Spaces:
Running
on
Zero
Running
on
Zero
test
Browse files
diffrhythm/infer/infer.py
CHANGED
@@ -8,6 +8,7 @@ from tqdm import tqdm
|
|
8 |
import random
|
9 |
import numpy as np
|
10 |
import time
|
|
|
11 |
|
12 |
from diffrhythm.infer.infer_utils import (
|
13 |
get_reference_latent,
|
@@ -17,6 +18,7 @@ from diffrhythm.infer.infer_utils import (
|
|
17 |
get_negative_style_prompt
|
18 |
)
|
19 |
|
|
|
20 |
def decode_audio(latents, vae_model, chunked=False, overlap=32, chunk_size=128):
|
21 |
downsampling_ratio = 2048
|
22 |
io_channels = 2
|
@@ -72,6 +74,7 @@ def decode_audio(latents, vae_model, chunked=False, overlap=32, chunk_size=128):
|
|
72 |
y_final[:,:,t_start:t_end] = y_chunk[:,:,chunk_start:chunk_end]
|
73 |
return y_final
|
74 |
|
|
|
75 |
def inference(cfm_model, vae_model, cond, text, duration, style_prompt, negative_style_prompt, steps, sway_sampling_coef, start_time):
|
76 |
# import pdb; pdb.set_trace()
|
77 |
s_t = time.time()
|
@@ -100,7 +103,7 @@ def inference(cfm_model, vae_model, cond, text, duration, style_prompt, negative
|
|
100 |
output = rearrange(output, "b d n -> d (b n)")
|
101 |
output_tensor = output.to(torch.float32).div(torch.max(torch.abs(output))).clamp(-1, 1).cpu()
|
102 |
output_np = output_tensor.numpy().T.astype(np.float32)
|
103 |
-
print(f"**** vae time : {time.
|
104 |
print(output_np.mean(), output_np.min(), output_np.max(), output_np.std())
|
105 |
return (44100, output_np)
|
106 |
|
|
|
8 |
import random
|
9 |
import numpy as np
|
10 |
import time
|
11 |
+
import spaces
|
12 |
|
13 |
from diffrhythm.infer.infer_utils import (
|
14 |
get_reference_latent,
|
|
|
18 |
get_negative_style_prompt
|
19 |
)
|
20 |
|
21 |
+
@spaces.GPU
|
22 |
def decode_audio(latents, vae_model, chunked=False, overlap=32, chunk_size=128):
|
23 |
downsampling_ratio = 2048
|
24 |
io_channels = 2
|
|
|
74 |
y_final[:,:,t_start:t_end] = y_chunk[:,:,chunk_start:chunk_end]
|
75 |
return y_final
|
76 |
|
77 |
+
@spaces.GPU
|
78 |
def inference(cfm_model, vae_model, cond, text, duration, style_prompt, negative_style_prompt, steps, sway_sampling_coef, start_time):
|
79 |
# import pdb; pdb.set_trace()
|
80 |
s_t = time.time()
|
|
|
103 |
output = rearrange(output, "b d n -> d (b n)")
|
104 |
output_tensor = output.to(torch.float32).div(torch.max(torch.abs(output))).clamp(-1, 1).cpu()
|
105 |
output_np = output_tensor.numpy().T.astype(np.float32)
|
106 |
+
print(f"**** vae time : {time.time()-e_t} ****")
|
107 |
print(output_np.mean(), output_np.min(), output_np.max(), output_np.std())
|
108 |
return (44100, output_np)
|
109 |
|
diffrhythm/infer/infer_utils.py
CHANGED
@@ -35,7 +35,9 @@ def prepare_model(device):
|
|
35 |
# prepare vae
|
36 |
vae_ckpt_path = hf_hub_download(repo_id="ASLP-lab/DiffRhythm-vae", filename="vae_model.pt")
|
37 |
vae = torch.jit.load(vae_ckpt_path).to(device)
|
38 |
-
|
|
|
|
|
39 |
return cfm, tokenizer, muq, vae
|
40 |
|
41 |
|
|
|
35 |
# prepare vae
|
36 |
vae_ckpt_path = hf_hub_download(repo_id="ASLP-lab/DiffRhythm-vae", filename="vae_model.pt")
|
37 |
vae = torch.jit.load(vae_ckpt_path).to(device)
|
38 |
+
print("********* vae.parameters() ", next(vae.parameters()).dtype)
|
39 |
+
vae = vae.half()
|
40 |
+
print("********* vae half .parameters() ", next(vae.parameters()).dtype)
|
41 |
return cfm, tokenizer, muq, vae
|
42 |
|
43 |
|