Spaces:
Running
on
Zero
Running
on
Zero
ameerazam08
commited on
Commit
•
2142ea2
1
Parent(s):
b3812f7
Update app.py
Browse files
app.py
CHANGED
@@ -1,3 +1,87 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import torch
|
2 |
import torchaudio
|
3 |
from einops import rearrange
|
@@ -10,12 +94,6 @@ import uuid
|
|
10 |
from stable_audio_tools import get_pretrained_model
|
11 |
from stable_audio_tools.inference.generation import generate_diffusion_cond
|
12 |
|
13 |
-
|
14 |
-
from huggingface_hub import login
|
15 |
-
|
16 |
-
hf_token = os.getenv('HF_TOKEN')
|
17 |
-
login(token=hf_token,add_to_git_credential=True)
|
18 |
-
|
19 |
# Load the model outside of the GPU-decorated function
|
20 |
def load_model():
|
21 |
print("Loading model...")
|
@@ -23,30 +101,43 @@ def load_model():
|
|
23 |
print("Model loaded successfully.")
|
24 |
return model, model_config
|
25 |
|
26 |
-
#
|
27 |
-
@spaces.GPU(duration=120)
|
28 |
-
def generate_audio(prompt,
|
|
|
|
|
|
|
29 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
|
30 |
|
31 |
-
#
|
|
|
|
|
|
|
|
|
32 |
model, model_config = load_model()
|
33 |
sample_rate = model_config["sample_rate"]
|
34 |
sample_size = model_config["sample_size"]
|
35 |
|
|
|
|
|
36 |
model = model.to(device)
|
|
|
37 |
|
38 |
# Set up text and timing conditioning
|
39 |
conditioning = [{
|
40 |
-
"prompt":
|
41 |
"seconds_start": 0,
|
42 |
"seconds_total": seconds_total
|
43 |
}]
|
|
|
44 |
|
45 |
# Generate stereo audio
|
|
|
46 |
output = generate_diffusion_cond(
|
47 |
model,
|
48 |
-
steps=
|
49 |
-
cfg_scale=
|
50 |
conditioning=conditioning,
|
51 |
sample_size=sample_size,
|
52 |
sigma_min=0.3,
|
@@ -54,30 +145,44 @@ def generate_audio(prompt, bpm, seconds_total):
|
|
54 |
sampler_type="dpmpp-3m-sde",
|
55 |
device=device
|
56 |
)
|
|
|
57 |
|
58 |
# Rearrange audio batch to a single sequence
|
59 |
output = rearrange(output, "b d n -> d (b n)")
|
|
|
60 |
|
61 |
-
# Peak normalize, clip, convert to int16
|
62 |
output = output.to(torch.float32).div(torch.max(torch.abs(output))).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
|
|
|
|
|
|
|
|
68 |
|
69 |
-
#
|
70 |
-
|
|
|
|
|
|
|
71 |
fn=generate_audio,
|
72 |
inputs=[
|
73 |
-
gr.Textbox(label="Prompt", placeholder="Enter
|
74 |
-
gr.
|
75 |
-
gr.
|
|
|
76 |
],
|
77 |
-
outputs=gr.Audio(label="Generated Audio"),
|
78 |
-
title="Stable Audio
|
79 |
-
description="Generate audio
|
80 |
)
