res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
+            res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name="weight")
+            self.res_skip_layers.append(res_skip_layer)
+
+    def forward(self, x, x_mask, g=None, **kwargs):
+        output = torch.zeros_like(x)
+        n_channels_tensor = torch.IntTensor([self.hidden_channels])
+
+        if g is not None:
+            g = self.cond_layer(g)
+
+        for i in range(self.n_layers):
+            x_in = self.in_layers[i](x)
+            if g is not None:
+                cond_offset = i * 2 * self.hidden_channels
+                g_l = g[:, cond_offset : cond_offset + 2 * self.hidden_channels, :]
+            else:
+                g_l = torch.zeros_like(x_in)
+
+            acts = commons.fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor)
+            acts = self.drop(acts)
+
+            res_skip_acts = self.res_skip_layers[i](acts)
+            if i < self.n_layers - 1:
+                res_acts = res_skip_acts[:, : self.hidden_channels, :]
+                x = (x + res_acts) * x_mask
+                output = output + res_skip_acts[:, self.hidden_channels :, :]
+            else:
+                output = output + res_skip_acts
+        return output * x_mask
+
+    def remove_weight_norm(self):
+        if self.gin_channels != 0:
+            torch.nn.utils.remove_weight_norm(self.cond_layer)
+        for l in self.in_layers:
+            torch.nn.utils.remove_weight_norm(l)
+        for l in self.res_skip_layers:
+            torch.nn.utils.remove_weight_norm(l)
+
+
+class ResBlock1(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__()
+        self.convs1 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[2],
+                        padding=get_padding(kernel_size, dilation[2]),
+                    )
+                ),
+            ]
+        )
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+            ]
+        )
+        self.convs2.apply(init_weights)
+
+    def forward(self, x, x_mask=None):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c2(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+
+class ResBlock2(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
+        super(ResBlock2, self).__init__()
+        self.convs = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+            ]
+        )
+        self.convs.apply(init_weights)
+
+    def forward(self, x, x_mask=None):
+        for c in self.convs:
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_weight_norm(l)
+
+
+class Log(nn.Module):
+    def forward(self, x, x_mask, reverse=False, **kwargs):
+        if not reverse:
+            y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
+            logdet = torch.sum(-y, [1, 2])
+            return y, logdet
+        else:
+            x = torch.exp(x) * x_mask
+            return x
+
+
+class Flip(nn.Module):
+    def forward(self, x, *args, reverse=False, **kwargs):
+        x = torch.flip(x, [1])
+        if not reverse:
+            logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
+            return x, logdet
+        else:
+            return x
+
+
+class ElementwiseAffine(nn.Module):
+    def __init__(self, channels):
+        super().__init__()
+        self.channels = channels
+        self.m = nn.Parameter(torch.zeros(channels, 1))
+        self.logs = nn.Parameter(torch.zeros(channels, 1))
+
+    def forward(self, x, x_mask, reverse=False, **kwargs):
+        if not reverse:
+            y = self.m + torch.exp(self.logs) * x
+            y = y * x_mask
+            logdet = torch.sum(self.logs * x_mask, [1, 2])
+            return y, logdet
+        else:
+            x = (x - self.m) * torch.exp(-self.logs) * x_mask
+            return x
+
+
+class ResidualCouplingLayer(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        p_dropout=0,
+        gin_channels=0,
+        mean_only=False,
+    ):
+        assert channels % 2 == 0, "channels should be divisible by 2"
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.half_channels = channels // 2
+        self.mean_only = mean_only
+
+        self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
+        self.enc = WN(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            p_dropout=p_dropout,
+            gin_channels=gin_channels,
+        )
+        self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
+        self.post.weight.data.zero_()
+        self.post.bias.data.zero_()
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0) * x_mask
+        h = self.enc(h, x_mask, g=g)
+        stats = self.post(h) * x_mask
+        if not self.mean_only:
+            m, logs = torch.split(stats, [self.half_channels] * 2, 1)
+        else:
+            m = stats
+            logs = torch.zeros_like(m)
+
+        if not reverse:
+            x1 = m + x1 * torch.exp(logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            logdet = torch.sum(logs, [1, 2])
+            return x, logdet
+        else:
+            x1 = (x1 - m) * torch.exp(-logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            return x
+
+    def remove_weight_norm(self):
+        self.enc.remove_weight_norm()
+
+
+class ConvFlow(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        filter_channels,
+        kernel_size,
+        n_layers,
+        num_bins=10,
+        tail_bound=5.0,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.num_bins = num_bins
+        self.tail_bound = tail_bound
+        self.half_channels = in_channels // 2
+
+        self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
+        self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.0)
+        self.proj = nn.Conv1d(
+            filter_channels, self.half_channels * (num_bins * 3 - 1), 1
+        )
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0)
+        h = self.convs(h, x_mask, g=g)
+        h = self.proj(h) * x_mask
+
+        b, c, t = x0.shape
+        h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2)  # [b, cx?, t] -> [b, c, t, ?]
+
+        unnormalized_widths = h[..., : self.num_bins] / math.sqrt(self.filter_channels)
+        unnormalized_heights = h[..., self.num_bins : 2 * self.num_bins] / math.sqrt(
+            self.filter_channels
+        )
+        unnormalized_derivatives = h[..., 2 * self.num_bins :]
+
+        x1, logabsdet = piecewise_rational_quadratic_transform(
+            x1,
+            unnormalized_widths,
+            unnormalized_heights,
+            unnormalized_derivatives,
+            inverse=reverse,
+            tails="linear",
+            tail_bound=self.tail_bound,
+        )
+
+        x = torch.cat([x0, x1], 1) * x_mask
+        logdet = torch.sum(logabsdet * x_mask, [1, 2])
+        if not reverse:
+            return x, logdet
+        else:
+            return x

