Spaces:
Runtime error
Runtime error
| # Import libraries | |
| from transformers import pipeline | |
| from numpy import random | |
| import gradio as gr | |
| import re | |
| import torch | |
| from torch import autocast | |
| import os | |
| # Array with song cover art styles | |
| image_input_styles = ["Random", "Pencil sketch", "Oil painting", "Pop art", "Piet Mondriaan"] | |
| # Get image type for image input | |
| """ | |
| The default setting for the art style dropdown is "Random". The below function determines which style is chosen | |
| If set to "Random", copy the art style array and remove "Random" to prevent "Random" from being a chosen art style | |
| """ | |
| def get_image_input(title, given_input_style): | |
| if given_input_style == 'Random': | |
| image_input_styles_new = image_input_styles.copy() | |
| image_input_styles_new.pop(0) | |
| random_choice = random.randint(len(image_input_styles_new)-1) | |
| final_style = image_input_styles_new[random_choice] | |
| else: | |
| final_style = given_input_style | |
| image_input = 'Cover for ' + title.lower() + ' in style of ' + final_style | |
| return image_input, final_style | |
| # Available models for generate lyrics pipeline | |
| # checkpoint = 'wvangils/GPT-Medium-Beatles-Lyrics-finetuned-newlyrics' | |
| # checkpoint = 'wvangils/GPT-Neo-125m-Beatles-Lyrics-finetuned-newlyrics' | |
| checkpoint = 'wvangils/BLOOM-560m-Beatles-Lyrics-finetuned' | |
| # Setup all the pipelines we need | |
| title_generator = pipeline('summarization', model='czearing/story-to-title') | |
| lyrics_generator = pipeline("text-generation", model=checkpoint) | |
| # For the image generator we use stable diffusion from an existing HuggingFace space, Gradio accelerated backend | |
| stable_diffusion = gr.Blocks.load(name="spaces/stabilityai/stable-diffusion-1") | |
| # Create 4 images for the given prompt and receive the first one | |
| # This function uses an existing HuggingFace space where the number of created images cannot be modified | |
| def get_image(prompt): | |
| gallery_dir = stable_diffusion(prompt, fn_index=2) | |
| images = [os.path.join(gallery_dir, img) for img in os.listdir(gallery_dir)] | |
| return [images[0]] | |
| # Lyrics generation | |
| def generate_beatles(input_prompt, temperature, top_p, given_input_style): | |
| # Create generator for different models | |
| generated_lyrics = lyrics_generator(input_prompt | |
| , max_length = 100 | |
| , num_return_sequences = 1 | |
| , return_full_text = True | |
| , temperature = temperature | |
| , top_p = top_p # Default 1.0 | |
| , no_repeat_ngram_size = 3 # Default = 0 | |
| , repetition_penalty = 1.0 # Default = 1.0 | |
| )[0]["generated_text"] | |
| # Put lyrics in the right form | |
| lyrics_sentences = re.sub('\n', '. ', generated_lyrics) | |
| # Create a title based on the generated lyrics | |
| title = title_generator(lyrics_sentences, min_length=1, max_length=10, repetition_penalty=2.5)[0]['summary_text'] | |
| # Create an image based on the generated title | |
| image_input, image_style = get_image_input(title, given_input_style) | |
| # Generate the image | |
| image = get_image(image_input) | |
| return (title, generated_lyrics, image, image_style) | |
| # Create textboxes for input and output | |
| input_box = gr.Textbox(label="Write the start of a song here", placeholder="Write the start of a new song here", value="Looking out of my window", lines=2, max_lines=5) | |
| gen_lyrics = gr.Textbox(label="Song lyrics", lines=15) | |
| gen_title = gr.Textbox(label="Proposed songtitle", lines=1) | |
| gen_image = gr.Gallery(label="Proposed song cover").style(grid=1, height="auto") | |
| gen_image_style = gr.Textbox(label="Image style", lines=1) | |
| # Layout and text around the app | |
| title='Beatles lyrics generator' | |
| description="<p style='text-align: center'>We've fine-tuned multiple language models on lyrics from The Beatles to generate Beatles-like text. Below are the results we obtained fine-tuning a GPT Neo model. After generation a title is generated using <a href='https://huggingface.co/czearing/story-to-title' target='_blank'>this model</a>. On top we use the generated title to suggest an album cover using <a href='https://huggingface.co/CompVis/stable-diffusion-v1-4' target='_blank'>Stable Diffusion 1.4</a>. Give it a try!</p>" | |
| article="""<p style='text-align: left'>These text generation models that output Beatles-like text were created by data scientists working for <a href='https://cmotions.nl/' target="_blank">Cmotions.</a> | |
| We tried several text generation models that we were able to load in Colab: a general <a href='https://huggingface.co/gpt2-medium' target='_blank'>GPT2-medium</a> model, the Eleuther AI small-sized GPT model <a href='https://huggingface.co/EleutherAI/gpt-neo-125M' target='_blank'>GPT-Neo</a> and the new kid on the block build by the <a href='https://bigscience.notion.site/BLOOM-BigScience-176B-Model-ad073ca07cdf479398d5f95d88e218c4' target='_blank'>Bigscience</a> initiative <a href='https://huggingface.co/bigscience/bloom-560m' target='_blank'>BLOOM 560m</a>. | |
| Further we've put together a <a href='https://huggingface.co/datasets/cmotions/Beatles_lyrics' target='_blank'> Huggingface dataset</a> containing all known lyrics created by The Beatles. Currently we are fine-tuning models and are evaluating the results. Once finished we will publish a blog at this <a href='https://www.theanalyticslab.nl/blogs/' target='_blank'>location </a> with all the steps we took including a Python notebook using Huggingface. | |
| The default output contains 100 tokens and has a repetition penalty of 1.0. | |
| </p>""" | |
| css = """ | |
| .gr-button-primary { | |
| text-indent: -9999px; | |
| line-height: 0; | |
| } | |
| .gr-button-primary:after { | |
| content: "Beatlify!"; | |
| text-indent: 0; | |
| display: block; | |
| line-height: initial; | |
| } | |
| """ | |
| # Let users select their own temperature and top-p | |
| temperature = gr.Slider(minimum=0.1, maximum=1.0, step=0.1, label="Change the temperature \r\n (higher temperature = more creative in lyrics generation, but posibbly less Beatly)", value=0.7, show_label=True) #high = sensitive for low probability tokens | |
| top_p = gr.Slider(minimum=0.1, maximum=1.0, step=0.1, label="Change top probability of the next word \n (higher top probability = more words to choose from for the next word, but possibly less Beatly)", value=0.5, show_label=True) | |
| given_input_style = gr.Dropdown(choices=image_input_styles, value="Random", label="Choose the art style for the lyrics cover", show_label=True) | |
| # Use generate Beatles function in demo-app Gradio | |
| gr.Interface(fn=generate_beatles | |
| , inputs=[input_box, temperature, top_p, given_input_style] | |
| , outputs=[gen_title, gen_lyrics, gen_image, gen_image_style] | |
| , title=title | |
| , css=css | |
| , description=description | |
| , article=article | |
| , allow_flagging='never' | |
| ).launch() |