Spaces:
Runtime error
Runtime error
[feat] add repainting & edit
Browse files- pipeline_ace_step.py +401 -50
- ui/components.py +254 -13
pipeline_ace_step.py
CHANGED
@@ -4,6 +4,7 @@ import os
|
|
4 |
import re
|
5 |
|
6 |
import torch
|
|
|
7 |
from loguru import logger
|
8 |
from tqdm import tqdm
|
9 |
import json
|
@@ -22,11 +23,11 @@ from models.ace_step_transformer import ACEStepTransformer2DModel
|
|
22 |
from models.lyrics_utils.lyric_tokenizer import VoiceBpeTokenizer
|
23 |
from apg_guidance import apg_forward, MomentumBuffer, cfg_forward, cfg_zero_star, cfg_double_condition_forward
|
24 |
import torchaudio
|
|
|
|
|
25 |
torch.backends.cudnn.benchmark = False
|
26 |
torch.set_float32_matmul_precision('high')
|
27 |
-
|
28 |
-
# Enable TF32 for faster training on Ampere GPUs,
|
29 |
-
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
|
30 |
torch.backends.cuda.matmul.allow_tf32 = True
|
31 |
|
32 |
|
@@ -49,9 +50,10 @@ def ensure_directory_exists(directory):
|
|
49 |
REPO_ID = "ACE-Step/ACE-Step-v1-3.5B"
|
50 |
|
51 |
|
|
|
52 |
class ACEStepPipeline:
|
53 |
|
54 |
-
def __init__(self, checkpoint_dir=None, device_id=0, dtype="bfloat16", text_encoder_checkpoint_path=None, persistent_storage_path=None, **kwargs):
|
55 |
if checkpoint_dir is None:
|
56 |
if persistent_storage_path is None:
|
57 |
checkpoint_dir = os.path.join(os.path.dirname(__file__), "checkpoints")
|
@@ -64,6 +66,7 @@ class ACEStepPipeline:
|
|
64 |
self.dtype = torch.bfloat16 if dtype == "bfloat16" else torch.float32
|
65 |
self.device = device
|
66 |
self.loaded = False
|
|
|
67 |
|
68 |
def load_checkpoint(self, checkpoint_dir=None):
|
69 |
device = self.device
|
@@ -157,9 +160,10 @@ class ACEStepPipeline:
|
|
157 |
self.loaded = True
|
158 |
|
159 |
# compile
|
160 |
-
|
161 |
-
|
162 |
-
|
|
|
163 |
|
164 |
def get_text_embeddings(self, texts, device, text_max_length=256):
|
165 |
inputs = self.text_tokenizer(texts, return_tensors="pt", padding=True, truncation=True, max_length=text_max_length)
|
@@ -226,7 +230,7 @@ class ACEStepPipeline:
|
|
226 |
|
227 |
def get_lang(self, text):
|
228 |
language = "en"
|
229 |
-
try:
|
230 |
_ = self.lang_segment.getTexts(text)
|
231 |
langCounts = self.lang_segment.getCounts()
|
232 |
language = langCounts[0][0]
|
@@ -267,6 +271,250 @@ class ACEStepPipeline:
|
|
267 |
print("tokenize error", e, "for line", line, "major_language", lang)
|
268 |
return lyric_token_idx
|
269 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
270 |
@torch.no_grad()
|
271 |
def text2music_diffusion_process(
|
272 |
self,
|
@@ -296,13 +544,16 @@ class ACEStepPipeline:
|
|
296 |
add_retake_noise=False,
|
297 |
guidance_scale_text=0.0,
|
298 |
guidance_scale_lyric=0.0,
|
|
|
|
|
|
|
299 |
):
|
300 |
|
301 |
logger.info("cfg_type: {}, guidance_scale: {}, omega_scale: {}".format(cfg_type, guidance_scale, omega_scale))
|
302 |
do_classifier_free_guidance = True
|
303 |
if guidance_scale == 0.0 or guidance_scale == 1.0:
|
304 |
do_classifier_free_guidance = False
|
305 |
-
|
306 |
do_double_condition_guidance = False
|
307 |
if guidance_scale_text is not None and guidance_scale_text > 1.0 and guidance_scale_lyric is not None and guidance_scale_lyric > 1.0:
|
308 |
do_double_condition_guidance = True
|
@@ -322,7 +573,10 @@ class ACEStepPipeline:
|
|
322 |
num_train_timesteps=1000,
|
323 |
shift=3.0,
|
324 |
)
|
|
|
325 |
frame_length = int(duration * 44100 / 512 / 8)
|
|
|
|
|
326 |
|
327 |
if len(oss_steps) > 0:
|
328 |
infer_steps = max(oss_steps)
|
@@ -337,16 +591,30 @@ class ACEStepPipeline:
|
|
337 |
logger.info(f"oss_steps: {oss_steps}, num_inference_steps: {num_inference_steps} after remapping to timesteps {timesteps}")
|
338 |
else:
|
339 |
timesteps, num_inference_steps = retrieve_timesteps(scheduler, num_inference_steps=infer_steps, device=device, timesteps=None)
|
340 |
-
|
341 |
target_latents = randn_tensor(shape=(bsz, 8, 16, frame_length), generator=random_generators, device=device, dtype=dtype)
|
|
|
|
|
342 |
if add_retake_noise:
|
343 |
retake_variance = torch.tensor(retake_variance * math.pi/2).to(device).to(dtype)
|
344 |
retake_latents = randn_tensor(shape=(bsz, 8, 16, frame_length), generator=retake_random_generators, device=device, dtype=dtype)
|
|
|
|
|
|
|
|
|
|
|
345 |
# to make sure mean = 0, std = 1
|
346 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
347 |
|
348 |
attention_mask = torch.ones(bsz, frame_length, device=device, dtype=dtype)
|
349 |
-
|
350 |
# guidance interval
|
351 |
start_idx = int(num_inference_steps * ((1 - guidance_interval) / 2))
|
352 |
end_idx = int(num_inference_steps * (guidance_interval / 2 + 0.5))
|
@@ -356,20 +624,20 @@ class ACEStepPipeline:
|
|
356 |
|
357 |
def forward_encoder_with_temperature(self, inputs, tau=0.01, l_min=4, l_max=6):
|
358 |
handlers = []
|
359 |
-
|
360 |
def hook(module, input, output):
|
361 |
output[:] *= tau
|
362 |
return output
|
363 |
-
|
364 |
for i in range(l_min, l_max):
|
365 |
handler = self.ace_step_transformer.lyric_encoder.encoders[i].self_attn.linear_q.register_forward_hook(hook)
|
366 |
handlers.append(handler)
|
367 |
-
|
368 |
encoder_hidden_states, encoder_hidden_mask = self.ace_step_transformer.encode(**inputs)
|
369 |
-
|
370 |
for hook in handlers:
|
371 |
hook.remove()
|
372 |
-
|
373 |
return encoder_hidden_states
|
374 |
|
375 |
# P(speaker, text, lyric)
|
@@ -402,7 +670,7 @@ class ACEStepPipeline:
|
|
402 |
torch.zeros_like(lyric_token_ids),
|
403 |
lyric_mask,
|
404 |
)
|
405 |
-
|
406 |
encoder_hidden_states_no_lyric = None
|
407 |
if do_double_condition_guidance:
|
408 |
# P(null_speaker, text, lyric_weaker)
|
@@ -429,11 +697,11 @@ class ACEStepPipeline:
|
|
429 |
|
430 |
def forward_diffusion_with_temperature(self, hidden_states, timestep, inputs, tau=0.01, l_min=15, l_max=20):
|
431 |
handlers = []
|
432 |
-
|
433 |
def hook(module, input, output):
|
434 |
output[:] *= tau
|
435 |
return output
|
436 |
-
|
437 |
for i in range(l_min, l_max):
|
438 |
handler = self.ace_step_transformer.transformer_blocks[i].attn.to_q.register_forward_hook(hook)
|
439 |
handlers.append(handler)
|
@@ -441,10 +709,10 @@ class ACEStepPipeline:
|
|
441 |
handlers.append(handler)
|
442 |
|
443 |
