Reuse a working space to make it work (#1)

- Reuse a working space to make it work (d004e30cb351fcdc74a023fd09c588884b21f0de)


Co-authored-by: Fabrice TIERCELIN <[email protected]>

app.py

@@ -1,44 +1,59 @@
-import spaces
 import torch
 import torchaudio
 from einops import rearrange
+import gradio as gr
+import spaces
+import os
+import uuid
+
+# Importing the model-related functions
 from stable_audio_tools import get_pretrained_model
 from stable_audio_tools.inference.generation import generate_diffusion_cond
-import os
 
-# Load model config from stable-audio-tools
-model, model_config = get_pretrained_model(
-    "stabilityai/stable-audio-open-1.0", config_filename="model_config.json"
-)
-sample_rate = model_config["sample_rate"]
-sample_size = model_config["sample_size"]
+# Load the model outside of the GPU-decorated function
+def load_model():
+    print("Loading model...")
+    model, model_config = get_pretrained_model("stabilityai/stable-audio-open-1.0")
+    print("Model loaded successfully.")
+    return model, model_config
+
+# Function to set up, generate, and process the audio
+@spaces.GPU(duration=120)  # Allocate GPU only when this function is called
+def generate_audio(prompt, seconds_total=30, steps=100, cfg_scale=7):
+    print(f"Prompt received: {prompt}")
+    print(f"Settings: Duration={seconds_total}s, Steps={steps}, CFG Scale={cfg_scale}")
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    print(f"Using device: {device}")
 
-# Load the model using the transformers library 
-token = os.environ.get("TOKEN")
-model = AutoModelForAudioClassification.from_pretrained(
-    "stabilityai/stable-audio-open-1.0", use_auth_token=token, cache_dir=None
-)
+    # Fetch the Hugging Face token from the environment variable
+    hf_token = os.getenv('HF_TOKEN')
+    print(f"Hugging Face token: {hf_token}")
 
-device = "cuda" if torch.cuda.is_available() else "cpu"
-model = model.to(device)
+    # Use pre-loaded model and configuration
+    model, model_config = load_model()
+    sample_rate = model_config["sample_rate"]
+    sample_size = model_config["sample_size"]
 
-# --- Gradio App ---
+    print(f"Sample rate: {sample_rate}, Sample size: {sample_size}")
 
-def generate_music(prompt, seconds_total, bpm, genre):
-    """Generates music from a prompt using Stable Diffusion."""
+    model = model.to(device)
+    print("Model moved to device.")
 
     # Set up text and timing conditioning
     conditioning = [{
-        "prompt": f"{bpm} BPM {genre} {prompt}",
+        "prompt": prompt,
         "seconds_start": 0,
         "seconds_total": seconds_total
     }]
+    print(f"Conditioning: {conditioning}")
 
     # Generate stereo audio
+    print("Generating audio...")
     output = generate_diffusion_cond(
         model,
-        steps=100,
-        cfg_scale=7,
+        steps=steps,
+        cfg_scale=cfg_scale,
         conditioning=conditioning,
         sample_size=sample_size,
         sigma_min=0.3,
@@ -46,37 +61,86 @@ def generate_music(prompt, seconds_total, bpm, genre):
         sampler_type="dpmpp-3m-sde",
         device=device
     )
+    print("Audio generated.")
 
     # Rearrange audio batch to a single sequence
     output = rearrange(output, "b d n -> d (b n)")
+    print("Audio rearranged.")
 
-    # Peak normalize, clip, convert to int16, and save to file
+    # Peak normalize, clip, convert to int16
     output = output.to(torch.float32).div(torch.max(torch.abs(output))).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
-    return output
+    print("Audio normalized and converted.")
 
-@spaces.GPU(duration=120)
-def generate_music_and_save(prompt, seconds_total, bpm, genre):
-    """Generates music, saves it to a file, and returns the file path."""
+    # Generate a unique filename for the output
+    unique_filename = f"output_{uuid.uuid4().hex}.wav"
+    print(f"Saving audio to file: {unique_filename}")
 
-    output = generate_music(prompt, seconds_total, bpm, genre)
-    filename = "output.wav"
-    torchaudio.save(filename, output, sample_rate)
-    return filename
+    # Save to file
+    torchaudio.save(unique_filename, output, sample_rate)
+    print(f"Audio saved: {unique_filename}")
 
-# Create Gradio interface
-iface = spaces.Interface(
-    generate_music_and_save,
+    # Return the path to the generated audio file
+    return unique_filename
+
+# Setting up the Gradio Interface
+interface = gr.Interface(
+    fn=generate_audio,
     inputs=[
-        spaces.Textbox(label="Prompt (e.g., 'upbeat drum loop')", lines=1),
-        spaces.Slider(label="Duration (seconds)", minimum=1, maximum=60, step=1),
-        spaces.Slider(label="BPM", minimum=60, maximum=200, step=1),
-        spaces.Dropdown(label="Genre", choices=["pop", "rock", "hip hop", "electronic", "classical"], value="pop")
+        gr.Textbox(label="Prompt", placeholder="Enter your text prompt here"),
+        gr.Slider(0, 47, value=30, label="Duration in Seconds"),
+        gr.Slider(10, 150, value=100, step=10, label="Number of Diffusion Steps"),
+        gr.Slider(1, 15, value=7, step=0.1, label="CFG Scale")
+    ],
+    outputs=gr.Audio(type="filepath", label="Generated Audio"),
+    title="Stable Audio Generator",
+    description="Generate variable-length stereo audio at 44.1kHz from text prompts using Stable Audio Open 1.0.",
+    examples=[
+    [
+        "Create a serene soundscape of a quiet beach at sunset.",  # Text prompt
+ 
+        45,  # Duration in Seconds
+        100,  # Number of Diffusion Steps
+        10,  # CFG Scale
+    ],
+    [
+        "Generate an energetic and bustling city street scene with distant traffic and close conversations.",  # Text prompt
+        
+        30,  # Duration in Seconds
+        120,  # Number of Diffusion Steps
+        5,  # CFG Scale
+    ],
+    [
+        "Simulate a forest ambiance with birds chirping and wind rustling through the leaves.",  # Text prompt
+        60,  # Duration in Seconds
+        140,  # Number of Diffusion Steps
+        7.5,  # CFG Scale
     ],
-    outputs=[
-        spaces.Audio(label="Generated Music")
+    [
+        "Recreate a gentle rainfall with distant thunder.",  # Text prompt
+       
+        35,  # Duration in Seconds
+        110,  # Number of Diffusion Steps
+        8,  # CFG Scale
+        
     ],
-    title="Stable Audio Open",
-    description="Generate music from text prompts using Stable Audio."
-)
+    [
+        "Imagine a jazz cafe environment with soft music and ambient chatter.",  # Text prompt
+        25,  # Duration in Seconds
+        90,  # Number of Diffusion Steps
+        6,  # CFG Scale
+  
+    ],
+    ["Rock beat played in a treated studio, session drumming on an acoustic kit.",
+        30,  # Duration in Seconds
+        100,  # Number of Diffusion Steps
+        7,  # CFG Scale
+     
+    ]
+])
+
+
+# Pre-load the model to avoid multiprocessing issues
+model, model_config = load_model()
 
-iface.launch(share=True)
+# Launch the Interface
+interface.launch()