|
81 |
|
82 |
-
#
|
83 |
-
|
|
|
|
|
|
|
|
|
|
1 |
+
# import torch
|
2 |
+
# import torchaudio
|
3 |
+
# from einops import rearrange
|
4 |
+
# import gradio as gr
|
5 |
+
# import spaces
|
6 |
+
# import os
|
7 |
+
# import uuid
|
8 |
+
|
9 |
+
# # Importing the model-related functions
|
10 |
+
# from stable_audio_tools import get_pretrained_model
|
11 |
+
# from stable_audio_tools.inference.generation import generate_diffusion_cond
|
12 |
+
|
13 |
+
|
14 |
+
# from huggingface_hub import login
|
15 |
+
|
16 |
+
# hf_token = os.getenv('HF_TOKEN')
|
17 |
+
# login(token=hf_token,add_to_git_credential=True)
|
18 |
+
|
19 |
+
# # Load the model outside of the GPU-decorated function
|
20 |
+
# def load_model():
|
21 |
+
# print("Loading model...")
|
22 |
+
# model, model_config = get_pretrained_model("stabilityai/stable-audio-open-1.0")
|
23 |
+
# print("Model loaded successfully.")
|
24 |
+
# return model, model_config
|
25 |
+
|
26 |
+
# # Define the function to generate audio
|
27 |
+
# @spaces.GPU(duration=120)
|
28 |
+
# def generate_audio(prompt, bpm, seconds_total):
|
29 |
+
# device = "cuda" if torch.cuda.is_available() else "cpu"
|
30 |
+
|
31 |
+
# # Download model
|
32 |
+
# model, model_config = load_model()
|
33 |
+
# sample_rate = model_config["sample_rate"]
|
34 |
+
# sample_size = model_config["sample_size"]
|
35 |
+
|
36 |
+
# model = model.to(device)
|
37 |
+
|
38 |
+
# # Set up text and timing conditioning
|
39 |
+
# conditioning = [{
|
40 |
+
# "prompt": f"{bpm} BPM {prompt}",
|
41 |
+
# "seconds_start": 0,
|
42 |
+
# "seconds_total": seconds_total
|
43 |
+
# }]
|
44 |
+
|
45 |
+
# # Generate stereo audio
|
46 |
+
# output = generate_diffusion_cond(
|
47 |
+
# model,
|
48 |
+
# steps=100,
|
49 |
+
# cfg_scale=7,
|
50 |
+
# conditioning=conditioning,
|
51 |
+
# sample_size=sample_size,
|
52 |
+
# sigma_min=0.3,
|
53 |
+
# sigma_max=500,
|
54 |
+
# sampler_type="dpmpp-3m-sde",
|
55 |
+
# device=device
|
56 |
+
# )
|
57 |
+
|
58 |
+
# # Rearrange audio batch to a single sequence
|
59 |
+
# output = rearrange(output, "b d n -> d (b n)")
|
60 |
+
|
61 |
+
# # Peak normalize, clip, convert to int16, and save to file
|
62 |
+
# output = output.to(torch.float32).div(torch.max(torch.abs(output))).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
|
63 |
+
|
64 |
+
# output_path = "output.wav"
|
65 |
+
# torchaudio.save(output_path, output, sample_rate)
|
66 |
+
|
67 |
+
# return output_path
|
68 |
+
|
69 |
+
# # Define the Gradio interface
|
70 |
+
# iface = gr.Interface(
|
71 |
+
# fn=generate_audio,
|
72 |
+
# inputs=[
|
73 |
+
# gr.Textbox(label="Prompt", placeholder="Enter the description of the audio (e.g., tech house drum loop)"),
|
74 |
+
# gr.Number(label="BPM", value=128),
|
75 |
+
# gr.Number(label="Duration (seconds)", value=30)
|
76 |
+
# ],
|
77 |
+
# outputs=gr.Audio(label="Generated Audio"),
|
78 |
+
# title="Stable Audio Generation",
|
79 |
+
# description="Generate audio based on a text prompt using stable audio tools.",
|
80 |
+
# )
|
81 |
+
|
82 |
+
# # Launch the interface
|
83 |
+
# iface.launch()
|
84 |
+
|
85 |
import torch
|
86 |
import torchaudio
|
87 |
from einops import rearrange
|
|
|
94 |
from stable_audio_tools import get_pretrained_model
|
95 |
from stable_audio_tools.inference.generation import generate_diffusion_cond
|
96 |
|
|
|
|
|
|
|
|
|
|
|
|
|
97 |
# Load the model outside of the GPU-decorated function
|
98 |
def load_model():
|
99 |
print("Loading model...")
|
|
|
101 |
print("Model loaded successfully.")
