Spaces:
Running
on
Zero
Running
on
Zero
Extra mode
Browse files- app.py +42 -8
- diffrhythm/infer/infer.py +5 -4
- diffrhythm/infer/infer_utils.py +19 -7
- diffrhythm/model/cfm.py +7 -17
- diffrhythm/model/dit.py +1 -17
app.py
CHANGED
@@ -18,7 +18,8 @@ import base64
|
|
18 |
from diffrhythm.infer.infer_utils import (
|
19 |
get_reference_latent,
|
20 |
get_lrc_token,
|
21 |
-
|
|
|
22 |
prepare_model,
|
23 |
get_negative_style_prompt
|
24 |
)
|
@@ -29,16 +30,19 @@ device='cuda'
|
|
29 |
cfm, tokenizer, muq, vae = prepare_model(device)
|
30 |
cfm = torch.compile(cfm)
|
31 |
|
32 |
-
|
33 |
-
def infer_music(lrc, ref_audio_path, seed=42, randomize_seed=False, steps=32, file_type='wav', max_frames=2048, device='cuda'):
|
34 |
|
35 |
if randomize_seed:
|
36 |
seed = random.randint(0, MAX_SEED)
|
37 |
torch.manual_seed(seed)
|
38 |
sway_sampling_coef = -1 if steps < 32 else None
|
|
|
39 |
try:
|
40 |
lrc_prompt, start_time = get_lrc_token(lrc, tokenizer, device)
|
41 |
-
|
|
|
|
|
|
|
42 |
except Exception as e:
|
43 |
raise gr.Error(f"Error: {str(e)}")
|
44 |
negative_style_prompt = get_negative_style_prompt(device)
|
@@ -53,7 +57,8 @@ def infer_music(lrc, ref_audio_path, seed=42, randomize_seed=False, steps=32, fi
|
|
53 |
steps=steps,
|
54 |
sway_sampling_coef=sway_sampling_coef,
|
55 |
start_time=start_time,
|
56 |
-
file_type=file_type
|
|
|
57 |
)
|
58 |
return generated_song
|
59 |
|
@@ -179,7 +184,23 @@ with gr.Blocks(css=css) as demo:
|
|
179 |
elem_classes="lyrics-scroll-box",
|
180 |
value="""[00:10.00]Moonlight spills through broken blinds\n[00:13.20]Your shadow dances on the dashboard shrine\n[00:16.85]Neon ghosts in gasoline rain\n[00:20.40]I hear your laughter down the midnight train\n[00:24.15]Static whispers through frayed wires\n[00:27.65]Guitar strings hum our cathedral choirs\n[00:31.30]Flicker screens show reruns of June\n[00:34.90]I'm drowning in this mercury lagoon\n[00:38.55]Electric veins pulse through concrete skies\n[00:42.10]Your name echoes in the hollow where my heartbeat lies\n[00:45.75]We're satellites trapped in parallel light\n[00:49.25]Burning through the atmosphere of endless night\n[01:00.00]Dusty vinyl spins reverse\n[01:03.45]Our polaroid timeline bleeds through the verse\n[01:07.10]Telescope aimed at dead stars\n[01:10.65]Still tracing constellations through prison bars\n[01:14.30]Electric veins pulse through concrete skies\n[01:17.85]Your name echoes in the hollow where my heartbeat lies\n[01:21.50]We're satellites trapped in parallel light\n[01:25.05]Burning through the atmosphere of endless night\n[02:10.00]Clockwork gears grind moonbeams to rust\n[02:13.50]Our fingerprint smudged by interstellar dust\n[02:17.15]Velvet thunder rolls through my veins\n[02:20.70]Chasing phantom trains through solar plane\n[02:24.35]Electric veins pulse through concrete skies\n[02:27.90]Your name echoes in the hollow where my heartbeat lies"""
|
181 |
)
|
