rrocap

README.md

@@ -1,6 +1,6 @@
 ---
 title: "MusicGen"
-python_version: "3.10"
+python_version: "3.9"
 tags:
   - "music generation"
   - "language models"
@@ -10,7 +10,7 @@ emoji: 🎵
 colorFrom: gray
 colorTo: blue
 sdk: gradio
-sdk_version: 5.49.1
+sdk_version: 3.34.0
 pinned: true
 license: "cc-by-nc-4.0"
 disable_embedding: true

demos/musicgen_app.py

@@ -8,19 +8,7 @@
 # also released under the MIT license.
 
 import argparse
-import subprocess
-import sys as _sys
 from concurrent.futures import ProcessPoolExecutor
-
-# audiocraft pulls demucs (for melody conditioning); its torchaudio<2.1 pin
-# would drag torch back and break `spaces`. Install demucs without deps so the
-# constraint doesn't propagate.
-subprocess.run(
-    [_sys.executable, "-m", "pip", "install", "--no-deps", "demucs"],
-    check=True,
-)
-
-import spaces
 import logging
 import os
 from pathlib import Path
@@ -94,41 +82,23 @@ class FileCleaner:
 file_cleaner = FileCleaner()
 
 
-
-
-ALL_MODELS = [
-    'facebook/musicgen-melody',
-    'facebook/musicgen-medium',
-    'facebook/musicgen-small',
-    'facebook/musicgen-large',
-    'facebook/musicgen-melody-large',
-    'facebook/musicgen-stereo-small',
-    'facebook/musicgen-stereo-medium',
-    'facebook/musicgen-stereo-melody',
-    'facebook/musicgen-stereo-large',
-    'facebook/musicgen-stereo-melody-large',
-]
-
-
-def _load_to_cuda(name):
-    print(f"Preloading {name} to cuda")
-    m = MusicGen.get_pretrained(name, device='cuda')
-    m.compression_model.to('cuda')
-    m.lm.to('cuda')
-    return m
-
-
-# All MusicGen variants (LM + EncodeC) fit on a 48GB ZeroGPU. Preload them once
-# on CUDA at module load so generation switches between models without
-# re-downloading or re-instantiating per request. `import spaces` hijacks CUDA,
-# so the .to('cuda') here gets routed to the real device.
-MODELS = {name: _load_to_cuda(name) for name in ALL_MODELS}
+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='facebook/musicgen-melody'):
     global MODEL
-    MODEL = MODELS[version]
-    print("Selected model", version)
+    print("Loading model", version)
+    if MODEL is None or MODEL.name != version:
+        del MODEL
+        MODEL = None  # in case loading would crash
+        MODEL = MusicGen.get_pretrained(version)
 
 
 def load_diffusion():
@@ -182,29 +152,32 @@ def _do_predictions(texts, melodies, duration, progress=False, gradio_progress=N
             outputs_diffusion = rearrange(outputs_diffusion, '(s b) c t -> b (s c) t', s=2)
         outputs = torch.cat([outputs[0], outputs_diffusion], dim=0)
     outputs = outputs.detach().cpu().float()
+    pending_videos = []
     out_wavs = []
     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)
+            pending_videos.append(pool.submit(make_waveform, file.name))
             out_wavs.append(file.name)
             file_cleaner.add(file.name)
+    out_videos = [pending_video.result() for pending_video in pending_videos]
+    for video in out_videos:
+        file_cleaner.add(video)
     print("batch finished", len(texts), time.time() - be)
     print("Tempfiles currently stored: ", len(file_cleaner.files))
-    return out_wavs
+    return out_videos, out_wavs
 
 
-@spaces.GPU(duration=120)
 def predict_batched(texts, melodies):
     max_text_length = 512
     texts = [text[:max_text_length] for text in texts]
     load_model('facebook/musicgen-stereo-melody')
-    wavs = _do_predictions(texts, melodies, BATCHED_DURATION)
-    return [wavs]
+    res = _do_predictions(texts, melodies, BATCHED_DURATION)
+    return res
 
 
-@spaces.GPU(duration=120)
 def predict_full(model, model_path, decoder, text, melody, duration, topk, topp, temperature, cfg_coef, progress=gr.Progress()):
     global INTERRUPTING
     global USE_DIFFUSION
@@ -244,17 +217,27 @@ def predict_full(model, model_path, decoder, text, melody, duration, topk, topp,
             raise gr.Error("Interrupted.")
     MODEL.set_custom_progress_callback(_progress)
 