sample = self.ace_step_transformer.decode(hidden_states=hidden_states, timestep=timestep, **inputs).sample
|
444 |
-
|
445 |
for hook in handlers:
|
446 |
hook.remove()
|
447 |
-
|
448 |
return sample
|
449 |
|
450 |
for i, t in tqdm(enumerate(timesteps), total=num_inference_steps):
|
@@ -551,7 +819,13 @@ class ACEStepPipeline:
|
|
551 |
).sample
|
552 |
|
553 |
target_latents = scheduler.step(model_output=noise_pred, timestep=t, sample=target_latents, return_dict=False, omega=omega_scale)[0]
|
|
|
|
|
|
|
|
|
|
|
554 |
|
|
|
555 |
return target_latents
|
556 |
|
557 |
def latents2audio(self, latents, target_wav_duration_second=30, sample_rate=48000, save_path=None, format="flac"):
|
@@ -570,7 +844,7 @@ class ACEStepPipeline:
|
|
570 |
def save_wav_file(self, target_wav, idx, save_path=None, sample_rate=48000, format="flac"):
|
571 |
if save_path is None:
|
572 |
logger.warning("save_path is None, using default path ./outputs/")
|
573 |
-
base_path = f"./outputs
|
574 |
ensure_directory_exists(base_path)
|
575 |
else:
|
576 |
base_path = save_path
|
@@ -581,6 +855,16 @@ class ACEStepPipeline:
|
|
581 |
torchaudio.save(output_path_flac, target_wav, sample_rate=sample_rate, format=format)
|
582 |
return output_path_flac
|
583 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
584 |
def __call__(
|
585 |
self,
|
586 |
audio_duration: float = 60.0,
|
@@ -604,6 +888,14 @@ class ACEStepPipeline:
|
|
604 |
retake_seeds: list = None,
|
605 |
retake_variance: float = 0.5,
|
606 |
task: str = "text2music",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
607 |
save_path: str = None,
|
608 |
format: str = "flac",
|
609 |
batch_size: int = 1,
|
@@ -626,7 +918,7 @@ class ACEStepPipeline:
|
|
626 |
oss_steps = list(map(int, oss_steps.split(",")))
|
627 |
else:
|
628 |
oss_steps = []
|
629 |
-
|
630 |
texts = [prompt]
|
631 |
encoder_text_hidden_states, text_attention_mask = self.get_text_embeddings(texts, self.device)
|
632 |
encoder_text_hidden_states = encoder_text_hidden_states.repeat(batch_size, 1, 1)
|
@@ -657,32 +949,83 @@ class ACEStepPipeline:
|
|
657 |
preprocess_time_cost = end_time - start_time
|
658 |
start_time = end_time
|
659 |
|
660 |
-
|
661 |
-
|
662 |
-
|
663 |
-
|
664 |
-
|
665 |
-
|
666 |
-
|
667 |
-
|
668 |
-
|
669 |
-
|
670 |
-
|
671 |
-
|
672 |
-
|
673 |
-
|
674 |
-
|
675 |
-
|
676 |
-
|
677 |
-
|
678 |
-
|
679 |
-
|
680 |
-
|
681 |
-
|
682 |
-
|
683 |
-
|
684 |
-
|
685 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
686 |
|
687 |
end_time = time.time()
|
688 |
diffusion_time_cost = end_time - start_time
|
@@ -726,6 +1069,14 @@ class ACEStepPipeline:
|
|
726 |
"retake_variance": retake_variance,
|
727 |
"guidance_scale_text": guidance_scale_text,
|
728 |
"guidance_scale_lyric": guidance_scale_lyric,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
729 |
}
|
730 |
# save input_params_json
|
731 |
for output_audio_path in output_paths:
|
|
|
4 |
import re
|
5 |
|
6 |
import torch
|
7 |
+
import torch.nn as nn
|
8 |
from loguru import logger
|
9 |
from tqdm import tqdm
|
10 |
import json
|
|
|
23 |
from models.lyrics_utils.lyric_tokenizer import VoiceBpeTokenizer
|
24 |
from apg_guidance import apg_forward, MomentumBuffer, cfg_forward, cfg_zero_star, cfg_double_condition_forward
|
25 |
import torchaudio
|
26 |
+
|
27 |
+
|
28 |
torch.backends.cudnn.benchmark = False
|
29 |
torch.set_float32_matmul_precision('high')
|
30 |
+
torch.backends.cudnn.deterministic = True
|
|
|
|
|
31 |
torch.backends.cuda.matmul.allow_tf32 = True
|
32 |
|
33 |
|
|
|
50 |
REPO_ID = "ACE-Step/ACE-Step-v1-3.5B"
|
51 |
|
52 |
|
53 |
+
# class ACEStepPipeline(DiffusionPipeline):
|
54 |
class ACEStepPipeline:
|
55 |
|
56 |
+
def __init__(self, checkpoint_dir=None, device_id=0, dtype="bfloat16", text_encoder_checkpoint_path=None, persistent_storage_path=None, torch_compile=False, **kwargs):
|
57 |
if checkpoint_dir is None:
|
58 |
if persistent_storage_path is None:
|
59 |
checkpoint_dir = os.path.join(os.path.dirname(__file__), "checkpoints")
|
|
|
66 |
self.dtype = torch.bfloat16 if dtype == "bfloat16" else torch.float32
|
67 |
self.device = device
|
68 |
self.loaded = False
|
69 |
+
self.torch_compile = torch_compile
|
70 |
|
71 |
def load_checkpoint(self, checkpoint_dir=None):
|
72 |
device = self.device
|
|
|
160 |
self.loaded = True
|
161 |
|
162 |
# compile
|
163 |
+
if self.torch_compile:
|
164 |
+
self.music_dcae = torch.compile(self.music_dcae)
|
165 |
+
self.ace_step_transformer = torch.compile(self.ace_step_transformer)
|
166 |
+
self.text_encoder_model = torch.compile(self.text_encoder_model)
|
167 |
|
168 |
def get_text_embeddings(self, texts, device, text_max_length=256):
|
169 |
inputs = self.text_tokenizer(texts, return_tensors="pt", padding=True, truncation=True, max_length=text_max_length)
|
|
|
230 |
|
231 |
def get_lang(self, text):
|
232 |
language = "en"
|
233 |
+
try:
|
234 |
_ = self.lang_segment.getTexts(text)
|
235 |
langCounts = self.lang_segment.getCounts()
|
236 |
language = langCounts[0][0]
|
|
|
271 |
print("tokenize error", e, "for line", line, "major_language", lang)
|
272 |
return lyric_token_idx
|
273 |
|
274 |
+
def calc_v(
|
275 |
+
self,
|
276 |
+
zt_src,
|
277 |
+
zt_tar,
|
278 |
+
t,
|
279 |
+
encoder_text_hidden_states,
|
280 |
+
text_attention_mask,
|
281 |
+
target_encoder_text_hidden_states,
|
282 |
+
target_text_attention_mask,
|
283 |
+
speaker_embds,
|
284 |
+
target_speaker_embeds,
|
285 |
+
lyric_token_ids,
|
286 |
+
lyric_mask,
|
287 |
+
target_lyric_token_ids,
|
288 |
+
target_lyric_mask,
|
289 |
+
do_classifier_free_guidance=False,
|
290 |
+
guidance_scale=1.0,
|
291 |
+
target_guidance_scale=1.0,
|
292 |
+
cfg_type="apg",
|
293 |
+
attention_mask=None,
|
294 |
+
momentum_buffer=None,
|
295 |
+
momentum_buffer_tar=None,
|
296 |
+
return_src_pred=True
|
297 |
+
):
|
298 |
+
noise_pred_src = None
|
299 |
+
if return_src_pred:
|
300 |
+
src_latent_model_input = torch.cat([zt_src, zt_src]) if do_classifier_free_guidance else zt_src
|
301 |
+
timestep = t.expand(src_latent_model_input.shape[0])
|
302 |
+
# source
|
303 |
+
noise_pred_src = self.ace_step_transformer(
|
304 |
+
hidden_states=src_latent_model_input,
|
305 |
+
attention_mask=attention_mask,
|
306 |
+
encoder_text_hidden_states=encoder_text_hidden_states,
|
307 |
+