182 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
183 |
|
184 |
with gr.Column():
|
185 |
with gr.Accordion("Best Practices Guide", open=True):
|
@@ -218,7 +239,7 @@ with gr.Blocks(css=css) as demo:
|
|
218 |
steps = gr.Slider(
|
219 |
minimum=10,
|
220 |
maximum=100,
|
221 |
-
value=32,
|
222 |
step=1,
|
223 |
label="Diffusion Steps",
|
224 |
interactive=True,
|
@@ -248,6 +269,19 @@ with gr.Blocks(css=css) as demo:
|
|
248 |
examples_per_page=13,
|
249 |
elem_id="audio-examples-container"
|
250 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
251 |
|
252 |
gr.Examples(
|
253 |
examples=[
|
@@ -352,7 +386,7 @@ with gr.Blocks(css=css) as demo:
|
|
352 |
|
353 |
lyrics_btn.click(
|
354 |
fn=infer_music,
|
355 |
-
inputs=[lrc, audio_prompt, seed, randomize_seed, steps, file_type],
|
356 |
outputs=audio_output
|
357 |
)
|
358 |
|
|
|
18 |
from diffrhythm.infer.infer_utils import (
|
19 |
get_reference_latent,
|
20 |
get_lrc_token,
|
21 |
+
get_audio_style_prompt,
|
22 |
+
get_text_style_prompt,
|
23 |
prepare_model,
|
24 |
get_negative_style_prompt
|
25 |
)
|
|
|
30 |
cfm, tokenizer, muq, vae = prepare_model(device)
|
31 |
cfm = torch.compile(cfm)
|
32 |
|
33 |
+
def infer_music(lrc, ref_audio_path, text_prompt, current_prompt_type, seed=42, randomize_seed=False, steps=32, file_type='wav', max_frames=2048, device='cuda'):
|
|
|
34 |
|
35 |
if randomize_seed:
|
36 |
seed = random.randint(0, MAX_SEED)
|
37 |
torch.manual_seed(seed)
|
38 |
sway_sampling_coef = -1 if steps < 32 else None
|
39 |
+
vocal_flag = False
|
40 |
try:
|
41 |
lrc_prompt, start_time = get_lrc_token(lrc, tokenizer, device)
|
42 |
+
if current_prompt_type == 'audio':
|
43 |
+
style_prompt, vocal_flag = get_audio_style_prompt(muq, ref_audio_path)
|
44 |
+
else:
|
45 |
+
style_prompt = get_text_style_prompt(muq, text_prompt)
|
46 |
except Exception as e:
|
47 |
raise gr.Error(f"Error: {str(e)}")
|
48 |
negative_style_prompt = get_negative_style_prompt(device)
|
|
|
57 |
steps=steps,
|
58 |
sway_sampling_coef=sway_sampling_coef,
|
59 |
start_time=start_time,
|
60 |
+
file_type=file_type,
|
61 |
+
vocal_flag=vocal_flag
|
62 |
)
|
63 |
return generated_song
|
64 |
|
|
|
184 |
elem_classes="lyrics-scroll-box",
|
185 |
value="""[00:10.00]Moonlight spills through broken blinds\n[00:13.20]Your shadow dances on the dashboard shrine\n[00:16.85]Neon ghosts in gasoline rain\n[00:20.40]I hear your laughter down the midnight train\n[00:24.15]Static whispers through frayed wires\n[00:27.65]Guitar strings hum our cathedral choirs\n[00:31.30]Flicker screens show reruns of June\n[00:34.90]I'm drowning in this mercury lagoon\n[00:38.55]Electric veins pulse through concrete skies\n[00:42.10]Your name echoes in the hollow where my heartbeat lies\n[00:45.75]We're satellites trapped in parallel light\n[00:49.25]Burning through the atmosphere of endless night\n[01:00.00]Dusty vinyl spins reverse\n[01:03.45]Our polaroid timeline bleeds through the verse\n[01:07.10]Telescope