src/infer_pack/predictor/FCPE.py

@@ -0,0 +1,1036 @@
+from typing import Union
+
+import torch.nn.functional as F
+import numpy as np
+import torch
+import torch.nn as nn
+from torch.nn.utils.parametrizations import weight_norm
+from torchaudio.transforms import Resample
+import os
+import librosa
+import soundfile as sf
+import torch.utils.data
+from librosa.filters import mel as librosa_mel_fn
+import math
+from functools import partial
+
+from einops import rearrange, repeat
+from local_attention import LocalAttention
+from torch import nn
+
+os.environ["LRU_CACHE_CAPACITY"] = "3"
+
+
+def load_wav_to_torch(full_path, target_sr=None, return_empty_on_exception=False):
+    sampling_rate = None
+    try:
+        data, sampling_rate = sf.read(full_path, always_2d=True)  # than soundfile.
+    except Exception as error:
+        print(f"'{full_path}' failed to load with {error}")
+        if return_empty_on_exception:
+            return [], sampling_rate or target_sr or 48000
+        else:
+            raise Exception(error)
+
+    if len(data.shape) > 1:
+        data = data[:, 0]
+        assert (
+            len(data) > 2
+        )  # check duration of audio file is > 2 samples (because otherwise the slice operation was on the wrong dimension)
+
+    if np.issubdtype(data.dtype, np.integer):  # if audio data is type int
+        max_mag = -np.iinfo(
+            data.dtype
+        ).min  # maximum magnitude = min possible value of intXX
+    else:  # if audio data is type fp32
+        max_mag = max(np.amax(data), -np.amin(data))
+        max_mag = (
+            (2**31) + 1
+            if max_mag > (2**15)
+            else ((2**15) + 1 if max_mag > 1.01 else 1.0)
+        )  # data should be either 16-bit INT, 32-bit INT or [-1 to 1] float32
+
+    data = torch.FloatTensor(data.astype(np.float32)) / max_mag
+
+    if (
+        torch.isinf(data) | torch.isnan(data)
+    ).any() and return_empty_on_exception:  # resample will crash with inf/NaN inputs. return_empty_on_exception will return empty arr instead of except
+        return [], sampling_rate or target_sr or 48000
+    if target_sr is not None and sampling_rate != target_sr:
+        data = torch.from_numpy(
+            librosa.core.resample(
+                data.numpy(), orig_sr=sampling_rate, target_sr=target_sr
+            )
+        )
+        sampling_rate = target_sr
+
+    return data, sampling_rate
+
+
+def dynamic_range_compression(x, C=1, clip_val=1e-5):
+    return np.log(np.clip(x, a_min=clip_val, a_max=None) * C)
+
+
+def dynamic_range_decompression(x, C=1):
+    return np.exp(x) / C
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression_torch(x, C=1):
+    return torch.exp(x) / C
+
+
+class STFT:
+    def __init__(
+        self,
+        sr=22050,
+        n_mels=80,
+        n_fft=1024,
+        win_size=1024,
+        hop_length=256,
+        fmin=20,
+        fmax=11025,
+        clip_val=1e-5,
+    ):
+        self.target_sr = sr
+
+        self.n_mels = n_mels
+        self.n_fft = n_fft
+        self.win_size = win_size
+        self.hop_length = hop_length
+        self.fmin = fmin
+        self.fmax = fmax
+        self.clip_val = clip_val
+        self.mel_basis = {}
+        self.hann_window = {}
+
+    def get_mel(self, y, keyshift=0, speed=1, center=False, train=False):
+        sampling_rate = self.target_sr
+        n_mels = self.n_mels
+        n_fft = self.n_fft
+        win_size = self.win_size
+        hop_length = self.hop_length
+        fmin = self.fmin
+        fmax = self.fmax
+        clip_val = self.clip_val
+
+        factor = 2 ** (keyshift / 12)
+        n_fft_new = int(np.round(n_fft * factor))
+        win_size_new = int(np.round(win_size * factor))
+        hop_length_new = int(np.round(hop_length * speed))
+        if not train:
+            mel_basis = self.mel_basis
+            hann_window = self.hann_window
+        else:
+            mel_basis = {}
+            hann_window = {}
+
+        mel_basis_key = str(fmax) + "_" + str(y.device)
+        if mel_basis_key not in mel_basis:
+            mel = librosa_mel_fn(
+                sr=sampling_rate, n_fft=n_fft, n_mels=n_mels, fmin=fmin, fmax=fmax
+            )
+            mel_basis[mel_basis_key] = torch.from_numpy(mel).float().to(y.device)
+
+        keyshift_key = str(keyshift) + "_" + str(y.device)
+        if keyshift_key not in hann_window:
+            hann_window[keyshift_key] = torch.hann_window(win_size_new).to(y.device)
+
+        pad_left = (win_size_new - hop_length_new) // 2
+        pad_right = max(
+            (win_size_new - hop_length_new + 1) // 2,
+            win_size_new - y.size(-1) - pad_left,
+        )
+        if pad_right < y.size(-1):
+            mode = "reflect"
+        else:
+            mode = "constant"
+        y = torch.nn.functional.pad(y.unsqueeze(1), (pad_left, pad_right), mode=mode)
+        y = y.squeeze(1)
+
+        spec = torch.stft(
+            y,
+            n_fft_new,
+            hop_length=hop_length_new,
+            win_length=win_size_new,
+            window=hann_window[keyshift_key],
+            center=center,
+            pad_mode="reflect",
+            normalized=False,
+            onesided=True,
+            return_complex=True,
+        )
+        spec = torch.sqrt(spec.real.pow(2) + spec.imag.pow(2) + (1e-9))
+        if keyshift != 0:
+            size = n_fft // 2 + 1
+            resize = spec.size(1)
+            if resize < size:
+                spec = F.pad(spec, (0, 0, 0, size - resize))
+            spec = spec[:, :size, :] * win_size / win_size_new
+        spec = torch.matmul(mel_basis[mel_basis_key], spec)
+        spec = dynamic_range_compression_torch(spec, clip_val=clip_val)
+        return spec
+
+    def __call__(self, audiopath):
+        audio, sr = load_wav_to_torch(audiopath, target_sr=self.target_sr)
+        spect = self.get_mel(audio.unsqueeze(0)).squeeze(0)
+        return spect
+
+
+stft = STFT()
+
+# import fast_transformers.causal_product.causal_product_cuda
+
+
+def softmax_kernel(
+    data, *, projection_matrix, is_query, normalize_data=True, eps=1e-4, device=None
+):
+    b, h, *_ = data.shape
+    # (batch size, head, length, model_dim)
+
+    # normalize model dim
+    data_normalizer = (data.shape[-1] ** -0.25) if normalize_data else 1.0
+
+    # what is ration?, projection_matrix.shape[0] --> 266
+
+    ratio = projection_matrix.shape[0] ** -0.5
+
+    projection = repeat(projection_matrix, "j d -> b h j d", b=b, h=h)
+    projection = projection.type_as(data)
+
+    # data_dash = w^T x
+    data_dash = torch.einsum("...id,...jd->...ij", (data_normalizer * data), projection)
+
+    # diag_data = D**2
+    diag_data = data**2
+    diag_data = torch.sum(diag_data, dim=-1)
+    diag_data = (diag_data / 2.0) * (data_normalizer**2)
+    diag_data = diag_data.unsqueeze(dim=-1)
+
+    if is_query:
+        data_dash = ratio * (
+            torch.exp(
+                data_dash
+                - diag_data
+                - torch.max(data_dash, dim=-1, keepdim=True).values
+            )
+            + eps
+        )
+    else:
+        data_dash = ratio * (
+            torch.exp(data_dash - diag_data + eps)
+        )  # - torch.max(data_dash)) + eps)
+
+    return data_dash.type_as(data)
+
+
+def orthogonal_matrix_chunk(cols, qr_uniform_q=False, device=None):
+    unstructured_block = torch.randn((cols, cols), device=device)
+    q, r = torch.linalg.qr(unstructured_block.cpu(), mode="reduced")
+    q, r = map(lambda t: t.to(device), (q, r))
+
+    # proposed by @Parskatt
+    # to make sure Q is uniform https://arxiv.org/pdf/math-ph/0609050.pdf
+    if qr_uniform_q:
+        d = torch.diag(r, 0)
+        q *= d.sign()
+    return q.t()
+
+
+def exists(val):
+    return val is not None
+
+
+def empty(tensor):
+    return tensor.numel() == 0
+
+
+def default(val, d):
+    return val if exists(val) else d
+
+
+def cast_tuple(val):
+    return (val,) if not isinstance(val, tuple) else val
+
+
+class PCmer(nn.Module):
+    """The encoder that is used in the Transformer model."""
+
+    def __init__(
+        self,
+        num_layers,
+        num_heads,
+        dim_model,
+        dim_keys,
+        dim_values,
+        residual_dropout,
+        attention_dropout,
+    ):
+        super().__init__()
+        self.num_layers = num_layers
+        self.num_heads = num_heads
+        self.dim_model = dim_model
+        self.dim_values = dim_values
+        self.dim_keys = dim_keys
+        self.residual_dropout = residual_dropout
+        self.attention_dropout = attention_dropout
+
+        self._layers = nn.ModuleList([_EncoderLayer(self) for _ in range(num_layers)])
+
+    #  METHODS  ########################################################################################################
+
+    def forward(self, phone, mask=None):
+
+        # apply all layers to the input
+        for i, layer in enumerate(self._layers):
+            phone = layer(phone, mask)
+        # provide the final sequence
+        return phone
+
+
+# ==================================================================================================================== #
+#  CLASS  _ E N C O D E R  L A Y E R                                                                                   #
+# ==================================================================================================================== #
+
+
+class _EncoderLayer(nn.Module):
+    """One layer of the encoder.
+
+    Attributes:
+        attn: (:class:`mha.MultiHeadAttention`): The attention mechanism that is used to read the input sequence.
+        feed_forward (:class:`ffl.FeedForwardLayer`): The feed-forward layer on top of the attention mechanism.
+    """
+
+    def __init__(self, parent: PCmer):
+        """Creates a new instance of ``_EncoderLayer``.
+
+        Args:
+            parent (Encoder): The encoder that the layers is created for.
+        """
+        super().__init__()
+
+        self.conformer = ConformerConvModule(parent.dim_model)
+        self.norm = nn.LayerNorm(parent.dim_model)
+        self.dropout = nn.Dropout(parent.residual_dropout)
+
+        # selfatt -> fastatt: performer!
+        self.attn = SelfAttention(
+            dim=parent.dim_model, heads=parent.num_heads, causal=False
+        )
+
+    #  METHODS  ########################################################################################################
+
+    def forward(self, phone, mask=None):
+
+        # compute attention sub-layer
+        phone = phone + (self.attn(self.norm(phone), mask=mask))
+
+        phone = phone + (self.conformer(phone))
+
+        return phone
+
+
+def calc_same_padding(kernel_size):
+    pad = kernel_size // 2
+    return (pad, pad - (kernel_size + 1) % 2)
+
+
+# helper classes
+
+
+class Swish(nn.Module):
+    def forward(self, x):
+        return x * x.sigmoid()
+
+
+class Transpose(nn.Module):
+    def __init__(self, dims):
+        super().__init__()
+        assert len(dims) == 2, "dims must be a tuple of two dimensions"
+        self.dims = dims
+
+    def forward(self, x):
+        return x.transpose(*self.dims)
+
+
+class GLU(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, x):
+        out, gate = x.chunk(2, dim=self.dim)
+        return out * gate.sigmoid()
+
+
+class DepthWiseConv1d(nn.Module):
+    def __init__(self, chan_in, chan_out, kernel_size, padding):
+        super().__init__()
+        self.padding = padding
+        self.conv = nn.Conv1d(chan_in, chan_out, kernel_size, groups=chan_in)
+
+    def forward(self, x):
+        x = F.pad(x, self.padding)
+        return self.conv(x)
+
+
+class ConformerConvModule(nn.Module):
+    def __init__(
+        self, dim, causal=False, expansion_factor=2, kernel_size=31, dropout=0.0
+    ):
+        super().__init__()
+
+        inner_dim = dim * expansion_factor
+        padding = calc_same_padding(kernel_size) if not causal else (kernel_size - 1, 0)
+
+        self.net = nn.Sequential(
+            nn.LayerNorm(dim),
+            Transpose((1, 2)),
+            nn.Conv1d(dim, inner_dim * 2, 1),
+            GLU(dim=1),
+            DepthWiseConv1d(
+                inner_dim, inner_dim, kernel_size=kernel_size, padding=padding
+            ),
+            # nn.BatchNorm1d(inner_dim) if not causal else nn.Identity(),
+            Swish(),
+            nn.Conv1d(inner_dim, dim, 1),
+            Transpose((1, 2)),
+            nn.Dropout(dropout),
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+def linear_attention(q, k, v):
+    if v is None:
+        out = torch.einsum("...ed,...nd->...ne", k, q)
+        return out
+
+    else:
+        k_cumsum = k.sum(dim=-2)
+        # k_cumsum = k.sum(dim = -2)
+        D_inv = 1.0 / (torch.einsum("...nd,...d->...n", q, k_cumsum.type_as(q)) + 1e-8)
+
+        context = torch.einsum("...nd,...ne->...de", k, v)
+        out = torch.einsum("...de,...nd,...n->...ne", context, q, D_inv)
+        return out
+
+
+def gaussian_orthogonal_random_matrix(
+    nb_rows, nb_columns, scaling=0, qr_uniform_q=False, device=None
+):
+    nb_full_blocks = int(nb_rows / nb_columns)
+    block_list = []
+
+    for _ in range(nb_full_blocks):
+        q = orthogonal_matrix_chunk(
+            nb_columns, qr_uniform_q=qr_uniform_q, device=device
+        )
+        block_list.append(q)
+
+    remaining_rows = nb_rows - nb_full_blocks * nb_columns
+    if remaining_rows > 0:
+        q = orthogonal_matrix_chunk(
+            nb_columns, qr_uniform_q=qr_uniform_q, device=device
+        )
+
+        block_list.append(q[:remaining_rows])
+
+    final_matrix = torch.cat(block_list)
+
+    if scaling == 0:
+        multiplier = torch.randn((nb_rows, nb_columns), device=device).norm(dim=1)
+    elif scaling == 1:
+        multiplier = math.sqrt((float(nb_columns))) * torch.ones(
+            (nb_rows,), device=device
+        )
+    else:
+        raise ValueError(f"Invalid scaling {scaling}")
+
+    return torch.diag(multiplier) @ final_matrix
+
+
+class FastAttention(nn.Module):
+    def __init__(
+        self,
+        dim_heads,
+        nb_features=None,
+        ortho_scaling=0,
+        causal=False,
+        generalized_attention=False,
+        kernel_fn=nn.ReLU(),
+        qr_uniform_q=False,
+        no_projection=False,
+    ):
+        super().__init__()
+        nb_features = default(nb_features, int(dim_heads * math.log(dim_heads)))
+
+        self.dim_heads = dim_heads
+        self.nb_features = nb_features
+        self.ortho_scaling = ortho_scaling
+
+        self.create_projection = partial(
+            gaussian_orthogonal_random_matrix,
+            nb_rows=self.nb_features,
+            nb_columns=dim_heads,
+            scaling=ortho_scaling,
+            qr_uniform_q=qr_uniform_q,
+        )
+        projection_matrix = self.create_projection()
+        self.register_buffer("projection_matrix", projection_matrix)
+
+        self.generalized_attention = generalized_attention
+        self.kernel_fn = kernel_fn
+
+        # if this is turned on, no projection will be used
+        # queries and keys will be softmax-ed as in the original efficient attention paper
+        self.no_projection = no_projection
+
+        self.causal = causal
+
+    @torch.no_grad()
+    def redraw_projection_matrix(self):
+        projections = self.create_projection()
+        self.projection_matrix.copy_(projections)
+        del projections
+
+    def forward(self, q, k, v):
+        device = q.device
+
+        if self.no_projection:
+            q = q.softmax(dim=-1)
+            k = torch.exp(k) if self.causal else k.softmax(dim=-2)
+        else:
+            create_kernel = partial(
+                softmax_kernel, projection_matrix=self.projection_matrix, device=device
+            )
+
+            q = create_kernel(q, is_query=True)
+            k = create_kernel(k, is_query=False)
+
+        attn_fn = linear_attention if not self.causal else self.causal_linear_fn
+        if v is None:
+            out = attn_fn(q, k, None)
+            return out
+        else:
+            out = attn_fn(q, k, v)
+            return out
+
+
+class SelfAttention(nn.Module):
+    def __init__(
+        self,
+        dim,
+        causal=False,
+        heads=8,
+        dim_head=64,
+        local_heads=0,
+        local_window_size=256,
+        nb_features=None,
+        feature_redraw_interval=1000,
+        generalized_attention=False,
+        kernel_fn=nn.ReLU(),
+        qr_uniform_q=False,
+        dropout=0.0,
+        no_projection=False,
+    ):
+        super().__init__()
+        assert dim % heads == 0, "dimension must be divisible by number of heads"
+        dim_head = default(dim_head, dim // heads)
+        inner_dim = dim_head * heads
+        self.fast_attention = FastAttention(
+            dim_head,
+            nb_features,
+            causal=causal,
+            generalized_attention=generalized_attention,
+            kernel_fn=kernel_fn,
+            qr_uniform_q=qr_uniform_q,
+            no_projection=no_projection,
+        )
+
+        self.heads = heads
