Update app.py

app.py

@@ -1,44 +1,343 @@
-import torch #needed only for GPU
-from PIL import Image
-from io import BytesIO
-import numpy as np
-from diffusers import StableDiffusionLatentUpscalePipeline, StableDiffusionUpscalePipeline
-import gradio as gr
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Updated to account for UI changes from https://github.com/rkfg/audiocraft/blob/long/app.py
+# also released under the MIT license.
+
+import argparse
+from concurrent.futures import ProcessPoolExecutor
+import os
+import subprocess as sp
+from tempfile import NamedTemporaryFile
+import time
+import warnings
 import modin.pandas as pd
+import torch
+import gradio as gr
+
+from audiocraft.data.audio_utils import convert_audio
+from audiocraft.data.audio import audio_write
+from audiocraft.models import MusicGen
+
+
+MODEL = None  # Last used model
+IS_BATCHED = "facebook/MusicGen" in os.environ.get('SPACE_ID', '')
+MAX_BATCH_SIZE = 6
+BATCHED_DURATION = 15
+INTERRUPTING = True
+# We have to wrap subprocess call to clean a bit the log when using gr.make_waveform
+_old_call = sp.call
+
+
+def _call_nostderr(*args, **kwargs):
+    # Avoid ffmpeg vomitting on the logs.
+    kwargs['stderr'] = sp.DEVNULL
+    kwargs['stdout'] = sp.DEVNULL
+    _old_call(*args, **kwargs)
+
+
+sp.call = _call_nostderr
+# Preallocating the pool of processes.
+pool = ProcessPoolExecutor(3)
+pool.__enter__()
+
+
+def interrupt():
+    global INTERRUPTING
+    INTERRUPTING = True
+
+
+def make_waveform(*args, **kwargs):
+    # Further remove some warnings.
+    be = time.time()
+    with warnings.catch_warnings():
+        warnings.simplefilter('ignore')
+        out = gr.make_waveform(*args, **kwargs)
+        print("Make a video took", time.time() - be)
+        return out
+
+
+def load_model(version='melody'):
+    global MODEL
+    print("Loading model", version)
+    if MODEL is None or MODEL.name != version:
+        MODEL = MusicGen.get_pretrained(version)
+
+
+def _do_predictions(texts, melodies, duration, progress=False, **gen_kwargs):
+    MODEL.set_generation_params(duration=duration, **gen_kwargs)
+    print("new batch", len(texts), texts, [None if m is None else (m[0], m[1].shape) for m in melodies])
+    be = time.time()
+    processed_melodies = []
+    target_sr = 32000
+    target_ac = 1
+    for melody in melodies:
+        if melody is None:
+            processed_melodies.append(None)
+        else:
+            sr, melody = melody[0], torch.from_numpy(melody[1]).to(MODEL.device).float().t()
+            if melody.dim() == 1:
+                melody = melody[None]
+            melody = melody[..., :int(sr * duration)]
+            melody = convert_audio(melody, sr, target_sr, target_ac)
+            processed_melodies.append(melody)
+
+    if any(m is not None for m in processed_melodies):
+        outputs = MODEL.generate_with_chroma(
+            descriptions=texts,
+            melody_wavs=processed_melodies,
+            melody_sample_rate=target_sr,
+            progress=progress,
+        )
+    else:
+        outputs = MODEL.generate(texts, progress=progress)
+
+    outputs = outputs.detach().cpu().float()
+    out_files = []
+    for output in outputs:
+        with NamedTemporaryFile("wb", suffix=".wav", delete=False) as file:
+            audio_write(
+                file.name, output, MODEL.sample_rate, strategy="loudness",
+                loudness_headroom_db=16, loudness_compressor=True, add_suffix=False)
+            out_files.append(pool.submit(make_waveform, file.name))
+    res = [out_file.result() for out_file in out_files]
+    print("batch finished", len(texts), time.time() - be)
+    return res
+
+
+def predict_batched(texts, melodies):
+    max_text_length = 512
+    texts = [text[:max_text_length] for text in texts]
+    load_model('melody')
+    res = _do_predictions(texts, melodies, BATCHED_DURATION)
+    return [res]
+
+
+def predict_full(model, text, melody, duration, topk, topp, temperature, cfg_coef, progress=gr.Progress()):
+    global INTERRUPTING
+    INTERRUPTING = False
+    topk = int(topk)
+    load_model(model)
+
+    def _progress(generated, to_generate):
+        progress((generated, to_generate))
+        if INTERRUPTING:
+            raise gr.Error("Interrupted.")