aimed at dead stars\n[01:10.65]Still tracing constellations through prison bars\n[01:14.30]Electric veins pulse through concrete skies\n[01:17.85]Your name echoes in the hollow where my heartbeat lies\n[01:21.50]We're satellites trapped in parallel light\n[01:25.05]Burning through the atmosphere of endless night\n[02:10.00]Clockwork gears grind moonbeams to rust\n[02:13.50]Our fingerprint smudged by interstellar dust\n[02:17.15]Velvet thunder rolls through my veins\n[02:20.70]Chasing phantom trains through solar plane\n[02:24.35]Electric veins pulse through concrete skies\n[02:27.90]Your name echoes in the hollow where my heartbeat lies"""
|
186 |
)
|
187 |
+
|
188 |
+
current_prompt_type = gr.State(value="audio")
|
189 |
+
with gr.Tabs() as inside_tabs:
|
190 |
+
with gr.Tab("Audio Prompt"):
|
191 |
+
audio_prompt = gr.Audio(label="Audio Prompt", type="filepath", value="./src/prompt/default.wav")
|
192 |
+
with gr.Tab("Text Prompt"):
|
193 |
+
text_prompt = gr.Textbox(
|
194 |
+
label="Text Prompt",
|
195 |
+
placeholder="Enter the Text Prompt, eg: emotional piano pop",
|
196 |
+
)
|
197 |
+
def update_prompt_type(evt: gr.SelectData):
|
198 |
+
return "audio" if evt.index == 0 else "text"
|
199 |
+
|
200 |
+
inside_tabs.select(
|
201 |
+
fn=update_prompt_type,
|
202 |
+
outputs=current_prompt_type
|
203 |
+
)
|
204 |
|
205 |
with gr.Column():
|
206 |
with gr.Accordion("Best Practices Guide", open=True):
|
|
|
239 |
steps = gr.Slider(
|
240 |
minimum=10,
|
241 |
maximum=100,
|
242 |
+
value=32,
|
243 |
step=1,
|
244 |
label="Diffusion Steps",
|
245 |
interactive=True,
|
|
|
269 |
examples_per_page=13,
|
270 |
elem_id="audio-examples-container"
|
271 |
)
|
272 |
+
|
273 |
+
gr.Examples(
|
274 |
+
examples=[
|
275 |
+
["Pop Emotional Piano"],
|
276 |
+
["流行 情感 钢琴"],
|
277 |
+
["Indie folk ballad, coming-of-age themes, acoustic guitar picking with harmonica interludes"],
|
278 |
+
["独立民谣, 成长主题, 原声吉他弹奏与口琴间奏"]
|
279 |
+
],
|
280 |
+
inputs=[text_prompt],
|
281 |
+
label="Text Examples",
|
282 |
+
examples_per_page=4,
|
283 |
+
elem_id="text-examples-container"
|
284 |
+
)
|
285 |
|
286 |
gr.Examples(
|
287 |
examples=[
|
|
|
386 |
|
387 |
lyrics_btn.click(
|
388 |
fn=infer_music,
|
389 |
+
inputs=[lrc, audio_prompt, text_prompt, current_prompt_type, seed, randomize_seed, steps, file_type],
|
390 |
outputs=audio_output
|
391 |
)
|
392 |
|
diffrhythm/infer/infer.py
CHANGED
@@ -14,7 +14,7 @@ import pydub
|
|
14 |
from diffrhythm.infer.infer_utils import (
|
15 |
get_reference_latent,
|
16 |
get_lrc_token,
|
17 |
-
|
18 |
prepare_model,
|
19 |
get_negative_style_prompt
|
20 |
)
|
@@ -74,7 +74,7 @@ def decode_audio(latents, vae_model, chunked=False, overlap=32, chunk_size=128):
|
|
74 |
y_final[:,:,t_start:t_end] = y_chunk[:,:,chunk_start:chunk_end]
|
75 |
return y_final
|
76 |
|
77 |
-
def inference(cfm_model, vae_model, cond, text, duration, style_prompt, negative_style_prompt, steps, sway_sampling_coef, start_time, file_type):