-    wavs = _do_predictions(
+    videos, wavs = _do_predictions(
         [text], [melody], duration, progress=True,
         top_k=topk, top_p=topp, temperature=temperature, cfg_coef=cfg_coef,
         gradio_progress=progress)
     if USE_DIFFUSION:
-        return wavs[0], wavs[1]
-    return wavs[0], None
+        return videos[0], wavs[0], videos[1], wavs[1]
+    return videos[0], wavs[0], None, None
+
+
+def toggle_audio_src(choice):
+    if choice == "mic":
+        return gr.update(source="microphone", value=None, label="Microphone")
+    else:
+        return gr.update(source="upload", value=None, label="File")
 
 
 def toggle_diffusion(choice):
-    return [gr.update(visible=choice == "MultiBand_Diffusion")]
+    if choice == "MultiBand_Diffusion":
+        return [gr.update(visible=True)] * 2
+    else:
+        return [gr.update(visible=False)] * 2
 
 
 def ui_full(launch_kwargs):
@@ -271,9 +254,11 @@ def ui_full(launch_kwargs):
             with gr.Column():
                 with gr.Row():
                     text = gr.Text(label="Input Text", interactive=True)
-                    melody = gr.Audio(sources=["upload", "microphone"], type="numpy",
-                                      label="Condition on a melody (optional)",
-                                      interactive=True, elem_id="melody-input")
+                    with gr.Column():
+                        radio = gr.Radio(["file", "mic"], value="file",
+                                         label="Condition on a melody (optional) File or Mic")
+                        melody = gr.Audio(source="upload", type="numpy", label="File",
+                                          interactive=True, elem_id="melody-input")
                 with gr.Row():
                     submit = gr.Button("Submit")
                     # Adapted from https://github.com/rkfg/audiocraft/blob/long/app.py, MIT license.
@@ -297,12 +282,15 @@ def ui_full(launch_kwargs):
                     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():
-                audio_output = gr.Audio(label="Generated Music", type='filepath')
-                audio_diffusion = gr.Audio(label="MultiBand Diffusion Decoder", type='filepath')
-        submit.click(toggle_diffusion, decoder, [audio_diffusion], queue=False,
+                output = gr.Video(label="Generated Music")
+                audio_output = gr.Audio(label="Generated Music (wav)", type='filepath')
+                diffusion_output = gr.Video(label="MultiBand Diffusion Decoder")
+                audio_diffusion = gr.Audio(label="MultiBand Diffusion Decoder (wav)", type='filepath')
+        submit.click(toggle_diffusion, decoder, [diffusion_output, audio_diffusion], queue=False,
                      show_progress=False).then(predict_full, inputs=[model, model_path, decoder, text, melody, duration, topk, topp,
                                                                      temperature, cfg_coef],
-                                               outputs=[audio_output, audio_diffusion])
+                                               outputs=[output, audio_output, diffusion_output, audio_diffusion])
+        radio.change(toggle_audio_src, radio, [melody], queue=False, show_progress=False)
 
         gr.Examples(
             fn=predict_full,
@@ -345,7 +333,7 @@ def ui_full(launch_kwargs):
                 ],
             ],
             inputs=[text, melody, model, decoder],
-            outputs=[audio_output]
+            outputs=[output]
         )
         gr.Markdown(
             """
@@ -416,15 +404,19 @@ def ui_batched(launch_kwargs):
             with gr.Column():
                 with gr.Row():
                     text = gr.Text(label="Describe your music", lines=2, interactive=True)
-                    melody = gr.Audio(sources=["upload", "microphone"], type="numpy",
-                                      label="Condition on a melody (optional)",
-                                      interactive=True, elem_id="melody-input")
+                    with gr.Column():
+                        radio = gr.Radio(["file", "mic"], value="file",
+                                         label="Condition on a melody (optional) File or Mic")
+                        melody = gr.Audio(source="upload", type="numpy", label="File",
+                                          interactive=True, elem_id="melody-input")
                 with gr.Row():
                     submit = gr.Button("Generate")
             with gr.Column():
-                audio_output = gr.Audio(label="Generated Music", type='filepath')
+                output = gr.Video(label="Generated Music")
+                audio_output = gr.Audio(label="Generated Music (wav)", type='filepath')
         submit.click(predict_batched, inputs=[text, melody],
-                     outputs=[audio_output], batch=True, max_batch_size=MAX_BATCH_SIZE)
+                     outputs=[output, audio_output], batch=True, max_batch_size=MAX_BATCH_SIZE)
+        radio.change(toggle_audio_src, radio, [melody], queue=False, show_progress=False)
         gr.Examples(
             fn=predict_batched,
             examples=[
@@ -450,7 +442,7 @@ def ui_batched(launch_kwargs):
                 ],
             ],
             inputs=[text, melody],
-            outputs=[audio_output]
+            outputs=[output]
         )
         gr.Markdown("""
         ### More details

requirements.txt

@@ -1,23 +1,23 @@
-torch==2.8.0
-torchaudio==2.8.0
+# please make sure you have already a pytorch install that is cuda enabled!
 av
 einops
 flashy>=0.0.1
 hydra-core>=1.1
 hydra_colorlog
 julius
-lameenc
-openunmix
-dora-search
-pyyaml
 num2words
 numpy
 sentencepiece
-spacy
+spacy==3.5.2
+torch<2.1.0
+torchaudio<2.1.0
+huggingface_hub
 tqdm
-transformers>=4.31.0
+transformers>=4.31.0  # need Encodec there.
 xformers
+demucs
 librosa
+gradio_client==0.2.6
 torchmetrics
 encodec
 protobuf