+        self.global_heads = heads - local_heads
+        self.local_attn = (
+            LocalAttention(
+                window_size=local_window_size,
+                causal=causal,
+                autopad=True,
+                dropout=dropout,
+                look_forward=int(not causal),
+                rel_pos_emb_config=(dim_head, local_heads),
+            )
+            if local_heads > 0
+            else None
+        )
+
+        self.to_q = nn.Linear(dim, inner_dim)
+        self.to_k = nn.Linear(dim, inner_dim)
+        self.to_v = nn.Linear(dim, inner_dim)
+        self.to_out = nn.Linear(inner_dim, dim)
+        self.dropout = nn.Dropout(dropout)
+
+    @torch.no_grad()
+    def redraw_projection_matrix(self):
+        self.fast_attention.redraw_projection_matrix()
+
+    def forward(
+        self,
+        x,
+        context=None,
+        mask=None,
+        context_mask=None,
+        name=None,
+        inference=False,
+        **kwargs,
+    ):
+        _, _, _, h, gh = *x.shape, self.heads, self.global_heads
+
+        cross_attend = exists(context)
+
+        context = default(context, x)
+        context_mask = default(context_mask, mask) if not cross_attend else context_mask
+        q, k, v = self.to_q(x), self.to_k(context), self.to_v(context)
+
+        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q, k, v))
+        (q, lq), (k, lk), (v, lv) = map(lambda t: (t[:, :gh], t[:, gh:]), (q, k, v))
+
+        attn_outs = []
+        if not empty(q):
+            if exists(context_mask):
+                global_mask = context_mask[:, None, :, None]
+                v.masked_fill_(~global_mask, 0.0)
+            if cross_attend:
+                pass
+            else:
+                out = self.fast_attention(q, k, v)
+            attn_outs.append(out)
+
+        if not empty(lq):
+            assert (
+                not cross_attend
+            ), "local attention is not compatible with cross attention"
+            out = self.local_attn(lq, lk, lv, input_mask=mask)
+            attn_outs.append(out)
+
+        out = torch.cat(attn_outs, dim=1)
+        out = rearrange(out, "b h n d -> b n (h d)")
+        out = self.to_out(out)
+        return self.dropout(out)
+
+
+def l2_regularization(model, l2_alpha):
+    l2_loss = []
+    for module in model.modules():
+        if type(module) is nn.Conv2d:
+            l2_loss.append((module.weight**2).sum() / 2.0)
+    return l2_alpha * sum(l2_loss)
+
+
+class FCPE(nn.Module):
+    def __init__(
+        self,
+        input_channel=128,
+        out_dims=360,
+        n_layers=12,
+        n_chans=512,
+        use_siren=False,
+        use_full=False,
+        loss_mse_scale=10,
+        loss_l2_regularization=False,
+        loss_l2_regularization_scale=1,
+        loss_grad1_mse=False,
+        loss_grad1_mse_scale=1,
+        f0_max=1975.5,
+        f0_min=32.70,
+        confidence=False,
+        threshold=0.05,
+        use_input_conv=True,
+    ):
+        super().__init__()
+        if use_siren is True:
+            raise ValueError("Siren is not supported yet.")
+        if use_full is True:
+            raise ValueError("Full model is not supported yet.")
+
+        self.loss_mse_scale = loss_mse_scale if (loss_mse_scale is not None) else 10
+        self.loss_l2_regularization = (
+            loss_l2_regularization if (loss_l2_regularization is not None) else False
+        )
+        self.loss_l2_regularization_scale = (
+            loss_l2_regularization_scale
+            if (loss_l2_regularization_scale is not None)
+            else 1
+        )
+        self.loss_grad1_mse = loss_grad1_mse if (loss_grad1_mse is not None) else False
+        self.loss_grad1_mse_scale = (
+            loss_grad1_mse_scale if (loss_grad1_mse_scale is not None) else 1
+        )
+        self.f0_max = f0_max if (f0_max is not None) else 1975.5
+        self.f0_min = f0_min if (f0_min is not None) else 32.70
+        self.confidence = confidence if (confidence is not None) else False
+        self.threshold = threshold if (threshold is not None) else 0.05
+        self.use_input_conv = use_input_conv if (use_input_conv is not None) else True
+
+        self.cent_table_b = torch.Tensor(
+            np.linspace(
+                self.f0_to_cent(torch.Tensor([f0_min]))[0],
+                self.f0_to_cent(torch.Tensor([f0_max]))[0],
+                out_dims,
+            )
+        )
+        self.register_buffer("cent_table", self.cent_table_b)
+
+        # conv in stack
+        _leaky = nn.LeakyReLU()
+        self.stack = nn.Sequential(
+            nn.Conv1d(input_channel, n_chans, 3, 1, 1),
+            nn.GroupNorm(4, n_chans),
+            _leaky,
+            nn.Conv1d(n_chans, n_chans, 3, 1, 1),
+        )
+
+        # transformer
+        self.decoder = PCmer(
+            num_layers=n_layers,
+            num_heads=8,
+            dim_model=n_chans,
+            dim_keys=n_chans,
+            dim_values=n_chans,
+            residual_dropout=0.1,
+            attention_dropout=0.1,
+        )
+        self.norm = nn.LayerNorm(n_chans)
+
+        # out
+        self.n_out = out_dims
+        self.dense_out = weight_norm(nn.Linear(n_chans, self.n_out))
+
+    def forward(
+        self, mel, infer=True, gt_f0=None, return_hz_f0=False, cdecoder="local_argmax"
+    ):
+        """
+        input:
+            B x n_frames x n_unit
+        return:
+            dict of B x n_frames x feat
+        """
+        if cdecoder == "argmax":
+            self.cdecoder = self.cents_decoder
+        elif cdecoder == "local_argmax":
+            self.cdecoder = self.cents_local_decoder
+        if self.use_input_conv:
+            x = self.stack(mel.transpose(1, 2)).transpose(1, 2)
+        else:
+            x = mel
+        x = self.decoder(x)
+        x = self.norm(x)
+        x = self.dense_out(x)  # [B,N,D]
+        x = torch.sigmoid(x)
+        if not infer:
+            gt_cent_f0 = self.f0_to_cent(gt_f0)  # mel f0  #[B,N,1]
+            gt_cent_f0 = self.gaussian_blurred_cent(gt_cent_f0)  # #[B,N,out_dim]
+            loss_all = self.loss_mse_scale * F.binary_cross_entropy(
+                x, gt_cent_f0
+            )  # bce loss
+            # l2 regularization
+            if self.loss_l2_regularization:
+                loss_all = loss_all + l2_regularization(
+                    model=self, l2_alpha=self.loss_l2_regularization_scale
+                )
+            x = loss_all
+        if infer:
+            x = self.cdecoder(x)
+            x = self.cent_to_f0(x)
+            if not return_hz_f0:
+                x = (1 + x / 700).log()
+        return x
+
+    def cents_decoder(self, y, mask=True):
+        B, N, _ = y.size()
+        ci = self.cent_table[None, None, :].expand(B, N, -1)
+        rtn = torch.sum(ci * y, dim=-1, keepdim=True) / torch.sum(
+            y, dim=-1, keepdim=True
+        )  # cents: [B,N,1]
+        if mask:
+            confident = torch.max(y, dim=-1, keepdim=True)[0]
+            confident_mask = torch.ones_like(confident)
+            confident_mask[confident <= self.threshold] = float("-INF")
+            rtn = rtn * confident_mask
+        if self.confidence:
+            return rtn, confident
+        else:
+            return rtn
+
+    def cents_local_decoder(self, y, mask=True):
+        B, N, _ = y.size()
+        ci = self.cent_table[None, None, :].expand(B, N, -1)
+        confident, max_index = torch.max(y, dim=-1, keepdim=True)
+        local_argmax_index = torch.arange(0, 9).to(max_index.device) + (max_index - 4)
+        local_argmax_index[local_argmax_index < 0] = 0
+        local_argmax_index[local_argmax_index >= self.n_out] = self.n_out - 1
+        ci_l = torch.gather(ci, -1, local_argmax_index)
+        y_l = torch.gather(y, -1, local_argmax_index)
+        rtn = torch.sum(ci_l * y_l, dim=-1, keepdim=True) / torch.sum(
+            y_l, dim=-1, keepdim=True
+        )  # cents: [B,N,1]
+        if mask:
+            confident_mask = torch.ones_like(confident)
+            confident_mask[confident <= self.threshold] = float("-INF")
+            rtn = rtn * confident_mask
+        if self.confidence:
+            return rtn, confident
+        else:
+            return rtn
+
+    def cent_to_f0(self, cent):
+        return 10.0 * 2 ** (cent / 1200.0)
+
+    def f0_to_cent(self, f0):
+        return 1200.0 * torch.log2(f0 / 10.0)
+
+    def gaussian_blurred_cent(self, cents):  # cents: [B,N,1]
+        mask = (cents > 0.1) & (cents < (1200.0 * np.log2(self.f0_max / 10.0)))
+        B, N, _ = cents.size()
+        ci = self.cent_table[None, None, :].expand(B, N, -1)
+        return torch.exp(-torch.square(ci - cents) / 1250) * mask.float()
+
+
+class FCPEInfer:
+    def __init__(self, model_path, device=None, dtype=torch.float32):
+        if device is None:
+            device = "cuda" if torch.cuda.is_available() else "cpu"
+        self.device = device
+        ckpt = torch.load(model_path, map_location=torch.device(self.device))
+        self.args = DotDict(ckpt["config"])
+        self.dtype = dtype
+        model = FCPE(
+            input_channel=self.args.model.input_channel,
+            out_dims=self.args.model.out_dims,
+            n_layers=self.args.model.n_layers,
+            n_chans=self.args.model.n_chans,
+            use_siren=self.args.model.use_siren,
+            use_full=self.args.model.use_full,
+            loss_mse_scale=self.args.loss.loss_mse_scale,
+            loss_l2_regularization=self.args.loss.loss_l2_regularization,
+            loss_l2_regularization_scale=self.args.loss.loss_l2_regularization_scale,
+            loss_grad1_mse=self.args.loss.loss_grad1_mse,
+            loss_grad1_mse_scale=self.args.loss.loss_grad1_mse_scale,
+            f0_max=self.args.model.f0_max,
+            f0_min=self.args.model.f0_min,
+            confidence=self.args.model.confidence,
+        )
+        model.to(self.device).to(self.dtype)
+        model.load_state_dict(ckpt["model"])
+        model.eval()
+        self.model = model
+        self.wav2mel = Wav2Mel(self.args, dtype=self.dtype, device=self.device)
+
+    @torch.no_grad()
+    def __call__(self, audio, sr, threshold=0.05):
+        self.model.threshold = threshold
+        audio = audio[None, :]
+        mel = self.wav2mel(audio=audio, sample_rate=sr).to(self.dtype)
+        f0 = self.model(mel=mel, infer=True, return_hz_f0=True)
+        return f0
+
+
+class Wav2Mel:
+
+    def __init__(self, args, device=None, dtype=torch.float32):
+        # self.args = args
+        self.sampling_rate = args.mel.sampling_rate
+        self.hop_size = args.mel.hop_size
+        if device is None:
+            device = "cuda" if torch.cuda.is_available() else "cpu"
+        self.device = device
+        self.dtype = dtype
+        self.stft = STFT(
+            args.mel.sampling_rate,
+            args.mel.num_mels,
+            args.mel.n_fft,
+            args.mel.win_size,
+            args.mel.hop_size,
+            args.mel.fmin,
+            args.mel.fmax,
+        )
+        self.resample_kernel = {}
+
+    def extract_nvstft(self, audio, keyshift=0, train=False):
+        mel = self.stft.get_mel(audio, keyshift=keyshift, train=train).transpose(
+            1, 2
+        )  # B, n_frames, bins
+        return mel
+
+    def extract_mel(self, audio, sample_rate, keyshift=0, train=False):
+        audio = audio.to(self.dtype).to(self.device)
+        # resample
+        if sample_rate == self.sampling_rate:
+            audio_res = audio
+        else:
+            key_str = str(sample_rate)
+            if key_str not in self.resample_kernel:
+                self.resample_kernel[key_str] = Resample(
+                    sample_rate, self.sampling_rate, lowpass_filter_width=128
+                )
+            self.resample_kernel[key_str] = (
+                self.resample_kernel[key_str].to(self.dtype).to(self.device)
+            )
+            audio_res = self.resample_kernel[key_str](audio)
+
+        # extract
+        mel = self.extract_nvstft(
+            audio_res, keyshift=keyshift, train=train
+        )  # B, n_frames, bins
+        n_frames = int(audio.shape[1] // self.hop_size) + 1
+        if n_frames > int(mel.shape[1]):
+            mel = torch.cat((mel, mel[:, -1:, :]), 1)
+        if n_frames < int(mel.shape[1]):
+            mel = mel[:, :n_frames, :]
+        return mel
+
+    def __call__(self, audio, sample_rate, keyshift=0, train=False):
+        return self.extract_mel(audio, sample_rate, keyshift=keyshift, train=train)
+
+
+class DotDict(dict):
+    def __getattr__(*args):
+        val = dict.get(*args)
+        return DotDict(val) if type(val) is dict else val
+
+    __setattr__ = dict.__setitem__
+    __delattr__ = dict.__delitem__
+
+
+class F0Predictor(object):
+    def compute_f0(self, wav, p_len):
+        """
+        input: wav:[signal_length]
+               p_len:int
+        output: f0:[signal_length//hop_length]
+        """
+        pass
+
+    def compute_f0_uv(self, wav, p_len):
+        """
+        input: wav:[signal_length]
+               p_len:int
+        output: f0:[signal_length//hop_length],uv:[signal_length//hop_length]
+        """
+        pass
+
+
+class FCPEF0Predictor(F0Predictor):
+    def __init__(
+        self,
+        model_path,
+        hop_length=512,
+        f0_min=50,
+        f0_max=1100,
+        dtype=torch.float32,
+        device=None,
+        sampling_rate=44100,
+        threshold=0.05,
+    ):
+        self.fcpe = FCPEInfer(model_path, device=device, dtype=dtype)
+        self.hop_length = hop_length
+        self.f0_min = f0_min
+        self.f0_max = f0_max
+        if device is None:
+            self.device = "cuda" if torch.cuda.is_available() else "cpu"
+        else:
+            self.device = device
+        self.threshold = threshold
+        self.sampling_rate = sampling_rate
+        self.dtype = dtype
+        self.name = "fcpe"
+
+    def repeat_expand(
+        self,
+        content: Union[torch.Tensor, np.ndarray],
+        target_len: int,
+        mode: str = "nearest",
+    ):
+        ndim = content.ndim
+
+        if content.ndim == 1:
+            content = content[None, None]
+        elif content.ndim == 2:
+            content = content[None]
+
+        assert content.ndim == 3
+
+        is_np = isinstance(content, np.ndarray)
+        if is_np:
+            content = torch.from_numpy(content)
+
+        results = torch.nn.functional.interpolate(content, size=target_len, mode=mode)
+
+        if is_np:
+            results = results.numpy()
+
+        if ndim == 1:
+            return results[0, 0]
+        elif ndim == 2:
+            return results[0]
+
+    def post_process(self, x, sampling_rate, f0, pad_to):
+        if isinstance(f0, np.ndarray):
+            f0 = torch.from_numpy(f0).float().to(x.device)
+
+        if pad_to is None:
+            return f0
+
+        f0 = self.repeat_expand(f0, pad_to)
+
+        vuv_vector = torch.zeros_like(f0)
+        vuv_vector[f0 > 0.0] = 1.0
+        vuv_vector[f0 <= 0.0] = 0.0
+
+        # 去掉0频率, 并线性插值
+        nzindex = torch.nonzero(f0).squeeze()
+        f0 = torch.index_select(f0, dim=0, index=nzindex).cpu().numpy()
+        time_org = self.hop_length / sampling_rate * nzindex.cpu().numpy()
+        time_frame = np.arange(pad_to) * self.hop_length / sampling_rate
+
+        vuv_vector = F.interpolate(vuv_vector[None, None, :], size=pad_to)[0][0]
+
+        if f0.shape[0] <= 0:
+            return (
+                torch.zeros(pad_to, dtype=torch.float, device=x.device).cpu().numpy(),
+                vuv_vector.cpu().numpy(),
+            )
+        if f0.shape[0] == 1:
+            return (
+                torch.ones(pad_to, dtype=torch.float, device=x.device) * f0[0]
+            ).cpu().numpy(), vuv_vector.cpu().numpy()
+
+        # 大概可以用 torch 重写?
+        f0 = np.interp(time_frame, time_org, f0, left=f0[0], right=f0[-1])
+        # vuv_vector = np.ceil(scipy.ndimage.zoom(vuv_vector,pad_to/len(vuv_vector),order = 0))
+
+        return f0, vuv_vector.cpu().numpy()
+
+    def compute_f0(self, wav, p_len=None):
+        x = torch.FloatTensor(wav).to(self.dtype).to(self.device)
+        if p_len is None:
+            print("fcpe p_len is None")
+            p_len = x.shape[0] // self.hop_length
+        f0 = self.fcpe(x, sr=self.sampling_rate, threshold=self.threshold)[0, :, 0]
+        if torch.all(f0 == 0):
+            rtn = f0.cpu().numpy() if p_len is None else np.zeros(p_len)
+            return rtn, rtn
+        return self.post_process(x, self.sampling_rate, f0, p_len)[0]
+
+    def compute_f0_uv(self, wav, p_len=None):
+        x = torch.FloatTensor(wav).to(self.dtype).to(self.device)
+        if p_len is None:
+            p_len = x.shape[0] // self.hop_length
+        f0 = self.fcpe(x, sr=self.sampling_rate, threshold=self.threshold)[0, :, 0]
+        if torch.all(f0 == 0):
+            rtn = f0.cpu().numpy() if p_len is None else np.zeros(p_len)
+            return rtn, rtn
+        return self.post_process(x, self.sampling_rate, f0, p_len)