|
78 |
|
79 |
with torch.inference_mode():
|
80 |
generated, _ = cfm_model.sample(
|
@@ -86,7 +86,8 @@ def inference(cfm_model, vae_model, cond, text, duration, style_prompt, negative
|
|
86 |
steps=steps,
|
87 |
cfg_strength=4.0,
|
88 |
sway_sampling_coef=sway_sampling_coef,
|
89 |
-
start_time=start_time
|
|
|
90 |
)
|
91 |
|
92 |
generated = generated.to(torch.float32)
|
@@ -133,7 +134,7 @@ if __name__ == "__main__":
|
|
133 |
lrc = f.read()
|
134 |
lrc_prompt, start_time = get_lrc_token(lrc, tokenizer, device)
|
135 |
|
136 |
-
style_prompt =
|
137 |
|
138 |
negative_style_prompt = get_negative_style_prompt(device)
|
139 |
|
|
|
14 |
from diffrhythm.infer.infer_utils import (
|
15 |
get_reference_latent,
|
16 |
get_lrc_token,
|
17 |
+
get_audio_style_prompt,
|
18 |
prepare_model,
|
19 |
get_negative_style_prompt
|
20 |
)
|
|
|
74 |
y_final[:,:,t_start:t_end] = y_chunk[:,:,chunk_start:chunk_end]
|
75 |
return y_final
|
76 |
|
77 |
+
def inference(cfm_model, vae_model, cond, text, duration, style_prompt, negative_style_prompt, steps, sway_sampling_coef, start_time, file_type, vocal_flag):
|
78 |
|
79 |
with torch.inference_mode():
|
80 |
generated, _ = cfm_model.sample(
|
|
|
86 |
steps=steps,
|
87 |
cfg_strength=4.0,
|
88 |
sway_sampling_coef=sway_sampling_coef,
|
89 |
+
start_time=start_time,
|
90 |
+
vocal_flag=vocal_flag,
|
91 |
)
|
92 |
|
93 |
generated = generated.to(torch.float32)
|
|
|
134 |
lrc = f.read()
|
135 |
lrc_prompt, start_time = get_lrc_token(lrc, tokenizer, device)
|
136 |
|
137 |
+
style_prompt = get_audio_style_prompt(muq, args.ref_audio_path)
|
138 |
|
139 |
negative_style_prompt = get_negative_style_prompt(device)
|
140 |
|
diffrhythm/infer/infer_utils.py
CHANGED
@@ -51,13 +51,14 @@ def get_negative_style_prompt(device):
|
|
51 |
|
52 |
return vocal_stlye
|
53 |
|
54 |
-
def
|
|
|
55 |
mulan = model
|
56 |
audio, _ = librosa.load(wav_path, sr=24000)
|
57 |
audio_len = librosa.get_duration(y=audio, sr=24000)
|
58 |
|
59 |
-
|
60 |
-
|
61 |
|
62 |
if audio_len > 10:
|
63 |
start_time = int(audio_len // 2 - 5)
|
@@ -70,10 +71,20 @@ def get_style_prompt(model, wav_path):
|
|
70 |
with torch.no_grad():
|
71 |
audio_emb = mulan(wavs = wav) # [1, 512]
|
72 |
|
73 |
-
audio_emb = audio_emb
|
74 |
audio_emb = audio_emb.half()
|
75 |
|
76 |
-
return audio_emb
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
77 |
|
78 |
def parse_lyrics(lyrics: str):
|
79 |
lyrics_with_time = []
|
@@ -94,7 +105,6 @@ class CNENTokenizer():
|
|
94 |
with open('./diffrhythm/g2p/g2p/vocab.json', 'r') as file:
|
95 |
self.phone2id:dict = json.load(file)['vocab']
|
96 |
self.id2phone = {v:k for (k, v) in self.phone2id.items()}
|
97 |
-
# from f5_tts.g2p.g2p_generation import chn_eng_g2p
|
98 |
from diffrhythm.g2p.g2p_generation import chn_eng_g2p
|
99 |
self.tokenizer = chn_eng_g2p
|
100 |
def encode(self, text):
|
@@ -115,6 +125,8 @@ def get_lrc_token(text, tokenizer, device):