text_attention_mask=text_attention_mask,
|
308 |
+
speaker_embeds=speaker_embds,
|
309 |
+
lyric_token_idx=lyric_token_ids,
|
310 |
+
lyric_mask=lyric_mask,
|
311 |
+
timestep=timestep,
|
312 |
+
).sample
|
313 |
+
|
314 |
+
if do_classifier_free_guidance:
|
315 |
+
noise_pred_with_cond_src, noise_pred_uncond_src = noise_pred_src.chunk(2)
|
316 |
+
if cfg_type == "apg":
|
317 |
+
noise_pred_src = apg_forward(
|
318 |
+
pred_cond=noise_pred_with_cond_src,
|
319 |
+
pred_uncond=noise_pred_uncond_src,
|
320 |
+
guidance_scale=guidance_scale,
|
321 |
+
momentum_buffer=momentum_buffer,
|
322 |
+
)
|
323 |
+
elif cfg_type == "cfg":
|
324 |
+
noise_pred_src = cfg_forward(
|
325 |
+
cond_output=noise_pred_with_cond_src,
|
326 |
+
uncond_output=noise_pred_uncond_src,
|
327 |
+
cfg_strength=guidance_scale,
|
328 |
+
)
|
329 |
+
|
330 |
+
tar_latent_model_input = torch.cat([zt_tar, zt_tar]) if do_classifier_free_guidance else zt_tar
|
331 |
+
timestep = t.expand(tar_latent_model_input.shape[0])
|
332 |
+
# target
|
333 |
+
noise_pred_tar = self.ace_step_transformer(
|
334 |
+
hidden_states=tar_latent_model_input,
|
335 |
+
attention_mask=attention_mask,
|
336 |
+
encoder_text_hidden_states=target_encoder_text_hidden_states,
|
337 |
+
text_attention_mask=target_text_attention_mask,
|
338 |
+
speaker_embeds=target_speaker_embeds,
|
339 |
+
lyric_token_idx=target_lyric_token_ids,
|
340 |
+
lyric_mask=target_lyric_mask,
|
341 |
+
timestep=timestep,
|
342 |
+
).sample
|
343 |
+
|
344 |
+
if do_classifier_free_guidance:
|
345 |
+
noise_pred_with_cond_tar, noise_pred_uncond_tar = noise_pred_tar.chunk(2)
|
346 |
+
if cfg_type == "apg":
|
347 |
+
noise_pred_tar = apg_forward(
|
348 |
+
pred_cond=noise_pred_with_cond_tar,
|
349 |
+
pred_uncond=noise_pred_uncond_tar,
|
350 |
+
guidance_scale=target_guidance_scale,
|
351 |
+
momentum_buffer=momentum_buffer_tar,
|
352 |
+
)
|
353 |
+
elif cfg_type == "cfg":
|
354 |
+
noise_pred_tar = cfg_forward(
|
355 |
+
cond_output=noise_pred_with_cond_tar,
|
356 |
+
uncond_output=noise_pred_uncond_tar,
|
357 |
+
cfg_strength=target_guidance_scale,
|
358 |
+
)
|
359 |
+
return noise_pred_src, noise_pred_tar
|
360 |
+
|
361 |
+
@torch.no_grad()
|
362 |
+
def flowedit_diffusion_process(
|
363 |
+
self,
|
364 |
+
encoder_text_hidden_states,
|
365 |
+
text_attention_mask,
|
366 |
+
speaker_embds,
|
367 |
+
lyric_token_ids,
|
368 |
+
lyric_mask,
|
369 |
+
target_encoder_text_hidden_states,
|
370 |
+
target_text_attention_mask,
|
371 |
+
target_speaker_embeds,
|
372 |
+
target_lyric_token_ids,
|
373 |
+
target_lyric_mask,
|
374 |
+
src_latents,
|
375 |
+
random_generators=None,
|
376 |
+
infer_steps=60,
|
377 |
+
guidance_scale=15.0,
|
378 |
+
n_min=0,
|
379 |
+
n_max=1.0,
|
380 |
+
n_avg=1,
|
381 |
+
):
|
382 |
+
|
383 |
+
do_classifier_free_guidance = True
|
384 |
+
if guidance_scale == 0.0 or guidance_scale == 1.0:
|
385 |
+
do_classifier_free_guidance = False
|
386 |
+
|
387 |
+
target_guidance_scale = guidance_scale
|
388 |
+
device = encoder_text_hidden_states.device
|
389 |
+
dtype = encoder_text_hidden_states.dtype
|
390 |
+
bsz = encoder_text_hidden_states.shape[0]
|
391 |
+
|
392 |
+
scheduler = FlowMatchEulerDiscreteScheduler(
|
393 |
+
num_train_timesteps=1000,
|
394 |
+
shift=3.0,
|
395 |
+
)
|
396 |
+
|
397 |
+
T_steps = infer_steps
|
398 |
+
frame_length = src_latents.shape[-1]
|
399 |
+
attention_mask = torch.ones(bsz, frame_length, device=device, dtype=dtype)
|
400 |
+
|
401 |
+
timesteps, T_steps = retrieve_timesteps(scheduler, T_steps, device, timesteps=None)
|
402 |
+
|
403 |
+
if do_classifier_free_guidance:
|
404 |
+
attention_mask = torch.cat([attention_mask] * 2, dim=0)
|
405 |
+
|
406 |
+
encoder_text_hidden_states = torch.cat([encoder_text_hidden_states, torch.zeros_like(encoder_text_hidden_states)], 0)
|
407 |
+
text_attention_mask = torch.cat([text_attention_mask] * 2, dim=0)
|
408 |
+
|
409 |
+
target_encoder_text_hidden_states = torch.cat([target_encoder_text_hidden_states, torch.zeros_like(target_encoder_text_hidden_states)], 0)
|
410 |
+
target_text_attention_mask = torch.cat([target_text_attention_mask] * 2, dim=0)
|
411 |
+
|
412 |
+
speaker_embds = torch.cat([speaker_embds, torch.zeros_like(speaker_embds)], 0)
|
413 |
+
target_speaker_embeds = torch.cat([target_speaker_embeds, torch.zeros_like(target_speaker_embeds)], 0)
|
414 |
+
|
415 |
+
lyric_token_ids = torch.cat([lyric_token_ids, torch.zeros_like(lyric_token_ids)], 0)
|
416 |
+
lyric_mask = torch.cat([lyric_mask, torch.zeros_like(lyric_mask)], 0)
|
417 |
+
|
418 |
+
target_lyric_token_ids = torch.cat([target_lyric_token_ids, torch.zeros_like(target_lyric_token_ids)], 0)
|
419 |
+
target_lyric_mask = torch.cat([target_lyric_mask, torch.zeros_like(target_lyric_mask)], 0)
|
420 |
+
|
421 |
+
momentum_buffer = MomentumBuffer()
|
422 |
+
momentum_buffer_tar = MomentumBuffer()
|
423 |
+
x_src = src_latents
|
424 |
+
zt_edit = x_src.clone()
|
425 |
+
xt_tar = None
|
426 |
+
n_min = int(infer_steps * n_min)
|
427 |
+
n_max = int(infer_steps * n_max)
|
428 |
+
|
429 |
+
logger.info("flowedit start from {} to {}".format(n_min, n_max))
|
430 |
+
|
431 |
+
for i, t in tqdm(enumerate(timesteps), total=T_steps):
|
432 |
+
|
433 |
+
if i < n_min:
|
434 |
+
continue
|
435 |
+
|
436 |
+
t_i = t/1000
|
437 |
+
|
438 |
+
if i+1 < len(timesteps):
|
439 |
+
t_im1 = (timesteps[i+1])/1000
|
440 |
+
else:
|
441 |
+
t_im1 = torch.zeros_like(t_i).to(t_i.device)
|
442 |
+
|
443 |
+
if i < n_max:
|
444 |
+
# Calculate the average of the V predictions
|
445 |
+
V_delta_avg = torch.zeros_like(x_src)
|
446 |
+
for k in range(n_avg):
|
447 |
+
fwd_noise = randn_tensor(shape=x_src.shape, generator=random_generators, device=device, dtype=dtype)
|
448 |
+
|
449 |
+
zt_src = (1 - t_i) * x_src + (t_i) * fwd_noise
|
450 |
+
|
451 |
+
zt_tar = zt_edit + zt_src - x_src
|
452 |
+
|
453 |
+
Vt_src, Vt_tar = self.calc_v(
|
454 |
+
zt_src=zt_src,
|
455 |
+
zt_tar=zt_tar,
|
456 |
+
t=t,
|
457 |
+
encoder_text_hidden_states=encoder_text_hidden_states,
|
458 |
+
text_attention_mask=text_attention_mask,
|
459 |
+
target_encoder_text_hidden_states=target_encoder_text_hidden_states,
|
460 |
+
target_text_attention_mask=target_text_attention_mask,
|
461 |
+
speaker_embds=speaker_embds,
|
462 |
+