src/infer_pack/predictor/RMVPE.py

@@ -0,0 +1,399 @@
+import torch.nn as nn
+import torch, numpy as np
+import torch.nn.functional as F
+from librosa.filters import mel
+
+
+class BiGRU(nn.Module):
+    def __init__(self, input_features, hidden_features, num_layers):
+        super(BiGRU, self).__init__()
+        self.gru = nn.GRU(
+            input_features,
+            hidden_features,
+            num_layers=num_layers,
+            batch_first=True,
+            bidirectional=True,
+        )
+
+    def forward(self, x):
+        return self.gru(x)[0]
+
+
+class ConvBlockRes(nn.Module):
+    def __init__(self, in_channels, out_channels, momentum=0.01):
+        super(ConvBlockRes, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=(3, 3),
+                stride=(1, 1),
+                padding=(1, 1),
+                bias=False,
+            ),
+            nn.BatchNorm2d(out_channels, momentum=momentum),
+            nn.ReLU(),
+            nn.Conv2d(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                kernel_size=(3, 3),
+                stride=(1, 1),
+                padding=(1, 1),
+                bias=False,
+            ),
+            nn.BatchNorm2d(out_channels, momentum=momentum),
+            nn.ReLU(),
+        )
+        if in_channels != out_channels:
+            self.shortcut = nn.Conv2d(in_channels, out_channels, (1, 1))
+            self.is_shortcut = True
+        else:
+            self.is_shortcut = False
+
+    def forward(self, x):
+        if self.is_shortcut:
+            return self.conv(x) + self.shortcut(x)
+        else:
+            return self.conv(x) + x
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        in_size,
+        n_encoders,
+        kernel_size,
+        n_blocks,
+        out_channels=16,
+        momentum=0.01,
+    ):
+        super(Encoder, self).__init__()
+        self.n_encoders = n_encoders
+        self.bn = nn.BatchNorm2d(in_channels, momentum=momentum)
+        self.layers = nn.ModuleList()
+        self.latent_channels = []
+        for i in range(self.n_encoders):
+            self.layers.append(
+                ResEncoderBlock(
+                    in_channels, out_channels, kernel_size, n_blocks, momentum=momentum
+                )
+            )
+            self.latent_channels.append([out_channels, in_size])
+            in_channels = out_channels
+            out_channels *= 2
+            in_size //= 2
+        self.out_size = in_size
+        self.out_channel = out_channels
+
+    def forward(self, x):
+        concat_tensors = []
+        x = self.bn(x)
+        for i in range(self.n_encoders):
+            _, x = self.layers[i](x)
+            concat_tensors.append(_)
+        return x, concat_tensors
+
+
+class ResEncoderBlock(nn.Module):
+    def __init__(
+        self, in_channels, out_channels, kernel_size, n_blocks=1, momentum=0.01
+    ):
+        super(ResEncoderBlock, self).__init__()
+        self.n_blocks = n_blocks
+        self.conv = nn.ModuleList()
+        self.conv.append(ConvBlockRes(in_channels, out_channels, momentum))
+        for i in range(n_blocks - 1):
+            self.conv.append(ConvBlockRes(out_channels, out_channels, momentum))
+        self.kernel_size = kernel_size
+        if self.kernel_size is not None:
+            self.pool = nn.AvgPool2d(kernel_size=kernel_size)
+
+    def forward(self, x):
+        for i in range(self.n_blocks):
+            x = self.conv[i](x)
+        if self.kernel_size is not None:
+            return x, self.pool(x)
+        else:
+            return x
+
+
+class Intermediate(nn.Module):  #
+    def __init__(self, in_channels, out_channels, n_inters, n_blocks, momentum=0.01):
+        super(Intermediate, self).__init__()
+        self.n_inters = n_inters
+        self.layers = nn.ModuleList()
+        self.layers.append(
+            ResEncoderBlock(in_channels, out_channels, None, n_blocks, momentum)
+        )
+        for i in range(self.n_inters - 1):
+            self.layers.append(
+                ResEncoderBlock(out_channels, out_channels, None, n_blocks, momentum)
+            )
+
+    def forward(self, x):
+        for i in range(self.n_inters):
+            x = self.layers[i](x)
+        return x
+
+
+class ResDecoderBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, stride, n_blocks=1, momentum=0.01):
+        super(ResDecoderBlock, self).__init__()
+        out_padding = (0, 1) if stride == (1, 2) else (1, 1)
+        self.n_blocks = n_blocks
+        self.conv1 = nn.Sequential(
+            nn.ConvTranspose2d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=(3, 3),
+                stride=stride,
+                padding=(1, 1),
+                output_padding=out_padding,
+                bias=False,
+            ),
+            nn.BatchNorm2d(out_channels, momentum=momentum),
+            nn.ReLU(),
+        )
+        self.conv2 = nn.ModuleList()
+        self.conv2.append(ConvBlockRes(out_channels * 2, out_channels, momentum))
+        for i in range(n_blocks - 1):
+            self.conv2.append(ConvBlockRes(out_channels, out_channels, momentum))
+
+    def forward(self, x, concat_tensor):
+        x = self.conv1(x)
+        x = torch.cat((x, concat_tensor), dim=1)
+        for i in range(self.n_blocks):
+            x = self.conv2[i](x)
+        return x
+
+
+class Decoder(nn.Module):
+    def __init__(self, in_channels, n_decoders, stride, n_blocks, momentum=0.01):
+        super(Decoder, self).__init__()
+        self.layers = nn.ModuleList()
+        self.n_decoders = n_decoders
+        for i in range(self.n_decoders):
+            out_channels = in_channels // 2
+            self.layers.append(
+                ResDecoderBlock(in_channels, out_channels, stride, n_blocks, momentum)
+            )
+            in_channels = out_channels
+
+    def forward(self, x, concat_tensors):
+        for i in range(self.n_decoders):
+            x = self.layers[i](x, concat_tensors[-1 - i])
+        return x
+
+
+class DeepUnet(nn.Module):
+    def __init__(
+        self,
+        kernel_size,
+        n_blocks,
+        en_de_layers=5,
+        inter_layers=4,
+        in_channels=1,
+        en_out_channels=16,
+    ):
+        super(DeepUnet, self).__init__()
+        self.encoder = Encoder(
+            in_channels, 128, en_de_layers, kernel_size, n_blocks, en_out_channels
+        )
+        self.intermediate = Intermediate(
+            self.encoder.out_channel // 2,
+            self.encoder.out_channel,
+            inter_layers,
+            n_blocks,
+        )
+        self.decoder = Decoder(
+            self.encoder.out_channel, en_de_layers, kernel_size, n_blocks
+        )
+
+    def forward(self, x):
+        x, concat_tensors = self.encoder(x)
+        x = self.intermediate(x)
+        x = self.decoder(x, concat_tensors)
+        return x
+
+
+class E2E(nn.Module):
+    def __init__(
+        self,
+        n_blocks,
+        n_gru,
+        kernel_size,
+        en_de_layers=5,
+        inter_layers=4,
+        in_channels=1,
+        en_out_channels=16,
+    ):
+        super(E2E, self).__init__()
+        self.unet = DeepUnet(
+            kernel_size,
+            n_blocks,
+            en_de_layers,
+            inter_layers,
+            in_channels,
+            en_out_channels,
+        )
+        self.cnn = nn.Conv2d(en_out_channels, 3, (3, 3), padding=(1, 1))
+        if n_gru:
+            self.fc = nn.Sequential(
+                BiGRU(3 * 128, 256, n_gru),
+                nn.Linear(512, 360),
+                nn.Dropout(0.25),
+                nn.Sigmoid(),
+            )
+
+    def forward(self, mel):
+        mel = mel.transpose(-1, -2).unsqueeze(1)
+        x = self.cnn(self.unet(mel)).transpose(1, 2).flatten(-2)
+        x = self.fc(x)
+        return x
+
+
+class MelSpectrogram(torch.nn.Module):
+    def __init__(
+        self,
+        is_half,
+        n_mel_channels,
+        sampling_rate,
+        win_length,
+        hop_length,
+        n_fft=None,
+        mel_fmin=0,
+        mel_fmax=None,
+        clamp=1e-5,
+    ):
+        super().__init__()
+        n_fft = win_length if n_fft is None else n_fft
+        self.hann_window = {}
+        mel_basis = mel(
+            sr=sampling_rate,
+            n_fft=n_fft,
+            n_mels=n_mel_channels,
+            fmin=mel_fmin,
+            fmax=mel_fmax,
+            htk=True,
+        )
+        mel_basis = torch.from_numpy(mel_basis).float()
+        self.register_buffer("mel_basis", mel_basis)
+        self.n_fft = win_length if n_fft is None else n_fft
+        self.hop_length = hop_length
+        self.win_length = win_length
+        self.sampling_rate = sampling_rate
+        self.n_mel_channels = n_mel_channels
+        self.clamp = clamp
+        self.is_half = is_half
+
+    def forward(self, audio, keyshift=0, speed=1, center=True):
+        factor = 2 ** (keyshift / 12)
+        n_fft_new = int(np.round(self.n_fft * factor))
+        win_length_new = int(np.round(self.win_length * factor))
+        hop_length_new = int(np.round(self.hop_length * speed))
+        keyshift_key = str(keyshift) + "_" + str(audio.device)
+        if keyshift_key not in self.hann_window:
+            self.hann_window[keyshift_key] = torch.hann_window(win_length_new).to(
+                audio.device
+            )
+        fft = torch.stft(
+            audio,
+            n_fft=n_fft_new,
+            hop_length=hop_length_new,
+            win_length=win_length_new,
+            window=self.hann_window[keyshift_key],
+            center=center,
+            return_complex=True,
+        )
+        magnitude = torch.sqrt(fft.real.pow(2) + fft.imag.pow(2))
+        if keyshift != 0:
+            size = self.n_fft // 2 + 1
+            resize = magnitude.size(1)
+            if resize < size:
+                magnitude = F.pad(magnitude, (0, 0, 0, size - resize))
+            magnitude = magnitude[:, :size, :] * self.win_length / win_length_new
+        mel_output = torch.matmul(self.mel_basis, magnitude)
+        if self.is_half == True:
+            mel_output = mel_output.half()
+        log_mel_spec = torch.log(torch.clamp(mel_output, min=self.clamp))
+        return log_mel_spec
+
+
+class RMVPE:
+    def __init__(self, model_path, is_half, device=None):
+        self.resample_kernel = {}
+        model = E2E(4, 1, (2, 2))
+        ckpt = torch.load(model_path, map_location="cpu")
+        model.load_state_dict(ckpt)
+        model.eval()
+        if is_half == True:
+            model = model.half()
+        self.model = model
+        self.resample_kernel = {}
+        self.is_half = is_half
+        if device is None:
+            device = "cuda" if torch.cuda.is_available() else "cpu"
+        self.device = device
+        self.mel_extractor = MelSpectrogram(
+            is_half, 128, 16000, 1024, 160, None, 30, 8000
+        ).to(device)
+        self.model = self.model.to(device)
+        cents_mapping = 20 * np.arange(360) + 1997.3794084376191
+        self.cents_mapping = np.pad(cents_mapping, (4, 4))  # 368
+
+    def mel2hidden(self, mel):
+        with torch.no_grad():
+            n_frames = mel.shape[-1]
+            mel = F.pad(
+                mel, (0, 32 * ((n_frames - 1) // 32 + 1) - n_frames), mode="reflect"
+            )
+            hidden = self.model(mel)
+            return hidden[:, :n_frames]
+
+    def decode(self, hidden, thred=0.03):
+        cents_pred = self.to_local_average_cents(hidden, thred=thred)
+        f0 = 10 * (2 ** (cents_pred / 1200))
+        f0[f0 == 10] = 0
+        return f0
+
+    def infer_from_audio(self, audio, thred=0.03):
+        audio = torch.from_numpy(audio).float().to(self.device).unsqueeze(0)
+        mel = self.mel_extractor(audio, center=True)
+        hidden = self.mel2hidden(mel)
+        hidden = hidden.squeeze(0).cpu().numpy()
+        if self.is_half == True:
+            hidden = hidden.astype("float32")
+        f0 = self.decode(hidden, thred=thred)
+        return f0
+
+    def to_local_average_cents(self, salience, thred=0.05):
+        center = np.argmax(salience, axis=1)
+        salience = np.pad(salience, ((0, 0), (4, 4)))
+        center += 4
+        todo_salience = []
+        todo_cents_mapping = []
+        starts = center - 4
+        ends = center + 5
+        for idx in range(salience.shape[0]):
+            todo_salience.append(salience[:, starts[idx] : ends[idx]][idx])
+            todo_cents_mapping.append(self.cents_mapping[starts[idx] : ends[idx]])
+        todo_salience = np.array(todo_salience)
+        todo_cents_mapping = np.array(todo_cents_mapping)
+        product_sum = np.sum(todo_salience * todo_cents_mapping, 1)
+        weight_sum = np.sum(todo_salience, 1)
+        devided = product_sum / weight_sum
+        maxx = np.max(salience, axis=1)
+        devided[maxx <= thred] = 0
+        return devided
+
+    def infer_from_audio_with_pitch(self, audio, thred=0.03, f0_min=50, f0_max=1100):
+        audio = torch.from_numpy(audio).float().to(self.device).unsqueeze(0)
+        mel = self.mel_extractor(audio, center=True)
+        hidden = self.mel2hidden(mel)
+        hidden = hidden.squeeze(0).cpu().numpy()
+        if self.is_half == True:
+            hidden = hidden.astype("float32")
+        f0 = self.decode(hidden, thred=thred)
+        f0[(f0 < f0_min) | (f0 > f0_max)] = 0  
+        return f0