+    MODEL.set_custom_progress_callback(_progress)
+
+    outs = _do_predictions(
+        [text], [melody], duration, progress=True,
+        top_k=topk, top_p=topp, temperature=temperature, cfg_coef=cfg_coef)
+    return outs[0]
+
+
+def ui_full(launch_kwargs):
+    with gr.Blocks() as interface:
+        gr.Markdown(
+            """
+            # MusicGen
+            This is a demo for [MusicGen](https://github.com/facebookresearch/audiocraft), a simple and controllable model for music generation
+            presented at: ["Simple and Controllable Music Generation"](https://huggingface.co/papers/2306.05284)
+            """
+        )
+        with gr.Row():
+            with gr.Column():
+                with gr.Row():
+                    text = gr.Text(label="Input Text", interactive=True)
+                    melody = gr.Audio(source="upload", type="numpy", label="Melody Condition (optional)", interactive=True)
+                with gr.Row():
+                    submit = gr.Button("Submit")
+                    # Adapted from https://github.com/rkfg/audiocraft/blob/long/app.py, MIT license.
+                    _ = gr.Button("Interrupt").click(fn=interrupt, queue=False)
+                with gr.Row():
+                    model = gr.Radio(["melody", "large", "medium", "small"], label="Model", value="melody", interactive=True)
+                with gr.Row():
+                    duration = gr.Slider(minimum=1, maximum=120, value=16, label="Duration", interactive=True)
+                with gr.Row():
+                    topk = gr.Number(label="Top-k", value=250, interactive=True)
+                    topp = gr.Number(label="Top-p", value=0, interactive=True)
+                    temperature = gr.Number(label="Temperature", value=1.0, interactive=True)
+                    cfg_coef = gr.Number(label="Classifier Free Guidance", value=3.0, interactive=True)
+            with gr.Column():
+                output = gr.Video(label="Generated Music")
+        submit.click(predict_full, inputs=[model, text, melody, duration, topk, topp, temperature, cfg_coef], outputs=[output])
+        gr.Examples(
+            fn=predict_full,
+            examples=[
+                [
+                    "An 80s driving pop song with heavy drums and synth pads in the background",
+                    "./assets/bach.mp3",
+                    "melody"
+                ],
+                [
+                    "A cheerful country song with acoustic guitars",
+                    "./assets/bolero_ravel.mp3",
+                    "melody"
+                ],
+                [
+                    "90s rock song with electric guitar and heavy drums",
+                    None,
+                    "medium"
+                ],
+                [
+                    "a light and cheerly EDM track, with syncopated drums, aery pads, and strong emotions",
+                    "./assets/bach.mp3",
+                    "melody"
+                ],
+                [
+                    "lofi slow bpm electro chill with organic samples",
+                    None,
+                    "medium",
+                ],
+            ],
+            inputs=[text, melody, model],
+            outputs=[output]
+        )
+        gr.Markdown(
+            """
+            ### More details
+
+            The model will generate a short music extract based on the description you provided.
+            The model can generate up to 30 seconds of audio in one pass. It is now possible
+            to extend the generation by feeding back the end of the previous chunk of audio.
+            This can take a long time, and the model might lose consistency. The model might also
+            decide at arbitrary positions that the song ends.
+
+            **WARNING:** Choosing long durations will take a long time to generate (2min might take ~10min). An overlap of 12 seconds
+            is kept with the previously generated chunk, and 18 "new" seconds are generated each time.