|
102 |
return model, model_config
|
103 |
|
104 |
+
# Function to set up, generate, and process the audio
|
105 |
+
@spaces.GPU(duration=120) # Allocate GPU only when this function is called
|
106 |
+
def generate_audio(prompt, seconds_total=30, steps=100, cfg_scale=7):
|
107 |
+
print(f"Prompt received: {prompt}")
|
108 |
+
print(f"Settings: Duration={seconds_total}s, Steps={steps}, CFG Scale={cfg_scale}")
|
109 |
+
|
110 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
111 |
+
print(f"Using device: {device}")
|
112 |
|
113 |
+
# Fetch the Hugging Face token from the environment variable
|
114 |
+
hf_token = os.getenv('HF_TOKEN')
|
115 |
+
print(f"Hugging Face token: {hf_token}")
|
116 |
+
|
117 |
+
# Use pre-loaded model and configuration
|
118 |
model, model_config = load_model()
|
119 |
sample_rate = model_config["sample_rate"]
|
120 |
sample_size = model_config["sample_size"]
|
121 |
|
122 |
+
print(f"Sample rate: {sample_rate}, Sample size: {sample_size}")
|
123 |
+
|
124 |
model = model.to(device)
|
125 |
+
print("Model moved to device.")
|
126 |
|
127 |
# Set up text and timing conditioning
|
128 |
conditioning = [{
|
129 |
+
"prompt": prompt,
|
130 |
"seconds_start": 0,
|
131 |
"seconds_total": seconds_total
|
132 |
}]
|
133 |
+
print(f"Conditioning: {conditioning}")
|
134 |
|
135 |
# Generate stereo audio
|
136 |
+
print("Generating audio...")
|
137 |
output = generate_diffusion_cond(
|
138 |
model,
|
139 |
+
steps=steps,
|
140 |
+
cfg_scale=cfg_scale,
|
141 |
conditioning=conditioning,
|
142 |
sample_size=sample_size,
|
143 |
sigma_min=0.3,
|
|
|
145 |
sampler_type="dpmpp-3m-sde",
|
146 |
device=device
|
147 |
)
|
148 |
+
print("Audio generated.")
|
149 |
|
150 |
# Rearrange audio batch to a single sequence
|
151 |
output = rearrange(output, "b d n -> d (b n)")
|
152 |
+
print("Audio rearranged.")
|
153 |
|
154 |
+
# Peak normalize, clip, convert to int16
|
155 |
output = output.to(torch.float32).div(torch.max(torch.abs(output))).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
|
156 |
+
print("Audio normalized and converted.")
|
157 |
+
|
158 |
+
# Generate a unique filename for the output
|
159 |
+
unique_filename = f"output_{uuid.uuid4().hex}.wav"
|
160 |
+
print(f"Saving audio to file: {unique_filename}")
|
161 |
+
|
162 |
+
# Save to file
|
163 |
+
torchaudio.save(unique_filename, output, sample_rate)
|
164 |
+
print(f"Audio saved: {unique_filename}")
|
165 |
|
166 |
+
# Return the path to the generated audio file
|
167 |
+
return unique_filename
|
168 |
+
|
169 |
+
# Setting up the Gradio Interface
|
170 |
+
interface = gr.Interface(
|
171 |
fn=generate_audio,
|
172 |
inputs=[
|
173 |
+
gr.Textbox(label="Prompt", placeholder="Enter your text prompt here"),
|
174 |
+
gr.Slider(0, 47, value=30, label="Duration in Seconds"),
|
175 |
+
gr.Slider(10, 150, value=100, step=10, label="Number of Diffusion Steps"),
|
176 |
+
gr.Slider(1, 15, value=7, step=0.1, label="CFG Scale")
|
177 |
],
|
178 |
+
outputs=gr.Audio(type="filepath", label="Generated Audio"),
|
179 |
+
title="Stable Audio Generator",
|
180 |
+
description="Generate variable-length stereo audio at 44.1kHz from text prompts using Stable Audio Open 1.0."
|
181 |
)
|
182 |
|
183 |
+
# Pre-load the model to avoid multiprocessing issues
|
184 |
+
model, model_config = load_model()
|
185 |
+
|
186 |
+
# Launch the Interface
|
187 |
+
interface.launch()
|
188 |
+
|