src/infer_pack/transforms.py

@@ -0,0 +1,209 @@
+import torch
+from torch.nn import functional as F
+
+import numpy as np
+
+
+DEFAULT_MIN_BIN_WIDTH = 1e-3
+DEFAULT_MIN_BIN_HEIGHT = 1e-3
+DEFAULT_MIN_DERIVATIVE = 1e-3
+
+
+def piecewise_rational_quadratic_transform(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    tails=None,
+    tail_bound=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    if tails is None:
+        spline_fn = rational_quadratic_spline
+        spline_kwargs = {}
+    else:
+        spline_fn = unconstrained_rational_quadratic_spline
+        spline_kwargs = {"tails": tails, "tail_bound": tail_bound}
+
+    outputs, logabsdet = spline_fn(
+        inputs=inputs,
+        unnormalized_widths=unnormalized_widths,
+        unnormalized_heights=unnormalized_heights,
+        unnormalized_derivatives=unnormalized_derivatives,
+        inverse=inverse,
+        min_bin_width=min_bin_width,
+        min_bin_height=min_bin_height,
+        min_derivative=min_derivative,
+        **spline_kwargs
+    )
+    return outputs, logabsdet
+
+
+def searchsorted(bin_locations, inputs, eps=1e-6):
+    bin_locations[..., -1] += eps
+    return torch.sum(inputs[..., None] >= bin_locations, dim=-1) - 1
+
+
+def unconstrained_rational_quadratic_spline(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    tails="linear",
+    tail_bound=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
+    outside_interval_mask = ~inside_interval_mask
+
+    outputs = torch.zeros_like(inputs)
+    logabsdet = torch.zeros_like(inputs)
+
+    if tails == "linear":
+        unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
+        constant = np.log(np.exp(1 - min_derivative) - 1)
+        unnormalized_derivatives[..., 0] = constant
+        unnormalized_derivatives[..., -1] = constant
+
+        outputs[outside_interval_mask] = inputs[outside_interval_mask]
+        logabsdet[outside_interval_mask] = 0
+    else:
+        raise RuntimeError("{} tails are not implemented.".format(tails))
+
+    (
+        outputs[inside_interval_mask],
+        logabsdet[inside_interval_mask],
+    ) = rational_quadratic_spline(
+        inputs=inputs[inside_interval_mask],
+        unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
+        unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
+        unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
+        inverse=inverse,
+        left=-tail_bound,
+        right=tail_bound,
+        bottom=-tail_bound,
+        top=tail_bound,
+        min_bin_width=min_bin_width,
+        min_bin_height=min_bin_height,
+        min_derivative=min_derivative,
+    )
+
+    return outputs, logabsdet
+
+
+def rational_quadratic_spline(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    left=0.0,
+    right=1.0,
+    bottom=0.0,
+    top=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    if torch.min(inputs) < left or torch.max(inputs) > right:
+        raise ValueError("Input to a transform is not within its domain")
+
+    num_bins = unnormalized_widths.shape[-1]
+
+    if min_bin_width * num_bins > 1.0:
+        raise ValueError("Minimal bin width too large for the number of bins")
+    if min_bin_height * num_bins > 1.0:
+        raise ValueError("Minimal bin height too large for the number of bins")
+
+    widths = F.softmax(unnormalized_widths, dim=-1)
+    widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
+    cumwidths = torch.cumsum(widths, dim=-1)
+    cumwidths = F.pad(cumwidths, pad=(1, 0), mode="constant", value=0.0)
+    cumwidths = (right - left) * cumwidths + left
+    cumwidths[..., 0] = left
+    cumwidths[..., -1] = right
+    widths = cumwidths[..., 1:] - cumwidths[..., :-1]
+
+    derivatives = min_derivative + F.softplus(unnormalized_derivatives)
+
+    heights = F.softmax(unnormalized_heights, dim=-1)
+    heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
+    cumheights = torch.cumsum(heights, dim=-1)
+    cumheights = F.pad(cumheights, pad=(1, 0), mode="constant", value=0.0)
+    cumheights = (top - bottom) * cumheights + bottom
+    cumheights[..., 0] = bottom
+    cumheights[..., -1] = top
+    heights = cumheights[..., 1:] - cumheights[..., :-1]
+
+    if inverse:
+        bin_idx = searchsorted(cumheights, inputs)[..., None]
+    else:
+        bin_idx = searchsorted(cumwidths, inputs)[..., None]
+
+    input_cumwidths = cumwidths.gather(-1, bin_idx)[..., 0]
+    input_bin_widths = widths.gather(-1, bin_idx)[..., 0]
+
+    input_cumheights = cumheights.gather(-1, bin_idx)[..., 0]
+    delta = heights / widths
+    input_delta = delta.gather(-1, bin_idx)[..., 0]
+
+    input_derivatives = derivatives.gather(-1, bin_idx)[..., 0]
+    input_derivatives_plus_one = derivatives[..., 1:].gather(-1, bin_idx)[..., 0]
+
+    input_heights = heights.gather(-1, bin_idx)[..., 0]
+
+    if inverse:
+        a = (inputs - input_cumheights) * (
+            input_derivatives + input_derivatives_plus_one - 2 * input_delta
+        ) + input_heights * (input_delta - input_derivatives)
+        b = input_heights * input_derivatives - (inputs - input_cumheights) * (
+            input_derivatives + input_derivatives_plus_one - 2 * input_delta
+        )
+        c = -input_delta * (inputs - input_cumheights)
+
+        discriminant = b.pow(2) - 4 * a * c
+        assert (discriminant >= 0).all()
+
+        root = (2 * c) / (-b - torch.sqrt(discriminant))
+        outputs = root * input_bin_widths + input_cumwidths
+
+        theta_one_minus_theta = root * (1 - root)
+        denominator = input_delta + (
+            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+            * theta_one_minus_theta
+        )
+        derivative_numerator = input_delta.pow(2) * (
+            input_derivatives_plus_one * root.pow(2)
+            + 2 * input_delta * theta_one_minus_theta
+            + input_derivatives * (1 - root).pow(2)
+        )
+        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+        return outputs, -logabsdet
+    else:
+        theta = (inputs - input_cumwidths) / input_bin_widths
+        theta_one_minus_theta = theta * (1 - theta)
+
+        numerator = input_heights * (
+            input_delta * theta.pow(2) + input_derivatives * theta_one_minus_theta
+        )
+        denominator = input_delta + (
+            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+            * theta_one_minus_theta
+        )
+        outputs = input_cumheights + numerator / denominator
+
+        derivative_numerator = input_delta.pow(2) * (
+            input_derivatives_plus_one * theta.pow(2)
+            + 2 * input_delta * theta_one_minus_theta
+            + input_derivatives * (1 - theta).pow(2)
+        )
+        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+        return outputs, logabsdet

src/main.py

@@ -0,0 +1,86 @@
+import gc
+import hashlib
+import os
+import shlex
+import subprocess
+import librosa
+import numpy as np
+import soundfile as sf
+import gradio as gr
+from rvc import Config, load_hubert, get_vc, rvc_infer
+
+BASE_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+rvc_models_dir = os.path.join(BASE_DIR, 'rvc_models')
+output_dir = os.path.join(BASE_DIR, 'song_output')
+
+def get_rvc_model(voice_model):
+    model_dir = os.path.join(rvc_models_dir, voice_model)
+    rvc_model_path = next((os.path.join(model_dir, f) for f in os.listdir(model_dir) if f.endswith('.pth')), None)
+    rvc_index_path = next((os.path.join(model_dir, f) for f in os.listdir(model_dir) if f.endswith('.index')), None)
+
+    if rvc_model_path is None:
+        error_msg = f'В каталоге {model_dir} отсутствует файл модели.'
+        raise Exception(error_msg)
+
+    return rvc_model_path, rvc_index_path
+
+def convert_to_stereo(audio_path):
+    wave, sr = librosa.load(audio_path, mono=False, sr=44100)
+    if type(wave[0]) != np.ndarray:
+        stereo_path = f'Voice_stereo.wav'
+        command = shlex.split(f'ffmpeg -y -loglevel error -i "{audio_path}" -ac 2 -f wav "{stereo_path}"')
+        subprocess.run(command)
+        return stereo_path
+    else:
+        return audio_path
+
+def get_hash(filepath):
+    with open(filepath, 'rb') as f:
+        file_hash = hashlib.blake2b()
+        while chunk := f.read(8192):
+            file_hash.update(chunk)
+
+    return file_hash.hexdigest()[:11]
+
+def display_progress(percent, message, progress=gr.Progress()):
+    progress(percent, desc=message)
+
+def voice_change(voice_model, vocals_path, output_path, pitch_change, f0_method, index_rate, filter_radius, rms_mix_rate, protect, crepe_hop_length):
+    rvc_model_path, rvc_index_path = get_rvc_model(voice_model)
+    device = 'cuda:0'
+    config = Config(device, True)
+    hubert_model = load_hubert(device, config.is_half, os.path.join(rvc_models_dir, 'hubert_base.pt'))
+    cpt, version, net_g, tgt_sr, vc = get_vc(device, config.is_half, config, rvc_model_path)
+
+    rvc_infer(rvc_index_path, index_rate, vocals_path, output_path, pitch_change, f0_method, cpt, version, net_g,
+              filter_radius, tgt_sr, rms_mix_rate, protect, crepe_hop_length, vc, hubert_model)
+    del hubert_model, cpt
+    gc.collect()
+
+def song_cover_pipeline(uploaded_file, voice_model, pitch_change, index_rate=0.5, filter_radius=3, rms_mix_rate=0.25, f0_method='rmvpe',
+                        crepe_hop_length=128, protect=0.33, output_format='mp3', progress=gr.Progress()):
+
+    if not uploaded_file or not voice_model:
+        raise Exception('Убедитесь, что поле ввода песни и поле модели голоса заполнены.')
+
+    display_progress(0, '[~] Запуск конвейера генерации AI-кавера...', progress)
+
+    if not os.path.exists(uploaded_file):
+        error_msg = f'{uploaded_file} не существует.'
+        raise Exception(error_msg)
+
+    song_id = get_hash(uploaded_file)
+    song_dir = os.path.join(output_dir, song_id)
+    os.makedirs(song_dir, exist_ok=True)
+
+    orig_song_path = convert_to_stereo(uploaded_file)
+    ai_cover_path = os.path.join(song_dir, f'Converted_Voice.{output_format}')
+
+    if os.path.exists(ai_cover_path):
+        os.remove(ai_cover_path)
+
+    display_progress(0.5, '[~] Преобразование вокала...', progress)
+    voice_change(voice_model, orig_song_path, ai_cover_path, pitch_change, f0_method, index_rate,
+                 filter_radius, rms_mix_rate, protect, crepe_hop_length)
+
+    return ai_cover_path

