Spaces:
Runtime error
Runtime error
Commit
•
b725215
1
Parent(s):
4c9245b
Update app.py
Browse files
app.py
CHANGED
@@ -12,23 +12,33 @@ from stable_audio_tools.inference.generation import generate_diffusion_cond
|
|
12 |
|
13 |
# Load the model outside of the GPU-decorated function
|
14 |
def load_model():
|
|
|
15 |
model, model_config = get_pretrained_model("stabilityai/stable-audio-open-1.0")
|
|
|
16 |
return model, model_config
|
17 |
|
18 |
# Function to set up, generate, and process the audio
|
19 |
@spaces.GPU(duration=120) # Allocate GPU only when this function is called
|
20 |
def generate_audio(prompt, seconds_total=30, steps=100, cfg_scale=7):
|
|
|
|
|
|
|
21 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
|
22 |
|
23 |
# Fetch the Hugging Face token from the environment variable
|
24 |
hf_token = os.getenv('HF_TOKEN')
|
|
|
25 |
|
26 |
# Use pre-loaded model and configuration
|
27 |
model, model_config = load_model()
|
28 |
sample_rate = model_config["sample_rate"]
|
29 |
sample_size = model_config["sample_size"]
|
30 |
|
|
|
|
|
31 |
model = model.to(device)
|
|
|
32 |
|
33 |
# Set up text and timing conditioning
|
34 |
conditioning = [{
|
@@ -36,8 +46,10 @@ def generate_audio(prompt, seconds_total=30, steps=100, cfg_scale=7):
|
|
36 |
"seconds_start": 0,
|
37 |
"seconds_total": seconds_total
|
38 |
}]
|
|
|
39 |
|
40 |
# Generate stereo audio
|
|
|
41 |
output = generate_diffusion_cond(
|
42 |
model,
|
43 |
steps=steps,
|
@@ -49,17 +61,23 @@ def generate_audio(prompt, seconds_total=30, steps=100, cfg_scale=7):
|
|
49 |
sampler_type="dpmpp-3m-sde",
|
50 |
device=device
|
51 |
)
|
|
|
52 |
|
53 |
# Rearrange audio batch to a single sequence
|
54 |
output = rearrange(output, "b d n -> d (b n)")
|
|
|
55 |
|
56 |
# Peak normalize, clip, convert to int16
|
57 |
output = output.to(torch.float32).div(torch.max(torch.abs(output))).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
|
|
|
58 |
|
59 |
# Generate a unique filename for the output
|
60 |
unique_filename = f"output_{uuid.uuid4().hex}.wav"
|
|
|
|
|
61 |
# Save to file
|
62 |
torchaudio.save(unique_filename, output, sample_rate)
|
|
|
63 |
|
64 |
# Return the path to the generated audio file
|
65 |
return unique_filename
|
@@ -70,7 +88,7 @@ interface = gr.Interface(
|
|
70 |
inputs=[
|
71 |
gr.Textbox(label="Prompt", placeholder="Enter your text prompt here"),
|
72 |
gr.Slider(0, 47, value=30, label="Duration in Seconds"),
|
73 |
-
gr.Slider(10,
|
74 |
gr.Slider(1, 15, value=7, step=0.1, label="CFG Scale")
|
75 |
],
|
76 |
outputs=gr.Audio(type="filepath", label="Generated Audio"),
|
|
|
12 |
|
13 |
# Load the model outside of the GPU-decorated function
|
14 |
def load_model():
|
15 |
+
print("Loading model...")
|
16 |
model, model_config = get_pretrained_model("stabilityai/stable-audio-open-1.0")
|
17 |
+
print("Model loaded successfully.")
|
18 |
return model, model_config
|
19 |
|
20 |
# Function to set up, generate, and process the audio
|
21 |
@spaces.GPU(duration=120) # Allocate GPU only when this function is called
|
22 |
def generate_audio(prompt, seconds_total=30, steps=100, cfg_scale=7):
|
23 |
+
print(f"Prompt received: {prompt}")
|
24 |
+
print(f"Settings: Duration={seconds_total}s, Steps={steps}, CFG Scale={cfg_scale}")
|
25 |
+
|
26 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
27 |
+
print(f"Using device: {device}")
|
28 |
|
29 |
# Fetch the Hugging Face token from the environment variable
|
30 |
hf_token = os.getenv('HF_TOKEN')
|
31 |
+
print(f"Hugging Face token: {hf_token}")
|
32 |
|
33 |
# Use pre-loaded model and configuration
|
34 |
model, model_config = load_model()
|
35 |
sample_rate = model_config["sample_rate"]
|
36 |
sample_size = model_config["sample_size"]
|
37 |
|
38 |
+
print(f"Sample rate: {sample_rate}, Sample size: {sample_size}")
|
39 |
+
|
40 |
model = model.to(device)
|
41 |
+
print("Model moved to device.")
|
42 |
|
43 |
# Set up text and timing conditioning
|
44 |
conditioning = [{
|
|
|
46 |
"seconds_start": 0,
|
47 |
"seconds_total": seconds_total
|
48 |
}]
|
49 |
+
print(f"Conditioning: {conditioning}")
|
50 |
|
51 |
# Generate stereo audio
|
52 |
+
print("Generating audio...")
|
53 |
output = generate_diffusion_cond(
|
54 |
model,
|
55 |
steps=steps,
|
|
|
61 |
sampler_type="dpmpp-3m-sde",
|
62 |
device=device
|
63 |
)
|
64 |
+
print("Audio generated.")
|
65 |
|
66 |
# Rearrange audio batch to a single sequence
|
67 |
output = rearrange(output, "b d n -> d (b n)")
|
68 |
+
print("Audio rearranged.")
|
69 |
|
70 |
# Peak normalize, clip, convert to int16
|
71 |
output = output.to(torch.float32).div(torch.max(torch.abs(output))).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
|
72 |
+
print("Audio normalized and converted.")
|
73 |
|
74 |
# Generate a unique filename for the output
|
75 |
unique_filename = f"output_{uuid.uuid4().hex}.wav"
|
76 |
+
print(f"Saving audio to file: {unique_filename}")
|
77 |
+
|
78 |
# Save to file
|
79 |
torchaudio.save(unique_filename, output, sample_rate)
|
80 |
+
print(f"Audio saved: {unique_filename}")
|
81 |
|
82 |
# Return the path to the generated audio file
|
83 |
return unique_filename
|
|
|
88 |
inputs=[
|
89 |
gr.Textbox(label="Prompt", placeholder="Enter your text prompt here"),
|
90 |
gr.Slider(0, 47, value=30, label="Duration in Seconds"),
|
91 |
+
gr.Slider(10, 150, value=100, step=10, label="Number of Diffusion Steps"),
|
92 |
gr.Slider(1, 15, value=7, step=0.1, label="CFG Scale")
|
93 |
],
|
94 |
outputs=gr.Audio(type="filepath", label="Generated Audio"),
|