target_speaker_embeds=target_speaker_embeds,
|
463 |
+
lyric_token_ids=lyric_token_ids,
|
464 |
+
lyric_mask=lyric_mask,
|
465 |
+
target_lyric_token_ids=target_lyric_token_ids,
|
466 |
+
target_lyric_mask=target_lyric_mask,
|
467 |
+
do_classifier_free_guidance=do_classifier_free_guidance,
|
468 |
+
guidance_scale=guidance_scale,
|
469 |
+
target_guidance_scale=target_guidance_scale,
|
470 |
+
attention_mask=attention_mask,
|
471 |
+
momentum_buffer=momentum_buffer
|
472 |
+
)
|
473 |
+
V_delta_avg += (1 / n_avg) * (Vt_tar - Vt_src) # - (hfg-1)*( x_src))
|
474 |
+
|
475 |
+
# propagate direct ODE
|
476 |
+
zt_edit = zt_edit.to(torch.float32)
|
477 |
+
zt_edit = zt_edit + (t_im1 - t_i) * V_delta_avg
|
478 |
+
zt_edit = zt_edit.to(V_delta_avg.dtype)
|
479 |
+
else: # i >= T_steps-n_min # regular sampling for last n_min steps
|
480 |
+
if i == n_max:
|
481 |
+
fwd_noise = randn_tensor(shape=x_src.shape, generator=random_generators, device=device, dtype=dtype)
|
482 |
+
scheduler._init_step_index(t)
|
483 |
+
sigma = scheduler.sigmas[scheduler.step_index]
|
484 |
+
xt_src = sigma * fwd_noise + (1.0 - sigma) * x_src
|
485 |
+
xt_tar = zt_edit + xt_src - x_src
|
486 |
+
|
487 |
+
_, Vt_tar = self.calc_v(
|
488 |
+
zt_src=None,
|
489 |
+
zt_tar=xt_tar,
|
490 |
+
t=t,
|
491 |
+
encoder_text_hidden_states=encoder_text_hidden_states,
|
492 |
+
text_attention_mask=text_attention_mask,
|
493 |
+
target_encoder_text_hidden_states=target_encoder_text_hidden_states,
|
494 |
+
target_text_attention_mask=target_text_attention_mask,
|
495 |
+
speaker_embds=speaker_embds,
|
496 |
+
target_speaker_embeds=target_speaker_embeds,
|
497 |
+
lyric_token_ids=lyric_token_ids,
|
498 |
+
lyric_mask=lyric_mask,
|
499 |
+
target_lyric_token_ids=target_lyric_token_ids,
|
500 |
+
target_lyric_mask=target_lyric_mask,
|
501 |
+
do_classifier_free_guidance=do_classifier_free_guidance,
|
502 |
+
guidance_scale=guidance_scale,
|
503 |
+
target_guidance_scale=target_guidance_scale,
|
504 |
+
attention_mask=attention_mask,
|
505 |
+
momentum_buffer_tar=momentum_buffer_tar,
|
506 |
+
return_src_pred=False,
|
507 |
+
)
|
508 |
+
|
509 |
+
dtype = Vt_tar.dtype
|
510 |
+
xt_tar = xt_tar.to(torch.float32)
|
511 |
+
prev_sample = xt_tar + (t_im1 - t_i) * Vt_tar
|
512 |
+
prev_sample = prev_sample.to(dtype)
|
513 |
+
xt_tar = prev_sample
|
514 |
+
|
515 |
+
target_latents = zt_edit if xt_tar is None else xt_tar
|
516 |
+
return target_latents
|
517 |
+
|
518 |
@torch.no_grad()
|
519 |
def text2music_diffusion_process(
|
520 |
self,
|
|
|
544 |
add_retake_noise=False,
|
545 |
guidance_scale_text=0.0,
|
546 |
guidance_scale_lyric=0.0,
|
547 |
+
repaint_start=0,
|
548 |
+
repaint_end=0,
|
549 |
+
src_latents=None,
|
550 |
):
|
551 |
|
552 |
logger.info("cfg_type: {}, guidance_scale: {}, omega_scale: {}".format(cfg_type, guidance_scale, omega_scale))
|
553 |
do_classifier_free_guidance = True
|
554 |
if guidance_scale == 0.0 or guidance_scale == 1.0:
|
555 |
do_classifier_free_guidance = False
|
556 |
+
|
557 |
do_double_condition_guidance = False
|
558 |
if guidance_scale_text is not None and guidance_scale_text > 1.0 and guidance_scale_lyric is not None and guidance_scale_lyric > 1.0:
|
559 |
do_double_condition_guidance = True
|
|
|
573 |
num_train_timesteps=1000,
|
574 |
shift=3.0,
|
575 |
)
|
576 |
+
|
577 |
frame_length = int(duration * 44100 / 512 / 8)
|
578 |
+
if src_latents is not None:
|
579 |
+
frame_length = src_latents.shape[-1]
|
580 |
|
581 |
if len(oss_steps) > 0:
|
582 |
infer_steps = max(oss_steps)
|
|
|
591 |
logger.info(f"oss_steps: {oss_steps}, num_inference_steps: {num_inference_steps} after remapping to timesteps {timesteps}")
|
592 |
else:
|
593 |
timesteps, num_inference_steps = retrieve_timesteps(scheduler, num_inference_steps=infer_steps, device=device, timesteps=None)
|
594 |
+
|
595 |
target_latents = randn_tensor(shape=(bsz, 8, 16, frame_length), generator=random_generators, device=device, dtype=dtype)
|
596 |
+
|
597 |
+
is_repaint = False
|
598 |
if add_retake_noise:
|
599 |
retake_variance = torch.tensor(retake_variance * math.pi/2).to(device).to(dtype)
|
600 |
retake_latents = randn_tensor(shape=(bsz, 8, 16, frame_length), generator=retake_random_generators, device=device, dtype=dtype)
|
601 |
+
repaint_start_frame = int(repaint_start * 44100 / 512 / 8)
|
602 |
+
repaint_end_frame = int(repaint_end * 44100 / 512 / 8)
|
603 |
+
|
604 |
+
# retake
|
605 |
+
is_repaint = repaint_end_frame - repaint_start_frame != frame_length
|
606 |
# to make sure mean = 0, std = 1
|
607 |
+
if not is_repaint:
|
608 |
+
target_latents = torch.cos(retake_variance) * target_latents + torch.sin(retake_variance) * retake_latents
|
609 |
+
else:
|
610 |
+
repaint_mask = torch.zeros((bsz, 8, 16, frame_length), device=device, dtype=dtype)
|
611 |
+
repaint_mask[:, :, :, repaint_start_frame:repaint_end_frame] = 1.0
|
612 |
+
repaint_noise = torch.cos(retake_variance) * target_latents + torch.sin(retake_variance) * retake_latents
|
613 |
+
repaint_noise = torch.where(repaint_mask == 1.0, repaint_noise, target_latents)
|
614 |
+
z0 = repaint_noise
|
615 |
|
616 |
attention_mask = torch.ones(bsz, frame_length, device=device, dtype=dtype)
|
617 |
+
|
618 |
# guidance interval
|
619 |
start_idx = int(num_inference_steps * ((1 - guidance_interval) / 2))
|
620 |
end_idx = int(num_inference_steps * (guidance_interval / 2 + 0.5))
|
|
|
624 |
|
625 |
def forward_encoder_with_temperature(self, inputs, tau=0.01, l_min=4, l_max=6):
|
626 |
handlers = []
|
627 |
+
|
628 |
def hook(module, input, output):
|
629 |
output[:] *= tau
|
630 |
return output
|
631 |
+
|
632 |
for i in range(l_min, l_max):
|
633 |
handler = self.ace_step_transformer.lyric_encoder.encoders[i].self_attn.linear_q.register_forward_hook(hook)
|
634 |
handlers.append(handler)
|
635 |
+
|
636 |
encoder_hidden_states, encoder_hidden_mask = self.ace_step_transformer.encode(**inputs)
|
637 |
+
|
638 |
for hook in handlers:
|
639 |
hook.remove()
|
640 |
+
|
641 |
return encoder_hidden_states
|
642 |
|
643 |
# P(speaker, text, lyric)
|
|
|
670 |
torch.zeros_like(lyric_token_ids),
|
671 |
lyric_mask,
|
672 |
)
|
673 |
+
|
674 |
encoder_hidden_states_no_lyric = None
|
675 |
if do_double_condition_guidance:
|
676 |