src/modules/file_processing.py

@@ -0,0 +1,4 @@
+import gradio as gr
+
+def process_file_upload(file):
+    return gr.update(value=file)

src/modules/model_management.py

@@ -0,0 +1,89 @@
+import os
+import shutil
+import urllib.request
+import zipfile
+import gdown
+import gradio as gr
+
+
+BASE_DIR = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+rvc_models_dir = os.path.join(BASE_DIR, 'rvc_models')
+
+
+def ignore_files(models_dir):
+    models_list = os.listdir(models_dir)
+    items_to_remove = ['hubert_base.pt', 'MODELS.txt', 'rmvpe.pt', 'fcpe.pt']
+    return [item for item in models_list if item not in items_to_remove]
+
+
+def update_models_list():
+    models_l = ignore_files(rvc_models_dir)
+    return gr.update(choices=models_l)
+
+
+def extract_zip(extraction_folder, zip_name):
+    os.makedirs(extraction_folder)
+    with zipfile.ZipFile(zip_name, 'r') as zip_ref:
+        zip_ref.extractall(extraction_folder)
+    os.remove(zip_name)
+
+    index_filepath, model_filepath = None, None
+    for root, dirs, files in os.walk(extraction_folder):
+        for name in files:
+            if name.endswith('.index') and os.stat(os.path.join(root, name)).st_size > 1024 * 100:
+                index_filepath = os.path.join(root, name)
+            if name.endswith('.pth') and os.stat(os.path.join(root, name)).st_size > 1024 * 1024 * 40:
+                model_filepath = os.path.join(root, name)
+
+    if not model_filepath:
+        raise gr.Error(f'Не найден файл модели .pth в распакованном zip-файле. Пожалуйста, проверьте {extraction_folder}.')
+
+    os.rename(model_filepath, os.path.join(extraction_folder, os.path.basename(model_filepath)))
+    if index_filepath:
+        os.rename(index_filepath, os.path.join(extraction_folder, os.path.basename(index_filepath)))
+
+    for filepath in os.listdir(extraction_folder):
+        if os.path.isdir(os.path.join(extraction_folder, filepath)):
+            shutil.rmtree(os.path.join(extraction_folder, filepath))
+
+
+def download_from_url(url, dir_name, progress=gr.Progress()):
+    try:
+        progress(0, desc=f'[~] Загрузка голосовой модели с именем {dir_name}...')
+        zip_name = url.split('/')[-1]
+        extraction_folder = os.path.join(rvc_models_dir, dir_name)
+        if os.path.exists(extraction_folder):
+            raise gr.Error(f'Директория голосовой модели {dir_name} уже существует! Выберите другое имя для вашей голосовой модели.')
+
+        if 'huggingface.co' in url:
+            urllib.request.urlretrieve(url, zip_name)
+        elif 'pixeldrain.com' in url:
+            zip_name = dir_name + '.zip'
+            url = f'https://pixeldrain.com/api/file/{zip_name}'
+            urllib.request.urlretrieve(url, zip_name)
+        elif 'drive.google.com' in url:
+            zip_name = dir_name + '.zip'
+            file_id = url.split('/')[-2]
+            output = os.path.join('.', f'{dir_name}.zip')
+            gdown.download(id=file_id, output=output, quiet=False)
+
+        progress(0.5, desc='[~] Распаковка zip-файла...')
+        extract_zip(extraction_folder, zip_name)
+        return f'[+] Модель {dir_name} успешно загружена!'
+    except Exception as e:
+        raise gr.Error(str(e))
+
+
+def upload_zip_model(zip_path, dir_name, progress=gr.Progress()):
+    try:
+        extraction_folder = os.path.join(rvc_models_dir, dir_name)
+        if os.path.exists(extraction_folder):
+            raise gr.Error(f'Директория голосовой модели {dir_name} уже существует! Выберите другое имя для вашей голосовой модели.')
+
+        zip_name = zip_path.name
+        progress(0.5, desc='[~] Распаковка zip-файла...')
+        extract_zip(extraction_folder, zip_name)
+        return f'[+] Модель {dir_name} успешно загружена!'
+
+    except Exception as e:
+        raise gr.Error(str(e))

src/modules/ui_updates.py

@@ -0,0 +1,25 @@
+import gradio as gr
+    
+
+def show_hop_slider(pitch_detection_algo):
+    if pitch_detection_algo in ['mangio-crepe']:
+        return gr.update(visible=True)
+    else:
+        return gr.update(visible=False)
+
+
+def update_f0_method(use_hybrid_methods):
+    if use_hybrid_methods:
+        return gr.update(choices=['hybrid[rmvpe+fcpe]', 'hybrid[rmvpe+crepe]', 'hybrid[crepe+rmvpe]', 'hybrid[crepe+fcpe]', 'hybrid[crepe+rmvpe+fcpe]'], value='hybrid[rmvpe+fcpe]')
+    else:
+        return gr.update(choices=['rmvpe+', 'fcpe', 'rmvpe', 'mangio-crepe', 'crepe'], value='rmvpe+')
+
+
+def update_button_text():
+    return gr.update(label="Загрузить другой аудио-файл")
+
+def update_button_text_voc():
+    return gr.update(label="Загрузить другой вокал")
+
+def update_button_text_inst():
+    return gr.update(label="Загрузить другой инструментал")

src/my_utils.py

@@ -0,0 +1,18 @@
+import ffmpeg
+import numpy as np
+
+
+def load_audio(file, sr):
+    try:
+        file = (
+            file.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+        )
+        out, _ = (
+            ffmpeg.input(file, threads=0)
+            .output("-", format="f32le", acodec="pcm_f32le", ac=1, ar=sr)
+            .run(cmd=["ffmpeg", "-nostdin"], capture_stdout=True, capture_stderr=True)
+        )
+    except Exception as e:
+        raise RuntimeError(f"Failed to load audio: {e}")
+
+    return np.frombuffer(out, np.float32).flatten()

src/rvc.py

@@ -0,0 +1,187 @@
+from multiprocessing import cpu_count
+from pathlib import Path
+
+import torch
+from fairseq import checkpoint_utils
+from scipy.io import wavfile
+
+from infer_pack.models import (
+    SynthesizerTrnMs256NSFsid,
+    SynthesizerTrnMs256NSFsid_nono,
+    SynthesizerTrnMs768NSFsid,
+    SynthesizerTrnMs768NSFsid_nono,
+)
+from my_utils import load_audio
+from vc_infer_pipeline import VC
+
+BASE_DIR = Path(__file__).resolve().parent.parent
+
+
+class Config:
+    def __init__(self, device, is_half):
+        self.device = device
+        self.is_half = is_half
+        self.n_cpu = 0
+        self.gpu_name = None
+        self.gpu_mem = None
+        self.x_pad, self.x_query, self.x_center, self.x_max = self.device_config()
+
+    def device_config(self) -> tuple:
+        if torch.cuda.is_available():
+            i_device = int(self.device.split(":")[-1])
+            self.gpu_name = torch.cuda.get_device_name(i_device)
+            if (
+                    ("16" in self.gpu_name and "V100" not in self.gpu_name.upper())
+                    or "P40" in self.gpu_name.upper()
+                    or "1060" in self.gpu_name
+                    or "1070" in self.gpu_name
+                    or "1080" in self.gpu_name
+            ):
+                print("16 series/10 series P40 forced single precision")
+                self.is_half = False
+                for config_file in ["32k.json", "40k.json", "48k.json"]:
+                    with open(BASE_DIR / "src" / "configs" / config_file, "r") as f:
+                        strr = f.read().replace("true", "false")
+                    with open(BASE_DIR / "src" / "configs" / config_file, "w") as f:
+                        f.write(strr)
+                with open(BASE_DIR / "src" / "trainset_preprocess_pipeline_print.py", "r") as f:
+                    strr = f.read().replace("3.7", "3.0")
+                with open(BASE_DIR / "src" / "trainset_preprocess_pipeline_print.py", "w") as f:
+                    f.write(strr)
+            else:
+                self.gpu_name = None
+            self.gpu_mem = int(
+                torch.cuda.get_device_properties(i_device).total_memory
+                / 1024
+                / 1024
+                / 1024
+                + 0.4
+            )
+            if self.gpu_mem <= 4:
+                with open(BASE_DIR / "src" / "trainset_preprocess_pipeline_print.py", "r") as f:
+                    strr = f.read().replace("3.7", "3.0")
+                with open(BASE_DIR / "src" / "trainset_preprocess_pipeline_print.py", "w") as f:
+                    f.write(strr)
+        elif torch.backends.mps.is_available():
+            print("No supported N-card found, use MPS for inference")
+            self.device = "mps"
+        else:
+            print("No supported N-card found, use CPU for inference")
+            self.device = "cpu"
+            self.is_half = True
+
+        if self.n_cpu == 0:
+            self.n_cpu = cpu_count()
+
+        if self.is_half:
+            # 6G memory config
+            x_pad = 3
+            x_query = 10
+            x_center = 60
+            x_max = 65
+        else:
+            # 5G memory config
+            x_pad = 1
+            x_query = 6
+            x_center = 38
+            x_max = 41
+
+        if self.gpu_mem != None and self.gpu_mem <= 4:
+            x_pad = 1
+            x_query = 5
+            x_center = 30
+            x_max = 32
+
+        return x_pad, x_query, x_center, x_max
+
+
+def load_hubert(device, is_half, model_path):
+    models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task([model_path], suffix='', )
+    hubert = models[0]
+    hubert = hubert.to(device)
+
+    if is_half:
+        hubert = hubert.half()
+    else:
+        hubert = hubert.float()
+
+    hubert.eval()
+    return hubert
+
+
+def get_vc(device, is_half, config, model_path):
+    cpt = torch.load(model_path, map_location='cpu')
+    if "config" not in cpt or "weight" not in cpt:
+        raise ValueError(f'Incorrect format for {model_path}. Use a voice model trained using RVC v2 instead.')
+
+    tgt_sr = cpt["config"][-1]
+    cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
+    if_f0 = cpt.get("f0", 1)
+    version = cpt.get("version", "v1")
+
+    if version == "v1":
+        if if_f0 == 1:
+            net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=is_half)
+        else:
+            net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
+    elif version == "v2":
+        if if_f0 == 1:
+            net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=is_half)
+        else:
+            net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
+
+    del net_g.enc_q
+    print(net_g.load_state_dict(cpt["weight"], strict=False))
+    net_g.eval().to(device)
+
+    if is_half:
+        net_g = net_g.half()
+    else:
+        net_g = net_g.float()
+
+    vc = VC(tgt_sr, config)
+    return cpt, version, net_g, tgt_sr, vc
+
+
+def rvc_infer(
+    index_path,
+    index_rate,
+    input_path,
+    output_path,
+    pitch_change,
+    f0_method,
+    cpt,
+    version,
+    net_g,
+    filter_radius,
+    tgt_sr,
+    rms_mix_rate,
+    protect,
+    crepe_hop_length,
+    vc,
+    hubert_model
+):
+    audio = load_audio(input_path, 16000)
+    times = [0, 0, 0]
+    if_f0 = cpt.get('f0', 1)
+    audio_opt = vc.pipeline(
+        hubert_model,
+        net_g,
+        0,
+        audio,
+        input_path,
+        times,
+        pitch_change,
+        f0_method,
+        index_path,
+        index_rate,
+        if_f0,
+        filter_radius,
+        tgt_sr,
+        0,
+        rms_mix_rate,
+        version,
+        protect,
+        crepe_hop_length
+    )
+    wavfile.write(output_path, tgt_sr, audio_opt)