+
+            We present 4 model variations:
+            1. Melody -- a music generation model capable of generating music condition on text and melody inputs. **Note**, you can also use text only.
+            2. Small -- a 300M transformer decoder conditioned on text only.
+            3. Medium -- a 1.5B transformer decoder conditioned on text only.
+            4. Large -- a 3.3B transformer decoder conditioned on text only (might OOM for the longest sequences.)
+
+            When using `melody`, you can optionaly provide a reference audio from
+            which a broad melody will be extracted. The model will then try to follow both the description and melody provided.
+
+            You can also use your own GPU or a Google Colab by following the instructions on our repo.
+            See [github.com/facebookresearch/audiocraft](https://github.com/facebookresearch/audiocraft)
+            for more details.
+            """
+        )
+
+        interface.queue(max_size=2).launch(**launch_kwargs)
+
+
+def ui_batched(launch_kwargs):
+    with gr.Blocks() as demo:
+        gr.Markdown(
+            """
+            # MusicGen
+
+            This is the demo for [MusicGen](https://github.com/facebookresearch/audiocraft), a simple and controllable model for music generation
+            presented at: ["Simple and Controllable Music Generation"](https://huggingface.co/papers/2306.05284).
+            <br/>
+            <a href="https://huggingface.co/spaces/facebook/MusicGen?duplicate=true" style="display: inline-block;margin-top: .5em;margin-right: .25em;" target="_blank">
+            <img style="margin-bottom: 0em;display: inline;margin-top: -.25em;" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
+            for longer sequences, more control and no queue.</p>
+            """
+        )
+        with gr.Row():
+            with gr.Column():
+                with gr.Row():
+                    text = gr.Text(label="Describe your music", lines=2, interactive=True)
+                    melody = gr.Audio(source="upload", type="numpy", label="Condition on a melody (optional)", interactive=True)
+                with gr.Row():
+                    submit = gr.Button("Generate")
+            with gr.Column():
+                output = gr.Video(label="Generated Music")
+        submit.click(predict_batched, inputs=[text, melody], outputs=[output], batch=True, max_batch_size=MAX_BATCH_SIZE)
+        gr.Examples(
+            fn=predict_batched,
+            examples=[
+                [
+                    "An 80s driving pop song with heavy drums and synth pads in the background",
+                    "./assets/bach.mp3",
+                ],
+                [
+                    "A cheerful country song with acoustic guitars",
+                    "./assets/bolero_ravel.mp3",
+                ],
+                [
+                    "90s rock song with electric guitar and heavy drums",
+                    None,
+                ],
+                [
+                    "a light and cheerly EDM track, with syncopated drums, aery pads, and strong emotions bpm: 130",
+                    "./assets/bach.mp3",
+                ],
+                [
+                    "lofi slow bpm electro chill with organic samples",
+                    None,
+                ],
+            ],
+            inputs=[text, melody],
+            outputs=[output]
+        )
+        gr.Markdown("""
+        ### More details
+
+        The model will generate 12 seconds of audio based on the description you provided.
+        You can optionaly provide a reference audio from which a broad melody will be extracted.
+        The model will then try to follow both the description and melody provided.
+        All samples are generated with the `melody` model.
+
+        You can also use your own GPU or a Google Colab by following the instructions on our repo.
+
+        See [github.com/facebookresearch/audiocraft](https://github.com/facebookresearch/audiocraft)
+        for more details.