# P(null_speaker, text, lyric_weaker)
|
|
|
697 |
|
698 |
def forward_diffusion_with_temperature(self, hidden_states, timestep, inputs, tau=0.01, l_min=15, l_max=20):
|
699 |
handlers = []
|
700 |
+
|
701 |
def hook(module, input, output):
|
702 |
output[:] *= tau
|
703 |
return output
|
704 |
+
|
705 |
for i in range(l_min, l_max):
|
706 |
handler = self.ace_step_transformer.transformer_blocks[i].attn.to_q.register_forward_hook(hook)
|
707 |
handlers.append(handler)
|
|
|
709 |
handlers.append(handler)
|
710 |
|
711 |
sample = self.ace_step_transformer.decode(hidden_states=hidden_states, timestep=timestep, **inputs).sample
|
712 |
+
|
713 |
for hook in handlers:
|
714 |
hook.remove()
|
715 |
+
|
716 |
return sample
|
717 |
|
718 |
for i, t in tqdm(enumerate(timesteps), total=num_inference_steps):
|
|
|
819 |
).sample
|
820 |
|
821 |
target_latents = scheduler.step(model_output=noise_pred, timestep=t, sample=target_latents, return_dict=False, omega=omega_scale)[0]
|
822 |
+
if is_repaint:
|
823 |
+
t_i = t / 1000
|
824 |
+
x0 = src_latents
|
825 |
+
xt = (1 - t_i) * x0 + t_i * z0
|
826 |
+
target_latents = torch.where(repaint_mask == 1.0, target_latents, xt)
|
827 |
|
828 |
+
|
829 |
return target_latents
|
830 |
|
831 |
def latents2audio(self, latents, target_wav_duration_second=30, sample_rate=48000, save_path=None, format="flac"):
|
|
|
844 |
def save_wav_file(self, target_wav, idx, save_path=None, sample_rate=48000, format="flac"):
|
845 |
if save_path is None:
|
846 |
logger.warning("save_path is None, using default path ./outputs/")
|
847 |
+
base_path = f"./outputs"
|
848 |
ensure_directory_exists(base_path)
|
849 |
else:
|
850 |
base_path = save_path
|
|
|
855 |
torchaudio.save(output_path_flac, target_wav, sample_rate=sample_rate, format=format)
|
856 |
return output_path_flac
|
857 |
|
858 |
+
def infer_latents(self, input_audio_path):
|
859 |
+
if input_audio_path is None:
|
860 |
+
return None
|
861 |
+
input_audio, sr = self.music_dcae.load_audio(input_audio_path)
|
862 |
+
input_audio = input_audio.unsqueeze(0)
|
863 |
+
device, dtype = self.device, self.dtype
|
864 |
+
input_audio = input_audio.to(device=device, dtype=dtype)
|
865 |
+
latents, _ = self.music_dcae.encode(input_audio, sr=sr)
|
866 |
+
return latents
|
867 |
+
|
868 |
def __call__(
|
869 |
self,
|
870 |
audio_duration: float = 60.0,
|
|
|
888 |
retake_seeds: list = None,
|
889 |
retake_variance: float = 0.5,
|
890 |
task: str = "text2music",
|
891 |
+
repaint_start: int = 0,
|
892 |
+
repaint_end: int = 0,
|
893 |
+
src_audio_path: str = None,
|
894 |
+
edit_target_prompt: str = None,
|
895 |
+
edit_target_lyrics: str = None,
|
896 |
+
edit_n_min: float = 0.0,
|
897 |
+
edit_n_max: float = 1.0,
|
898 |
+
edit_n_avg: int = 1,
|
899 |
save_path: str = None,
|
900 |
format: str = "flac",
|
901 |
batch_size: int = 1,
|
|
|
918 |
oss_steps = list(map(int, oss_steps.split(",")))
|
919 |
else:
|
920 |
oss_steps = []
|
921 |
+
|
922 |
texts = [prompt]
|
923 |
encoder_text_hidden_states, text_attention_mask = self.get_text_embeddings(texts, self.device)
|
924 |
encoder_text_hidden_states = encoder_text_hidden_states.repeat(batch_size, 1, 1)
|
|
|
949 |
preprocess_time_cost = end_time - start_time
|
950 |
start_time = end_time
|
951 |
|
952 |
+
add_retake_noise = task in ("retake", "repaint")
|
953 |
+
# retake equal to repaint
|
954 |
+
if task == "retake":
|
955 |
+
repaint_start = 0
|
956 |
+
repaint_end = audio_duration
|
957 |
+
|
958 |
+
src_latents = None
|
959 |
+
if src_audio_path is not None:
|
960 |
+
assert src_audio_path is not None and task in ("repaint", "edit"), "src_audio_path is required for repaint task"
|
961 |
+
assert os.path.exists(src_audio_path), f"src_audio_path {src_audio_path} does not exist"
|
962 |
+
src_latents = self.infer_latents(src_audio_path)
|
963 |
+
|
964 |
+
if task == "edit":
|
965 |
+
texts = [edit_target_prompt]
|
966 |
+
target_encoder_text_hidden_states, target_text_attention_mask = self.get_text_embeddings(texts, self.device)
|
967 |
+
target_encoder_text_hidden_states = target_encoder_text_hidden_states.repeat(batch_size, 1, 1)
|
968 |
+
target_text_attention_mask = target_text_attention_mask.repeat(batch_size, 1)
|
969 |
+
|
970 |
+
target_lyric_token_idx = torch.tensor([0]).repeat(batch_size, 1).to(self.device).long()
|
971 |
+
target_lyric_mask = torch.tensor([0]).repeat(batch_size, 1).to(self.device).long()
|
972 |
+
if len(edit_target_lyrics) > 0:
|
973 |
+
target_lyric_token_idx = self.tokenize_lyrics(edit_target_lyrics, debug=True)
|
974 |
+
target_lyric_mask = [1] * len(target_lyric_token_idx)
|
975 |
+
target_lyric_token_idx = torch.tensor(target_lyric_token_idx).unsqueeze(0).to(self.device).repeat(batch_size, 1)
|
976 |
+
target_lyric_mask = torch.tensor(target_lyric_mask).unsqueeze(0).to(self.device).repeat(batch_size, 1)
|
977 |
+
|
978 |
+
target_speaker_embeds = speaker_embeds.clone()
|
979 |
+
|
980 |
+
target_latents = self.flowedit_diffusion_process(
|
981 |
+
encoder_text_hidden_states=encoder_text_hidden_states,
|
982 |
+
text_attention_mask=text_attention_mask,
|
983 |
+
speaker_embds=speaker_embeds,
|
984 |
+
lyric_token_ids=lyric_token_idx,
|
985 |
+
lyric_mask=lyric_mask,
|
986 |
+
target_encoder_text_hidden_states=target_encoder_text_hidden_states,
|
987 |
+
target_text_attention_mask=target_text_attention_mask,
|
988 |
+
target_speaker_embeds=target_speaker_embeds,
|
989 |
+
target_lyric_token_ids=target_lyric_token_idx,
|
990 |
+
target_lyric_mask=target_lyric_mask,
|
991 |
+
src_latents=src_latents,
|
992 |
+
random_generators=random_generators,
|
993 |
+
infer_steps=infer_step,
|
994 |
+
guidance_scale=guidance_scale,
|
995 |
+
n_min=edit_n_min,
|
996 |
+
n_max=edit_n_max,
|
997 |
+
n_avg=edit_n_avg,
|
998 |
+
)
|
999 |
+
else:
|
1000 |
+
target_latents = self.text2music_diffusion_process(
|
1001 |
+
duration=audio_duration,
|
1002 |
+
encoder_text_hidden_states=encoder_text_hidden_states,
|
1003 |
+
text_attention_mask=text_attention_mask,
|
1004 |
+
speaker_embds=speaker_embeds,
|
1005 |
+
lyric_token_ids=lyric_token_idx,
|
1006 |
+
lyric_mask=lyric_mask,
|
1007 |
+
guidance_scale=guidance_scale,
|
1008 |
+
omega_scale=omega_scale,
|
1009 |
+
infer_steps=infer_step,
|
1010 |
+
random_generators=random_generators,