src/trainset_preprocess_pipeline_print.py

@@ -0,0 +1,146 @@
+import sys, os, multiprocessing
+from scipy import signal
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+inp_root = sys.argv[1]
+sr = int(sys.argv[2])
+n_p = int(sys.argv[3])
+exp_dir = sys.argv[4]
+noparallel = sys.argv[5] == "True"
+import numpy as np, os, traceback
+from slicer2 import Slicer
+import librosa, traceback
+from scipy.io import wavfile
+import multiprocessing
+from my_utils import load_audio
+import tqdm
+
+DoFormant = False
+Quefrency = 1.0
+Timbre = 1.0
+
+mutex = multiprocessing.Lock()
+f = open("%s/preprocess.log" % exp_dir, "a+")
+
+
+def println(strr):
+    mutex.acquire()
+    print(strr)
+    f.write("%s\n" % strr)
+    f.flush()
+    mutex.release()
+
+
+class PreProcess:
+    def __init__(self, sr, exp_dir):
+        self.slicer = Slicer(
+            sr=sr,
+            threshold=-42,
+            min_length=1500,
+            min_interval=400,
+            hop_size=15,
+            max_sil_kept=500,
+        )
+        self.sr = sr
+        self.bh, self.ah = signal.butter(N=5, Wn=48, btype="high", fs=self.sr)
+        self.per = 3.0
+        self.overlap = 0.3
+        self.tail = self.per + self.overlap
+        self.max = 0.9
+        self.alpha = 0.75
+        self.exp_dir = exp_dir
+        self.gt_wavs_dir = "%s/0_gt_wavs" % exp_dir
+        self.wavs16k_dir = "%s/1_16k_wavs" % exp_dir
+        os.makedirs(self.exp_dir, exist_ok=True)
+        os.makedirs(self.gt_wavs_dir, exist_ok=True)
+        os.makedirs(self.wavs16k_dir, exist_ok=True)
+
+    def norm_write(self, tmp_audio, idx0, idx1):
+        tmp_max = np.abs(tmp_audio).max()
+        if tmp_max > 2.5:
+            print("%s-%s-%s-filtered" % (idx0, idx1, tmp_max))
+            return
+        tmp_audio = (tmp_audio / tmp_max * (self.max * self.alpha)) + (
+            1 - self.alpha
+        ) * tmp_audio
+        wavfile.write(
+            "%s/%s_%s.wav" % (self.gt_wavs_dir, idx0, idx1),
+            self.sr,
+            tmp_audio.astype(np.float32),
+        )
+        tmp_audio = librosa.resample(
+            tmp_audio, orig_sr=self.sr, target_sr=16000
+        )  # , res_type="soxr_vhq"
+        wavfile.write(
+            "%s/%s_%s.wav" % (self.wavs16k_dir, idx0, idx1),
+            16000,
+            tmp_audio.astype(np.float32),
+        )
+
+    def pipeline(self, path, idx0):
+        try:
+            audio = load_audio(path, self.sr, DoFormant, Quefrency, Timbre)
+            # zero phased digital filter cause pre-ringing noise...
+            # audio = signal.filtfilt(self.bh, self.ah, audio)
+            audio = signal.lfilter(self.bh, self.ah, audio)
+
+            idx1 = 0
+            for audio in self.slicer.slice(audio):
+                i = 0
+                while 1:
+                    start = int(self.sr * (self.per - self.overlap) * i)
+                    i += 1
+                    if len(audio[start:]) > self.tail * self.sr:
+                        tmp_audio = audio[start : start + int(self.per * self.sr)]
+                        self.norm_write(tmp_audio, idx0, idx1)
+                        idx1 += 1
+                    else:
+                        tmp_audio = audio[start:]
+                        idx1 += 1
+                        break
+                self.norm_write(tmp_audio, idx0, idx1)
+            # println("%s->Suc." % path)
+        except:
+            println("%s->%s" % (path, traceback.format_exc()))
+
+    def pipeline_mp(self, infos, thread_n):
+        for path, idx0 in tqdm.tqdm(
+            infos, position=thread_n, leave=True, desc="thread:%s" % thread_n
+        ):
+            self.pipeline(path, idx0)
+
+    def pipeline_mp_inp_dir(self, inp_root, n_p):
+        try:
+            infos = [
+                ("%s/%s" % (inp_root, name), idx)
+                for idx, name in enumerate(sorted(list(os.listdir(inp_root))))
+            ]
+            if noparallel:
+                for i in range(n_p):
+                    self.pipeline_mp(infos[i::n_p])
+            else:
+                ps = []
+                for i in range(n_p):
+                    p = multiprocessing.Process(
+                        target=self.pipeline_mp, args=(infos[i::n_p], i)
+                    )
+                    ps.append(p)
+                    p.start()
+                for i in range(n_p):
+                    ps[i].join()
+        except:
+            println("Fail. %s" % traceback.format_exc())
+
+
+def preprocess_trainset(inp_root, sr, n_p, exp_dir):
+    pp = PreProcess(sr, exp_dir)
+    println("start preprocess")
+    println(sys.argv)
+    pp.pipeline_mp_inp_dir(inp_root, n_p)
+    println("end preprocess")
+
+
+if __name__ == "__main__":
+    preprocess_trainset(inp_root, sr, n_p, exp_dir)

