Create generate.py
Browse files- generate.py +38 -0
generate.py
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from transformers import AutoProcessor, MusicgenForConditionalGeneration
|
| 3 |
+
import numpy as np
|
| 4 |
+
|
| 5 |
+
MODEL_ID = "facebook/musicgen-small"
|
| 6 |
+
device = "cpu"
|
| 7 |
+
|
| 8 |
+
print(f"[*] Initializing Engine: Loading {MODEL_ID}...")
|
| 9 |
+
processor = AutoProcessor.from_pretrained(MODEL_ID)
|
| 10 |
+
model = MusicgenForConditionalGeneration.from_pretrained(MODEL_ID, torch_dtype=torch.float32)
|
| 11 |
+
model.to(device)
|
| 12 |
+
print("[+] Engine Ready.")
|
| 13 |
+
|
| 14 |
+
def generate_music(prompt, duration):
|
| 15 |
+
if not prompt:
|
| 16 |
+
return None
|
| 17 |
+
|
| 18 |
+
print(f"[#] Generating: {prompt} ({duration}s)")
|
| 19 |
+
try:
|
| 20 |
+
duration = min(int(duration), 30)
|
| 21 |
+
inputs = processor(text=[prompt], padding=True, return_tensors="pt").to(device)
|
| 22 |
+
max_tokens = int(duration * 50)
|
| 23 |
+
|
| 24 |
+
with torch.no_grad():
|
| 25 |
+
audio_values = model.generate(
|
| 26 |
+
**inputs,
|
| 27 |
+
max_new_tokens=max_tokens,
|
| 28 |
+
do_sample=True,
|
| 29 |
+
guidance_scale=3.0
|
| 30 |
+
)
|
| 31 |
+
|
| 32 |
+
sampling_rate = model.config.audio_encoder.sampling_rate
|
| 33 |
+
audio_data = audio_values[0, 0].cpu().numpy()
|
| 34 |
+
return sampling_rate, audio_data
|
| 35 |
+
except Exception as e:
|
| 36 |
+
print(f"[!] Error: {str(e)}")
|
| 37 |
+
return None, None
|
| 38 |
+
|