|
1011 |
+
scheduler_type=scheduler_type,
|
1012 |
+
cfg_type=cfg_type,
|
1013 |
+
guidance_interval=guidance_interval,
|
1014 |
+
guidance_interval_decay=guidance_interval_decay,
|
1015 |
+
min_guidance_scale=min_guidance_scale,
|
1016 |
+
oss_steps=oss_steps,
|
1017 |
+
encoder_text_hidden_states_null=encoder_text_hidden_states_null,
|
1018 |
+
use_erg_lyric=use_erg_lyric,
|
1019 |
+
use_erg_diffusion=use_erg_diffusion,
|
1020 |
+
retake_random_generators=retake_random_generators,
|
1021 |
+
retake_variance=retake_variance,
|
1022 |
+
add_retake_noise=add_retake_noise,
|
1023 |
+
guidance_scale_text=guidance_scale_text,
|
1024 |
+
guidance_scale_lyric=guidance_scale_lyric,
|
1025 |
+
repaint_start=repaint_start,
|
1026 |
+
repaint_end=repaint_end,
|
1027 |
+
src_latents=src_latents,
|
1028 |
+
)
|
1029 |
|
1030 |
end_time = time.time()
|
1031 |
diffusion_time_cost = end_time - start_time
|
|
|
1069 |
"retake_variance": retake_variance,
|
1070 |
"guidance_scale_text": guidance_scale_text,
|
1071 |
"guidance_scale_lyric": guidance_scale_lyric,
|
1072 |
+
"repaint_start": repaint_start,
|
1073 |
+
"repaint_end": repaint_end,
|
1074 |
+
"edit_n_min": edit_n_min,
|
1075 |
+
"edit_n_max": edit_n_max,
|
1076 |
+
"edit_n_avg": edit_n_avg,
|
1077 |
+
"src_audio_path": src_audio_path,
|
1078 |
+
"edit_target_prompt": edit_target_prompt,
|
1079 |
+
"edit_target_lyrics": edit_target_lyrics,
|
1080 |
}
|
1081 |
# save input_params_json
|
1082 |
for output_audio_path in output_paths:
|
ui/components.py
CHANGED
@@ -63,15 +63,15 @@ def create_text2music_ui(
|
|
63 |
):
|
64 |
with gr.Row():
|
65 |
with gr.Column():
|
66 |
-
|
67 |
with gr.Row(equal_height=True):
|
|
|
68 |
audio_duration = gr.Slider(-1, 240.0, step=0.00001, value=180, label="Audio Duration", interactive=True, info="-1 means random duration (30 ~ 240).", scale=9)
|
69 |
sample_bnt = gr.Button("Sample", variant="primary", scale=1)
|
70 |
|
71 |
-
prompt = gr.Textbox(lines=2, label="Tags", max_lines=4, placeholder=TAG_PLACEHOLDER, info="Support tags, descriptions, and scene. Use commas to separate different tags
|
72 |
lyrics = gr.Textbox(lines=9, label="Lyrics", max_lines=13, placeholder=LYRIC_PLACEHOLDER, info="Support lyric structure tags like [verse], [chorus], and [bridge] to separate different parts of the lyrics.\nUse [instrumental] or [inst] to generate instrumental music. Not support genre structure tag in lyrics")
|
73 |
|
74 |
-
with gr.Accordion("Basic Settings", open=
|
75 |
infer_step = gr.Slider(minimum=1, maximum=1000, step=1, value=60, label="Infer Steps", interactive=True)
|
76 |
guidance_scale = gr.Slider(minimum=0.0, maximum=200.0, step=0.1, value=15.0, label="Guidance Scale", interactive=True, info="When guidance_scale_lyric > 1 and guidance_scale_text > 1, the guidance scale will not be applied.")
|
77 |
guidance_scale_text = gr.Slider(minimum=0.0, maximum=10.0, step=0.1, value=5.0, label="Guidance Scale Text", interactive=True, info="Guidance scale for text condition. It can only apply to cfg. set guidance_scale_text=5.0, guidance_scale_lyric=1.5 for start")
|
@@ -93,14 +93,14 @@ def create_text2music_ui(
|
|
93 |
min_guidance_scale = gr.Slider(minimum=0.0, maximum=200.0, step=0.1, value=3.0, label="Min Guidance Scale", interactive=True, info="Min guidance scale for guidance interval decay's end scale")
|
94 |
oss_steps = gr.Textbox(label="OSS Steps", placeholder="16, 29, 52, 96, 129, 158, 172, 183, 189, 200", value=None, info="Optimal Steps for the generation. But not test well")
|
95 |
|
96 |
-
text2music_bnt = gr.Button(variant="primary")
|
97 |
|
98 |
with gr.Column():
|
99 |
outputs, input_params_json = create_output_ui()
|
100 |
with gr.Tab("retake"):
|
101 |
-
retake_variance = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, value=0.2, label="variance"
|
102 |
-
retake_seeds = gr.Textbox(label="retake seeds (default None)", placeholder="", value=None
|
103 |
-
retake_bnt = gr.Button(variant="primary")
|
104 |
retake_outputs, retake_input_params_json = create_output_ui("Retake")
|
105 |
|
106 |
def retake_process_func(json_data, retake_variance, retake_seeds):
|
@@ -138,9 +138,251 @@ def create_text2music_ui(
|
|
138 |
outputs=retake_outputs + [retake_input_params_json],
|
139 |
)
|
140 |
with gr.Tab("repainting"):
|
141 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
142 |
with gr.Tab("edit"):
|
143 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
144 |
|
145 |
def sample_data():
|
146 |
json_data = sample_data_func()
|
@@ -219,13 +461,12 @@ def create_main_demo_ui(
|
|
219 |
sample_data_func=dump_func,
|
220 |
):
|
221 |
with gr.Blocks(
|
222 |
-
title="
|
223 |
) as demo:
|
224 |
gr.Markdown(
|
225 |
"""
|
226 |
-
<h1 style="text-align: center;">
|
227 |
-
|
228 |
-
)
|
229 |
|
230 |
with gr.Tab("text2music"):
|
231 |
create_text2music_ui(
|
|
|
63 |
):
|
64 |
with gr.Row():
|
65 |
with gr.Column():
|
|
|
66 |
with gr.Row(equal_height=True):
|
67 |
+
# add markdown, tags and lyrics examples are from ai music generation community
|
68 |
audio_duration = gr.Slider(-1, 240.0, step=0.00001, value=180, label="Audio Duration", interactive=True, info="-1 means random duration (30 ~ 240).", scale=9)
|
69 |
sample_bnt = gr.Button("Sample", variant="primary", scale=1)
|
70 |
|
71 |
+
prompt = gr.Textbox(lines=2, label="Tags", max_lines=4, placeholder=TAG_PLACEHOLDER, info="Support tags, descriptions, and scene. Use commas to separate different tags.\ntags and lyrics examples are from ai music generation community")
|
72 |
lyrics = gr.Textbox(lines=9, label="Lyrics", max_lines=13, placeholder=LYRIC_PLACEHOLDER, info="Support lyric structure tags like [verse], [chorus], and [bridge] to separate different parts of the lyrics.\nUse [instrumental] or [inst] to generate instrumental music. Not support genre structure tag in lyrics")
|
73 |
|
74 |
+
with gr.Accordion("Basic Settings", open=False):
|
75 |
infer_step = gr.Slider(minimum=1, maximum=1000, step=1, value=60, label="Infer Steps", interactive=True)
|
76 |
guidance_scale = gr.Slider(minimum=0.0, maximum=200.0, step=0.1, value=15.0, label="Guidance Scale", interactive=True, info="When guidance_scale_lyric > 1 and guidance_scale_text > 1, the guidance scale will not be applied.")