src/vc_infer_pipeline.py

@@ -0,0 +1,606 @@
+from functools import lru_cache
+import numpy as np, parselmouth, torch, pdb, sys, os
+from time import time as ttime
+import torch.nn.functional as F
+import torchcrepe
+from scipy import signal
+from torch import Tensor
+import pyworld, os, faiss, librosa, torchcrepe
+import random
+import gc
+import re
+
+BASE_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+now_dir = os.path.join(BASE_DIR, 'src')
+sys.path.append(now_dir)
+
+from infer_pack.predictor.FCPE import FCPEF0Predictor
+from infer_pack.predictor.RMVPE import RMVPE
+
+bh, ah = signal.butter(N=5, Wn=48, btype="high", fs=16000)
+
+input_audio_path2wav = {}
+
+
+@lru_cache
+def cache_harvest_f0(input_audio_path, fs, f0max, f0min, frame_period):
+    audio = input_audio_path2wav[input_audio_path]
+    f0, t = pyworld.harvest(
+        audio,
+        fs=fs,
+        f0_ceil=f0max,
+        f0_floor=f0min,
+        frame_period=frame_period,
+    )
+    f0 = pyworld.stonemask(audio, f0, t, fs)
+    return f0
+
+
+def change_rms(data1, sr1, data2, sr2, rate):
+    rms1 = librosa.feature.rms(y=data1, frame_length=sr1 // 2 * 2, hop_length=sr1 // 2)
+    rms2 = librosa.feature.rms(y=data2, frame_length=sr2 // 2 * 2, hop_length=sr2 // 2)
+    rms1 = torch.from_numpy(rms1)
+    rms1 = F.interpolate(rms1.unsqueeze(0), size=data2.shape[0], mode="linear").squeeze()
+    rms2 = torch.from_numpy(rms2)
+    rms2 = F.interpolate(rms2.unsqueeze(0), size=data2.shape[0], mode="linear").squeeze()
+    rms2 = torch.max(rms2, torch.zeros_like(rms2) + 1e-6)
+    data2 *= (torch.pow(rms1, torch.tensor(1 - rate))* torch.pow(rms2, torch.tensor(rate - 1))).numpy()
+    return data2
+
+
+class VC(object):
+    def __init__(self, tgt_sr, config):
+        self.x_pad, self.x_query, self.x_center, self.x_max, self.is_half = (
+            config.x_pad,
+            config.x_query,
+            config.x_center,
+            config.x_max,
+            config.is_half,
+        )
+        self.sr = 16000
+        self.window = 160
+        self.t_pad = self.sr * self.x_pad
+        self.t_pad_tgt = tgt_sr * self.x_pad
+        self.t_pad2 = self.t_pad * 2
+        self.t_query = self.sr * self.x_query
+        self.t_center = self.sr * self.x_center
+        self.t_max = self.sr * self.x_max
+        self.device = config.device
+        
+
+    def get_optimal_torch_device(self, index: int = 0) -> torch.device:
+        if torch.cuda.is_available():
+            return torch.device(f"cuda:{index % torch.cuda.device_count()}")
+        elif torch.backends.mps.is_available():
+            return torch.device("mps")
+        return torch.device("cpu")
+
+    def get_f0_crepe_computation(
+        self,
+        x,
+        f0_min,
+        f0_max,
+        p_len,
+        hop_length=160,
+        model="full",
+    ):
+        x = x.astype(np.float32)
+        x /= np.quantile(np.abs(x), 0.999)
+        torch_device = self.get_optimal_torch_device()
+        audio = torch.from_numpy(x).to(torch_device, copy=True)
+        audio = torch.unsqueeze(audio, dim=0)
+        if audio.ndim == 2 and audio.shape[0] > 1:
+            audio = torch.mean(audio, dim=0, keepdim=True).detach()
+        audio = audio.detach()
+        pitch: Tensor = torchcrepe.predict(
+            audio,
+            self.sr,
+            hop_length,
+            f0_min,
+            f0_max,
+            model,
+            batch_size=hop_length * 2,
+            device=torch_device,
+            pad=True,
+        )
+        p_len = p_len or x.shape[0] // hop_length
+        source = np.array(pitch.squeeze(0).cpu().float().numpy())
+        source[source < 0.001] = np.nan
+        target = np.interp(
+            np.arange(0, len(source) * p_len, len(source)) / p_len,
+            np.arange(0, len(source)),
+            source,
+        )
+        f0 = np.nan_to_num(target)
+        return f0
+
+    def get_f0_official_crepe_computation(
+        self,
+        x,
+        f0_min,
+        f0_max,
+        model="full",
+    ):
+        batch_size = 512
+        audio = torch.tensor(np.copy(x))[None].float()
+        f0, pd = torchcrepe.predict(
+            audio,
+            self.sr,
+            self.window,
+            f0_min,
+            f0_max,
+            model,
+            batch_size=batch_size,
+            device=self.device,
+            return_periodicity=True,
+        )
+        pd = torchcrepe.filter.median(pd, 3)
+        f0 = torchcrepe.filter.mean(f0, 3)
+        f0[pd < 0.1] = 0
+        f0 = f0[0].cpu().numpy()
+        return f0
+
+    def get_f0_hybrid_computation(
+        self,
+        methods_str,
+        input_audio_path,
+        x,
+        f0_min,
+        f0_max,
+        p_len,
+        filter_radius,
+        crepe_hop_length,
+        time_step,
+    ):
+        methods_str = re.search("hybrid\[(.+)\]", methods_str)
+        if methods_str:
+            methods = [method.strip() for method in methods_str.group(1).split("+")]
+        f0_computation_stack = []
+        print(f"Calculating f0 pitch estimations for methods {str(methods)}")
+        x = x.astype(np.float32)
+        x /= np.quantile(np.abs(x), 0.999)
+        for method in methods:
+            f0 = None
+            if method == "mangio-crepe":
+                f0 = self.get_f0_crepe_computation(
+                    x, f0_min, f0_max, p_len, crepe_hop_length
+                )
+            elif method == "rmvpe":
+                if hasattr(self, "model_rmvpe") == False:
+
+                    self.model_rmvpe = RMVPE(
+                        os.path.join(BASE_DIR, 'rvc_models', 'rmvpe.pt'), is_half=self.is_half, device=self.device
+                    )
+                f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
+                f0 = f0[1:]
+            elif method == "fcpe":
+                self.model_fcpe = FCPEF0Predictor(
+                    os.path.join(BASE_DIR, 'rvc_models', 'fcpe.pt'),
+                    f0_min=int(f0_min),
+                    f0_max=int(f0_max),
+                    dtype=torch.float32,
+                    device=self.device,
+                    sampling_rate=self.sr,
+                    threshold=0.03,
+                )
+                f0 = self.model_fcpe.compute_f0(x, p_len=p_len)
+                del self.model_fcpe
+                gc.collect()
+            f0_computation_stack.append(f0)
+
+        print(f"Calculating hybrid median f0 from the stack of {str(methods)}")
+        f0_computation_stack = [fc for fc in f0_computation_stack if fc is not None]
+        f0_median_hybrid = None
+        if len(f0_computation_stack) == 1:
+            f0_median_hybrid = f0_computation_stack[0]
+        else:
+            f0_median_hybrid = np.nanmedian(f0_computation_stack, axis=0)
+        return f0_median_hybrid
+
+    def get_f0(
+        self,
+        input_audio_path,
+        x,
+        p_len,
+        f0_up_key,
+        f0_method,
+        filter_radius,
+        crepe_hop_length,
+        inp_f0=None,
+    ):
+        global input_audio_path2wav
+        time_step = self.window / self.sr * 1000
+        f0_min = 50
+        f0_max = 1100
+        f0_mel_min = 1127 * np.log(1 + f0_min / 700)
+        f0_mel_max = 1127 * np.log(1 + f0_max / 700)
+        if f0_method == "pm":
+            f0 = (
+                parselmouth.Sound(x, self.sr)
+                .to_pitch_ac(
+                    time_step=time_step / 1000,
+                    voicing_threshold=0.6,
+                    pitch_floor=f0_min,
+                    pitch_ceiling=f0_max,
+                )
+                .selected_array["frequency"]
+            )
+            pad_size = (p_len - len(f0) + 1) // 2
+            if pad_size > 0 or p_len - len(f0) - pad_size > 0:
+                f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
+        
+        elif f0_method == "harvest":
+            input_audio_path2wav[input_audio_path] = x.astype(np.double)
+            f0 = cache_harvest_f0(input_audio_path, self.sr, f0_max, f0_min, 10)
+            if int(filter_radius) > 2:
+                f0 = signal.medfilt(f0, 3)
+        
+        elif f0_method == "dio":
+            f0, t = pyworld.dio(
+                x.astype(np.double),
+                fs=self.sr,
+                f0_ceil=f0_max,
+                f0_floor=f0_min,
+                frame_period=10,
+            )
+            f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.sr)
+            f0 = signal.medfilt(f0, 3)
+        
+        elif f0_method == "crepe":
+            f0 = self.get_f0_official_crepe_computation(x, f0_min, f0_max)
+
+        elif f0_method == "mangio-crepe":
+            f0 = self.get_f0_crepe_computation(x, f0_min, f0_max, p_len, crepe_hop_length)
+        
+        elif f0_method == "rmvpe":
+            if hasattr(self, "model_rmvpe") == False:
+
+                self.model_rmvpe = RMVPE(
+                    os.path.join(BASE_DIR, 'rvc_models', 'rmvpe.pt'), is_half=self.is_half, device=self.device
+                )
+            f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
+        
+        elif f0_method == "rmvpe+":
+            params = {'x': x, 'p_len': p_len, 'f0_up_key': f0_up_key, 'f0_min': f0_min, 
+                      'f0_max': f0_max, 'time_step': time_step, 'filter_radius': filter_radius, 
+                      'crepe_hop_length': crepe_hop_length, 'model': "full"
+                      }
+            f0 = self.get_pitch_dependant_rmvpe(**params)
+        
+        elif f0_method == "fcpe":
+            self.model_fcpe = FCPEF0Predictor(
+                os.path.join(BASE_DIR, 'rvc_models', 'fcpe.pt'),
+                f0_min=int(f0_min),
+                f0_max=int(f0_max),
+                dtype=torch.float32,
+                device=self.device,
+                sampling_rate=self.sr,
+                threshold=0.03,
+            )
+            f0 = self.model_fcpe.compute_f0(x, p_len=p_len)
+            del self.model_fcpe
+            gc.collect()
+        
+        elif "hybrid" in f0_method:
+            input_audio_path2wav[input_audio_path] = x.astype(np.double)
+            f0 = self.get_f0_hybrid_computation(
+                f0_method,
+                input_audio_path,
+                x,
+                f0_min,
+                f0_max,
+                p_len,
+                filter_radius,
+                crepe_hop_length,
+                time_step,
+            )
+
+        f0 *= pow(2, f0_up_key / 12)
+        tf0 = self.sr // self.window
+        if inp_f0 is not None:
+            delta_t = np.round(
+                (inp_f0[:, 0].max() - inp_f0[:, 0].min()) * tf0 + 1
+            ).astype("int16")
+            replace_f0 = np.interp(
+                list(range(delta_t)), inp_f0[:, 0] * 100, inp_f0[:, 1]
+            )
+            shape = f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)].shape[0]
+            f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)] = replace_f0[
+                :shape
+            ]
+        f0bak = f0.copy()
+        f0_mel = 1127 * np.log(1 + f0 / 700)
+        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
+            f0_mel_max - f0_mel_min
+        ) + 1
+        f0_mel[f0_mel <= 1] = 1
+        f0_mel[f0_mel > 255] = 255
+        f0_coarse = np.rint(f0_mel).astype(int)
+
+        return f0_coarse, f0bak
+
+    def get_pitch_dependant_rmvpe(self, x, f0_min=1, f0_max=40000, *args, **kwargs):
+        if not hasattr(self, "model_rmvpe"):
+            
+            self.model_rmvpe = RMVPE(
+                os.path.join(BASE_DIR, 'rvc_models', 'rmvpe.pt'),
+                is_half=self.is_half,
+                device=self.device,
+            )
+            
+        f0 = self.model_rmvpe.infer_from_audio_with_pitch(x, thred=0.03, f0_min=f0_min, f0_max=f0_max)   
+            
+        return f0
+
+    
+    def vc(
+        self,
+        model,
+        net_g,
+        sid,
+        audio0,
+        pitch,
+        pitchf,
+        times,
+        index,
+        big_npy,
+        index_rate,
+        version,
+        protect,
+    ):
+        feats = torch.from_numpy(audio0)
+        if self.is_half:
+            feats = feats.half()
+        else:
+            feats = feats.float()
+        if feats.dim() == 2:
+            feats = feats.mean(-1)
+        assert feats.dim() == 1, feats.dim()
+        feats = feats.view(1, -1)
+        padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False)
+
+        inputs = {
+            "source": feats.to(self.device),
+            "padding_mask": padding_mask,
+            "output_layer": 9 if version == "v1" else 12,
+        }
+        t0 = ttime()
+        with torch.no_grad():
+            logits = model.extract_features(**inputs)
+            feats = model.final_proj(logits[0]) if version == "v1" else logits[0]
+        if protect < 0.5 and pitch != None and pitchf != None:
+            feats0 = feats.clone()
+        if (
+            isinstance(index, type(None)) == False
+            and isinstance(big_npy, type(None)) == False
+            and index_rate != 0
+        ):
+            npy = feats[0].cpu().numpy()
+            if self.is_half:
+                npy = npy.astype("float32")
+
+            score, ix = index.search(npy, k=8)
+            weight = np.square(1 / score)
+            weight /= weight.sum(axis=1, keepdims=True)
+            npy = np.sum(big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)
+
+            if self.is_half:
+                npy = npy.astype("float16")
+            feats = (
+                torch.from_numpy(npy).unsqueeze(0).to(self.device) * index_rate
+                + (1 - index_rate) * feats
+            )
+
+        feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
+        if protect < 0.5 and pitch != None and pitchf != None:
+            feats0 = F.interpolate(feats0.permute(0, 2, 1), scale_factor=2).permute(
+                0, 2, 1
+            )
+        t1 = ttime()
+        p_len = audio0.shape[0] // self.window
+        if feats.shape[1] < p_len:
+            p_len = feats.shape[1]
+            if pitch != None and pitchf != None:
+                pitch = pitch[:, :p_len]
+                pitchf = pitchf[:, :p_len]
+
+        if protect < 0.5 and pitch != None and pitchf != None:
+            pitchff = pitchf.clone()
+            pitchff[pitchf > 0] = 1
+            pitchff[pitchf < 1] = protect
+            pitchff = pitchff.unsqueeze(-1)
+            feats = feats * pitchff + feats0 * (1 - pitchff)
+            feats = feats.to(feats0.dtype)
+        p_len = torch.tensor([p_len], device=self.device).long()
+        with torch.no_grad():
+            if pitch != None and pitchf != None:
+                audio1 = (
+                    (net_g.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0])
+                    .data.cpu()
+                    .float()
+                    .numpy()
+                )
+            else:
+                audio1 = (
+                    (net_g.infer(feats, p_len, sid)[0][0, 0]).data.cpu().float().numpy()
+                )
+        del feats, p_len, padding_mask
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        t2 = ttime()
+        times[0] += t1 - t0
+        times[2] += t2 - t1
+        return audio1
+
+    def pipeline(
+        self,
+        model,
+        net_g,
+        sid,
+        audio,
+        input_audio_path,
+        times,
+        f0_up_key,
+        f0_method,
+        file_index,
+        index_rate,
+        if_f0,
+        filter_radius,
+        tgt_sr,
+        resample_sr,
+        rms_mix_rate,
+        version,
+        protect,
+        crepe_hop_length,
+        f0_file=None,
+    ):
+        if file_index != "" and os.path.exists(file_index) == True and index_rate != 0:
+            try:
+                index = faiss.read_index(file_index)
+                big_npy = index.reconstruct_n(0, index.ntotal)
+            except Exception as error:
+                print(error)
+                index = big_npy = None
+        else:
+            index = big_npy = None
+        audio = signal.filtfilt(bh, ah, audio)
+        audio_pad = np.pad(audio, (self.window // 2, self.window // 2), mode="reflect")
+        opt_ts = []
+        if audio_pad.shape[0] > self.t_max:
+            audio_sum = np.zeros_like(audio)
+            for i in range(self.window):
+                audio_sum += audio_pad[i : i - self.window]
+            for t in range(self.t_center, audio.shape[0], self.t_center):
+                opt_ts.append(
+                    t
+                    - self.t_query
+                    + np.where(
+                        np.abs(audio_sum[t - self.t_query : t + self.t_query])
+                        == np.abs(audio_sum[t - self.t_query : t + self.t_query]).min()
+                    )[0][0]
+                )
+        s = 0
+        audio_opt = []
+        t = None
+        t1 = ttime()
+        audio_pad = np.pad(audio, (self.t_pad, self.t_pad), mode="reflect")
+        p_len = audio_pad.shape[0] // self.window
+        inp_f0 = None
+        if hasattr(f0_file, "name") == True:
+            try:
+                with open(f0_file.name, "r") as f:
+                    lines = f.read().strip("\n").split("\n")
+                inp_f0 = []
+                for line in lines:
+                    inp_f0.append([float(i) for i in line.split(",")])
+                inp_f0 = np.array(inp_f0, dtype="float32")
+            except Exception as error:
+                print(error)
+        sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
+        pitch, pitchf = None, None
+        if if_f0 == 1:
+            pitch, pitchf = self.get_f0(
+                input_audio_path,
+                audio_pad,
+                p_len,
+                f0_up_key,
+                f0_method,
+                filter_radius,
+                crepe_hop_length,
+                inp_f0,
+            )
+            pitch = pitch[:p_len]
+            pitchf = pitchf[:p_len]
+            if self.device == "mps":
+                pitchf = pitchf.astype(np.float32)
+            pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
+            pitchf = torch.tensor(pitchf, device=self.device).unsqueeze(0).float()
+        t2 = ttime()
+        times[1] += t2 - t1
+        for t in opt_ts:
+            t = t // self.window * self.window
+            if if_f0 == 1:
+                audio_opt.append(
+                    self.vc(
+                        model,
+                        net_g,
+                        sid,
+                        audio_pad[s : t + self.t_pad2 + self.window],
+                        pitch[:, s // self.window : (t + self.t_pad2) // self.window],
+                        pitchf[:, s // self.window : (t + self.t_pad2) // self.window],
+                        times,
+                        index,
+                        big_npy,
+                        index_rate,
+                        version,
+                        protect,
+                    )[self.t_pad_tgt : -self.t_pad_tgt]
+                )
+            else:
+                audio_opt.append(
+                    self.vc(
+                        model,
+                        net_g,
+                        sid,
+                        audio_pad[s : t + self.t_pad2 + self.window],
+                        None,
+                        None,
+                        times,
+                        index,
+                        big_npy,
+                        index_rate,
+                        version,
+                        protect,
+                    )[self.t_pad_tgt : -self.t_pad_tgt]
+                )
+            s = t
+        if if_f0 == 1:
+            audio_opt.append(
+                self.vc(
+                    model,
+                    net_g,
+                    sid,
+                    audio_pad[t:],
+                    pitch[:, t // self.window :] if t is not None else pitch,
+                    pitchf[:, t // self.window :] if t is not None else pitchf,
+                    times,
+                    index,
+                    big_npy,
+                    index_rate,
+                    version,
+                    protect,
+                )[self.t_pad_tgt : -self.t_pad_tgt]
+            )
+        else:
+            audio_opt.append(
+                self.vc(
+                    model,
+                    net_g,
+                    sid,
+                    audio_pad[t:],
+                    None,
+                    None,
+                    times,
+                    index,
+                    big_npy,
+                    index_rate,
+                    version,
+                    protect,
+                )[self.t_pad_tgt : -self.t_pad_tgt]
+            )
+        audio_opt = np.concatenate(audio_opt)
+        if rms_mix_rate != 1:
+            audio_opt = change_rms(audio, 16000, audio_opt, tgt_sr, rms_mix_rate)
+        if resample_sr >= 16000 and tgt_sr != resample_sr:
+            audio_opt = librosa.resample(
+                audio_opt, orig_sr=tgt_sr, target_sr=resample_sr
+            )
+        audio_max = np.abs(audio_opt).max() / 0.99
+        max_int16 = 32768
+        if audio_max > 1:
+            max_int16 /= audio_max
+        audio_opt = (audio_opt * max_int16).astype(np.int16)
+        del pitch, pitchf, sid
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        return audio_opt