+        """)
+
+        demo.queue(max_size=3).launch(**launch_kwargs)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        '--listen',
+        type=str,
+        default='0.0.0.0' if 'SPACE_ID' in os.environ else '127.0.0.1',
+        help='IP to listen on for connections to Gradio',
+    )
+    parser.add_argument(
+        '--username', type=str, default='', help='Username for authentication'
+    )
+    parser.add_argument(
+        '--password', type=str, default='', help='Password for authentication'
+    )
+    parser.add_argument(
+        '--server_port',
+        type=int,
+        default=0,
+        help='Port to run the server listener on',
+    )
+    parser.add_argument(
+        '--inbrowser', action='store_true', help='Open in browser'
+    )
+    parser.add_argument(
+        '--share', action='store_true', help='Share the gradio UI'
+    )
+
+    args = parser.parse_args()
+
+    launch_kwargs = {}
+    launch_kwargs['server_name'] = args.listen
+
+    if args.username and args.password:
+        launch_kwargs['auth'] = (args.username, args.password)
+    if args.server_port:
+        launch_kwargs['server_port'] = args.server_port
+    if args.inbrowser:
+        launch_kwargs['inbrowser'] = args.inbrowser
+    if args.share:
+        launch_kwargs['share'] = args.share
 
-# load model for CPU or GPU
-
-model_2x = "stabilityai/sd-x2-latent-upscaler"
-model_4x = "stabilityai/stable-diffusion-x4-upscaler"
-device = "cuda" if torch.cuda.is_available() else "cpu"
-upscaler2x = StableDiffusionLatentUpscalePipeline.from_pretrained(model_2x, torch_dtype=torch.float16) if torch.cuda.is_available() else StableDiffusionLatentUpscalePipeline.from_pretrained(model_2x)
-upscaler4x = StableDiffusionUpscalePipeline.from_pretrained(model_4x, torch_dtype=torch.float16, revision="fp16") if torch.cuda.is_available() else StableDiffusionUpscalePipeline.from_pretrained(model_4x)
-upscaler2x = upscaler2x.to(device)
-upscaler4x = upscaler4x.to(device)
-
-#define interface 
-
-def upscale(raw_img, model, prompt, negative_prompt, scale, steps, Seed):
-    generator = torch.manual_seed(Seed)
-    raw_img = Image.open(raw_img).convert("RGB")
-    if model == "Upscaler 4x":
-        low_res_img = raw_img.resize((128, 128))
-        upscaled_image = upscaler4x(prompt=prompt, negative_prompt=negative_prompt, image=low_res_img, guidance_scale=scale, num_inference_steps=steps).images[0]
-    else: 
-        upscaled_image = upscaler2x(prompt=prompt, negative_prompt=negative_prompt, image=raw_img, guidance_scale=scale, num_inference_steps=steps).images[0]
-    return upscaled_image
-    
-#launch interface
-    
-gr.Interface(fn=upscale, inputs=[
-    gr.Image(type="filepath", label='Lower Resolution Image'), 
-    gr.Radio(['Upscaler 2x','Upscaler 4x'], label="Models"),
-    gr.Textbox(label="Optional: Enter a Prompt to Guide the AI's Enhancement, this can have an Img2Img Effect"), 
-    gr.Textbox(label='Experimental: Influence What you do not want the AI to Enhance. Such as Blur, Smudges, or Pixels'), 
-    gr.Slider(1, 15, 1, step=1, label='Guidance Scale: How much the AI influences the Upscaling.'), 
-    gr.Slider(5, 50, 5, step=1, label='Number of Iterations'),
-    gr.Slider(minimum=1, maximum=999999999999999999, randomize=True, step=1)], 
-    outputs=gr.Image(type="filepath", label = 'Upscaled Image'), 
-    title='SD Upscaler', 
-    description='2x Latent Upscaler using SD 2.0 And 4x Upscaler using SD 2.1. This version runs on CPU or GPU and is currently running on a T4 GPU. For 4x Upscaling use images lower than 512x512, ideally 128x128 or smaller for 512x512 output. For 2x Upscaling use up to 512x512 images for 1024x1024 output.<br><br><b>Notice: Largest Accepted Resolution is 512x512', 
-    article = "Code Monkey: <a href=\"https://huggingface.co/Manjushri\">Manjushri</a>").launch(max_threads=True, debug=True)
+    # Show the interface
+    if IS_BATCHED:
+        ui_batched(launch_kwargs)
+    else:
+        ui_full(launch_kwargs)