Load model once, remove gradients from accentor

app.py

@@ -2,7 +2,6 @@ import tempfile
 
 import gradio as gr
 
-from TTS.utils.manage import ModelManager
 from TTS.utils.synthesizer import Synthesizer
 import requests
 from os.path import exists
@@ -11,19 +10,11 @@ from datetime import datetime
 from stress import sentence_to_stress
 from enum import Enum
 import torch
-import gc
 
 class StressOption(Enum):
     ManualStress = "Наголоси вручну"
     AutomaticStress = "Автоматичні наголоси (Beta)"
 
-MODEL_NAMES = [
-    "uk/mykyta/vits-tts"
-]
-MODELS = {}
-
-manager = ModelManager()
-
 
 def download(url, file_name):
     if not exists(file_name):
@@ -35,39 +26,36 @@ def download(url, file_name):
         print(f"Found {file_name}. Skipping download...")
 
 
-for MODEL_NAME in MODEL_NAMES:
-    print(f"downloading {MODEL_NAME}")
-    release_number = "v2.0.0-beta"
-    model_link = f"https://github.com/robinhad/ukrainian-tts/releases/download/{release_number}/model-inference.pth"
-    config_link = f"https://github.com/robinhad/ukrainian-tts/releases/download/{release_number}/config.json"
+print("downloading uk/mykyta/vits-tts")
+release_number = "v2.0.0-beta"
+model_link = f"https://github.com/robinhad/ukrainian-tts/releases/download/{release_number}/model-inference.pth"
+config_link = f"https://github.com/robinhad/ukrainian-tts/releases/download/{release_number}/config.json"
 
-    model_path = "model.pth"
-    config_path = "config.json"
+model_path = "model.pth"
+config_path = "config.json"
 
-    download(model_link, model_path)
-    download(config_link, config_path)
+download(model_link, model_path)
+download(config_link, config_path)
 
-    
-    #MODELS[MODEL_NAME] = synthesizer
 
+synthesizer = Synthesizer(
+    model_path, config_path, None, None, None,
+)
+
+if synthesizer is None:
+    raise NameError("model not found")
 
 def tts(text: str, stress: str):
     text = preprocess_text(text)
     text_limit = 1200
     text = text if len(text) < text_limit else text[0:text_limit] # mitigate crashes on hf space
     text = sentence_to_stress(text) if stress == StressOption.AutomaticStress.value else text
-    print(text, datetime.utcnow())
+    print(text, stress, datetime.utcnow())
 
     with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as fp:
         with torch.no_grad():
-            synthesizer = Synthesizer(
-                model_path, config_path, None, None, None,
-            )
-            if synthesizer is None:
-                raise NameError("model not found")
             wavs = synthesizer.tts(text)
             synthesizer.save_wav(wavs, fp)
-        gc.collect()
         return fp.name
 
 
@@ -91,7 +79,8 @@ iface = gr.Interface(
     "Github: [https://github.com/robinhad/ukrainian-tts](https://github.com/robinhad/ukrainian-tts)",
     examples=[
         ["Введ+іть, б+удь л+аска, сво+є р+ечення.", StressOption.ManualStress.value],
-        ["Привіт, як тебе звати?", StressOption.AutomaticStress.value]
+        ["Введіть, будь ласка, своє речення.", StressOption.ManualStress.value],
+        ["Привіт, як тебе звати?", StressOption.AutomaticStress.value],
     ]
 )
 iface.launch(enable_queue=True, prevent_thread_lock=True)

stress.py

@@ -14,7 +14,8 @@ replace_accents = importer.load_pickle("uk-accentor", "replace_accents")
 alphabet = "абгґдеєжзиіїйклмнопрстуфхцчшщьюя"
 
 def accent_word(word):
-    stressed_words = accentor.predict([word], mode='stress')
+    with torch.no_grad():
+        stressed_words = accentor.predict([word], mode='stress')
     plused_words = [replace_accents(x) for x in stressed_words]
     return plused_words[0]