|
77 |
guidance_scale_text = gr.Slider(minimum=0.0, maximum=10.0, step=0.1, value=5.0, label="Guidance Scale Text", interactive=True, info="Guidance scale for text condition. It can only apply to cfg. set guidance_scale_text=5.0, guidance_scale_lyric=1.5 for start")
|
|
|
93 |
min_guidance_scale = gr.Slider(minimum=0.0, maximum=200.0, step=0.1, value=3.0, label="Min Guidance Scale", interactive=True, info="Min guidance scale for guidance interval decay's end scale")
|
94 |
oss_steps = gr.Textbox(label="OSS Steps", placeholder="16, 29, 52, 96, 129, 158, 172, 183, 189, 200", value=None, info="Optimal Steps for the generation. But not test well")
|
95 |
|
96 |
+
text2music_bnt = gr.Button("Generate", variant="primary")
|
97 |
|
98 |
with gr.Column():
|
99 |
outputs, input_params_json = create_output_ui()
|
100 |
with gr.Tab("retake"):
|
101 |
+
retake_variance = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, value=0.2, label="variance")
|
102 |
+
retake_seeds = gr.Textbox(label="retake seeds (default None)", placeholder="", value=None)
|
103 |
+
retake_bnt = gr.Button("Retake", variant="primary")
|
104 |
retake_outputs, retake_input_params_json = create_output_ui("Retake")
|
105 |
|
106 |
def retake_process_func(json_data, retake_variance, retake_seeds):
|
|
|
138 |
outputs=retake_outputs + [retake_input_params_json],
|
139 |
)
|
140 |
with gr.Tab("repainting"):
|
141 |
+
retake_variance = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, value=0.2, label="variance")
|
142 |
+
retake_seeds = gr.Textbox(label="retake seeds (default None)", placeholder="", value=None)
|
143 |
+
repaint_start = gr.Slider(minimum=0.0, maximum=240.0, step=0.01, value=0.0, label="Repaint Start Time", interactive=True)
|
144 |
+
repaint_end = gr.Slider(minimum=0.0, maximum=240.0, step=0.01, value=30.0, label="Repaint End Time", interactive=True)
|
145 |
+
repaint_source = gr.Radio(["text2music", "last_repaint", "upload"], value="text2music", label="Repaint Source", elem_id="repaint_source")
|
146 |
+
|
147 |
+
repaint_source_audio_upload = gr.Audio(label="Upload Audio", type="filepath", visible=False, elem_id="repaint_source_audio_upload")
|
148 |
+
repaint_source.change(
|
149 |
+
fn=lambda x: gr.update(visible=x == "upload", elem_id="repaint_source_audio_upload"),
|
150 |
+
inputs=[repaint_source],
|
151 |
+
outputs=[repaint_source_audio_upload],
|
152 |
+
)
|
153 |
+
|
154 |
+
repaint_bnt = gr.Button("Repaint", variant="primary")
|
155 |
+
repaint_outputs, repaint_input_params_json = create_output_ui("Repaint")
|
156 |
+
|
157 |
+
def repaint_process_func(
|
158 |
+
text2music_json_data,
|
159 |
+
repaint_json_data,
|
160 |
+
retake_variance,
|
161 |
+
retake_seeds,
|
162 |
+
repaint_start,
|
163 |
+
repaint_end,
|
164 |
+
repaint_source,
|
165 |
+
repaint_source_audio_upload,
|
166 |
+
prompt,
|
167 |
+
lyrics,
|
168 |
+
infer_step,
|
169 |
+
guidance_scale,
|
170 |
+
scheduler_type,
|
171 |
+
cfg_type,
|
172 |
+
omega_scale,
|
173 |
+
manual_seeds,
|
174 |
+
guidance_interval,
|
175 |
+
guidance_interval_decay,
|
176 |
+
min_guidance_scale,
|
177 |
+
use_erg_tag,
|
178 |
+
use_erg_lyric,
|
179 |
+
use_erg_diffusion,
|
180 |
+
oss_steps,
|
181 |
+
guidance_scale_text,
|
182 |
+
guidance_scale_lyric,
|
183 |
+
):
|
184 |
+
if repaint_source == "upload":
|
185 |
+
src_audio_path = repaint_source_audio_upload
|
186 |
+
json_data = text2music_json_data
|
187 |
+
elif repaint_source == "text2music":
|
188 |
+
json_data = text2music_json_data
|
189 |
+
src_audio_path = json_data["audio_path"]
|
190 |
+
elif repaint_source == "last_repaint":
|
191 |
+
json_data = repaint_json_data
|
192 |
+
src_audio_path = json_data["audio_path"]
|
193 |
+
|
194 |
+
return text2music_process_func(
|
195 |
+
json_data["audio_duration"],
|
196 |
+
prompt,
|
197 |
+
lyrics,
|
198 |
+
infer_step,
|
199 |
+
guidance_scale,
|
200 |
+
scheduler_type,
|
201 |
+
cfg_type,
|
202 |
+
omega_scale,
|
203 |
+
manual_seeds,
|
204 |
+
guidance_interval,
|
205 |
+
guidance_interval_decay,
|
206 |
+
min_guidance_scale,
|
207 |
+
use_erg_tag,
|
208 |
+
use_erg_lyric,
|
209 |
+
use_erg_diffusion,
|
210 |
+
oss_steps,
|
211 |
+
guidance_scale_text,
|
212 |
+
guidance_scale_lyric,
|
213 |
+
retake_seeds=retake_seeds,
|
214 |
+
retake_variance=retake_variance,
|
215 |
+
task="repaint",
|
216 |
+
repaint_start=repaint_start,
|
217 |
+
repaint_end=repaint_end,
|
218 |
+
src_audio_path=src_audio_path,
|
219 |
+
)
|
220 |
+
|
221 |
+
repaint_bnt.click(
|
222 |
+
fn=repaint_process_func,
|
223 |
+
inputs=[
|
224 |
+
input_params_json,
|
225 |
+
repaint_input_params_json,
|
226 |
+
retake_variance,
|
227 |
+
retake_seeds,
|
228 |
+
repaint_start,
|
229 |
+
repaint_end,
|
230 |
+
repaint_source,
|
231 |
+
repaint_source_audio_upload,
|
232 |
+
prompt,
|
233 |
+
lyrics,
|
234 |
+
infer_step,
|
235 |
+
guidance_scale,
|
236 |
+
scheduler_type,
|
237 |
+
cfg_type,
|
238 |
+
omega_scale,
|
239 |
+
manual_seeds,
|
240 |
+
guidance_interval,
|
241 |
+
guidance_interval_decay,
|
242 |
+
min_guidance_scale,
|
243 |
+
use_erg_tag,
|
244 |
+
use_erg_lyric,
|
245 |
+
use_erg_diffusion,
|
246 |
+
oss_steps,
|
247 |
+
guidance_scale_text,
|
248 |
+
guidance_scale_lyric,
|
249 |
+
],
|
250 |
+
outputs=repaint_outputs + [repaint_input_params_json],
|
251 |
+
)
|
252 |
with gr.Tab("edit"):
|
253 |
+
edit_prompt = gr.Textbox(lines=2, label="Edit Tags", max_lines=4)
|
254 |
+
edit_lyrics = gr.Textbox(lines=9, label="Edit Lyrics", max_lines=13)
|
255 |
+
|
256 |
+
edit_type = gr.Radio(["only_lyrics", "remix"], value="only_lyrics", label="Edit Type", elem_id="edit_type", info="`only_lyrics` will keep the whole song the same except lyrics difference. Make your diffrence smaller, e.g. one lyrc line change.\nremix can change the song melody and genre")
|
257 |
+
edit_n_min = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, value=0.8, label="edit_n_min", interactive=True)
|
258 |
+
edit_n_max = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, value=1.0, label="edit_n_max", interactive=True)
|
259 |
+
|
260 |
+
def edit_type_change_func(edit_type):
|
261 |
+
if edit_type == "only_lyrics":
|
262 |
+
n_min = 0.8
|
263 |
+
n_max = 1.0
|
264 |
+
elif edit_type == "remix":
|
265 |
+
n_min = 0.2
|
266 |
+
n_max = 0.4
|
267 |
+
return n_min, n_max
|
268 |
+
|
269 |
+
edit_type.change(
|
270 |
+
edit_type_change_func,
|
271 |
+
inputs=[edit_type],
|
272 |
+
outputs=[edit_n_min, edit_n_max]
|
273 |
+
)
|
274 |
+
|
275 |
+
edit_source = gr.Radio(["text2music", "last_edit", "upload"], value="text2music", label="Edit Source", elem_id="edit_source")
|
276 |
+
edit_source_audio_upload = gr.Audio(label="Upload Audio", type="filepath", visible=False, elem_id="edit_source_audio_upload")
|
277 |
+
edit_source.change(
|
278 |
+
fn=lambda x: gr.update(visible=x == "upload", elem_id="edit_source_audio_upload"),
|
279 |
+
inputs=[edit_source],
|
280 |
+
outputs=[edit_source_audio_upload],
|
281 |
+
)
|
282 |
+
|
283 |
+
edit_bnt = gr.Button("Edit", variant="primary")
|
284 |
+
edit_outputs, edit_input_params_json = create_output_ui("Edit")
|
285 |
+
|
286 |
+
def edit_process_func(
|
287 |
+
text2music_json_data,
|
288 |
+
edit_input_params_json,
|
289 |
+
edit_source,
|
290 |
+
edit_source_audio_upload,
|
291 |
+
prompt,
|
292 |
+
lyrics,
|
293 |
+
edit_prompt,
|
294 |
+
edit_lyrics,
|
295 |
+
edit_n_min,
|
296 |
+
edit_n_max,
|
297 |
+
infer_step,
|
298 |
+
guidance_scale,
|
299 |
+
scheduler_type,
|
300 |
+
cfg_type,
|
301 |
+
omega_scale,
|
302 |
+
manual_seeds,
|
303 |
+
guidance_interval,
|
304 |
+
guidance_interval_decay,
|
305 |
+
min_guidance_scale,
|
306 |
+
use_erg_tag,
|
307 |
+
use_erg_lyric,
|
308 |
+
use_erg_diffusion,
|
309 |
+
oss_steps,
|
310 |
+
guidance_scale_text,
|
311 |
+
guidance_scale_lyric,
|
312 |
+
):
|
313 |
+
if edit_source == "upload":
|
314 |
+
src_audio_path = edit_source_audio_upload
|
315 |
+
json_data = text2music_json_data
|
316 |
+
elif edit_source == "text2music":
|
317 |
+
json_data = text2music_json_data
|
318 |
+
src_audio_path = json_data["audio_path"]
|
319 |
+
elif edit_source == "last_edit":
|
320 |
+
json_data = edit_input_params_json
|
321 |
+
src_audio_path = json_data["audio_path"]
|
322 |
+
|
323 |
+
if not edit_prompt:
|
324 |
+
edit_prompt = prompt
|
325 |
+
if not edit_lyrics:
|
326 |
+
edit_lyrics = lyrics
|
327 |
+
|
328 |
+
return text2music_process_func(
|
329 |
+
json_data["audio_duration"],
|
330 |
+
prompt,
|
331 |
+
lyrics,
|
332 |
+
infer_step,
|
333 |
+
guidance_scale,
|
334 |
+
scheduler_type,
|
335 |
+
cfg_type,
|
336 |
+
omega_scale,
|
337 |
+
manual_seeds,
|
338 |
+
guidance_interval,
|
339 |
+
guidance_interval_decay,
|
340 |
+
min_guidance_scale,
|
341 |
+
use_erg_tag,
|
342 |
+
use_erg_lyric,
|
343 |
+
use_erg_diffusion,
|
344 |
+
oss_steps,
|
345 |
+
guidance_scale_text,
|
346 |
+
guidance_scale_lyric,
|
347 |
+
task="edit",
|
348 |
+
src_audio_path=src_audio_path,
|
349 |
+
edit_target_prompt=edit_prompt,
|
350 |
+
edit_target_lyrics=edit_lyrics,
|
351 |
+
edit_n_min=edit_n_min,
|
352 |
+
edit_n_max=edit_n_max
|
353 |
+
)
|
354 |
+
|
355 |
+
edit_bnt.click(
|
356 |
+
fn=edit_process_func,
|
357 |
+
inputs=[
|
358 |
+
input_params_json,
|
359 |
+
edit_input_params_json,
|
360 |
+
edit_source,
|
361 |
+
edit_source_audio_upload,
|
362 |
+
prompt,
|
363 |
+
lyrics,
|
364 |
+
edit_prompt,
|
365 |
+
edit_lyrics,
|
366 |
+
edit_n_min,
|
367 |
+
edit_n_max,
|
368 |
+
infer_step,
|
369 |
+
guidance_scale,
|
370 |
+
scheduler_type,
|
371 |
+
cfg_type,
|
372 |
+
omega_scale,
|
373 |
+
manual_seeds,
|
374 |
+
guidance_interval,
|
375 |
+
guidance_interval_decay,
|
376 |
+
min_guidance_scale,
|
377 |
+
use_erg_tag,
|
378 |
+
use_erg_lyric,
|
379 |
+
use_erg_diffusion,
|
380 |
+
oss_steps,
|
381 |
+
guidance_scale_text,
|
382 |
+
guidance_scale_lyric,
|
383 |
+
],
|
384 |
+
outputs=edit_outputs + [edit_input_params_json],
|
385 |
+
)
|
386 |
|
387 |
def sample_data():
|
388 |
json_data = sample_data_func()
|
|
|
461 |
sample_data_func=dump_func,
|
462 |
):
|
463 |
with gr.Blocks(
|
464 |
+
title="ACE-Step Model 1.0 DEMO",
|
465 |
) as demo:
|
466 |
gr.Markdown(
|
467 |
"""
|
468 |
+
<h1 style="text-align: center;">ACE-Step: A Step Towards Music Generation Foundation Model</h1>
|
469 |
+
""")
|
|
|
470 |
|
471 |
with gr.Tab("text2music"):
|
472 |
create_text2music_ui(
|