|
|
115 |
pad_token_id = 0
|
116 |
comma_token_id = 1
|
117 |
period_token_id = 2
|
|
|
|
|
118 |
|
119 |
lrc_with_time = parse_lyrics(text)
|
120 |
|
@@ -146,7 +158,7 @@ def get_lrc_token(text, tokenizer, device):
|
|
146 |
frame_start = max(gt_frame_start - frame_shift, last_end_pos)
|
147 |
frame_len = min(num_tokens, max_frames - frame_start)
|
148 |
|
149 |
-
|
150 |
|
151 |
lrc[frame_start:frame_start + frame_len] = tokens[:frame_len]
|
152 |
|
|
|
51 |
|
52 |
return vocal_stlye
|
53 |
|
54 |
+
def get_audio_style_prompt(model, wav_path):
|
55 |
+
vocal_flag = False
|
56 |
mulan = model
|
57 |
audio, _ = librosa.load(wav_path, sr=24000)
|
58 |
audio_len = librosa.get_duration(y=audio, sr=24000)
|
59 |
|
60 |
+
if audio_len <= 1:
|
61 |
+
vocal_flag = True
|
62 |
|
63 |
if audio_len > 10:
|
64 |
start_time = int(audio_len // 2 - 5)
|
|
|
71 |
with torch.no_grad():
|
72 |
audio_emb = mulan(wavs = wav) # [1, 512]
|
73 |
|
|
|
74 |
audio_emb = audio_emb.half()
|
75 |
|
76 |
+
return audio_emb, vocal_flag
|
77 |
+
|
78 |
+
def get_text_style_prompt(model, text_prompt):
|
79 |
+
mulan = model
|
80 |
+
|
81 |
+
with torch.no_grad():
|
82 |
+
text_emb = mulan(texts = text_prompt) # [1, 512]
|
83 |
+
text_emb = text_emb.half()
|
84 |
+
|
85 |
+
return text_emb
|
86 |
+
|
87 |
+
|
88 |
|
89 |
def parse_lyrics(lyrics: str):
|
90 |
lyrics_with_time = []
|
|
|
105 |
with open('./diffrhythm/g2p/g2p/vocab.json', 'r') as file:
|
106 |
self.phone2id:dict = json.load(file)['vocab']
|
107 |
self.id2phone = {v:k for (k, v) in self.phone2id.items()}
|
|
|
108 |
from diffrhythm.g2p.g2p_generation import chn_eng_g2p
|
109 |
self.tokenizer = chn_eng_g2p
|
110 |
def encode(self, text):
|
|
|
125 |
pad_token_id = 0
|
126 |
comma_token_id = 1
|
127 |
period_token_id = 2
|
128 |
+
if text == "":
|
129 |
+
return torch.zeros((max_frames,), dtype=torch.long).unsqueeze(0).to(device), torch.tensor(0.).unsqueeze(0).to(device).half()
|
130 |
|
131 |
lrc_with_time = parse_lyrics(text)
|
132 |
|
|
|
158 |
frame_start = max(gt_frame_start - frame_shift, last_end_pos)
|
159 |
frame_len = min(num_tokens, max_frames - frame_start)
|
160 |
|
161 |
+
|
162 |
|
163 |
lrc[frame_start:frame_start + frame_len] = tokens[:frame_len]
|
164 |
|
diffrhythm/model/cfm.py
CHANGED
@@ -42,10 +42,7 @@ class CFM(nn.Module):
|
|
42 |
transformer: nn.Module,
|
43 |
sigma=0.0,
|
44 |
odeint_kwargs: dict = dict(
|
45 |
-
# atol = 1e-5,
|
46 |
-
# rtol = 1e-5,
|
47 |
method="euler" # 'midpoint'
|
48 |
-
# method="adaptive_heun" # dopri5
|
49 |
),
|
50 |
odeint_options: dict = dict(
|
51 |
min_step=0.05
|
@@ -71,8 +68,6 @@ class CFM(nn.Module):
|
|
71 |
self.style_drop_prob = style_drop_prob
|
72 |
self.lrc_drop_prob = lrc_drop_prob
|
73 |
|
74 |
-
print(f"audio drop prob -> {self.audio_drop_prob}; style_drop_prob -> {self.style_drop_prob}; lrc_drop_prob: {self.lrc_drop_prob}")
|
75 |
-
|
76 |
# transformer
|
77 |
self.transformer = transformer
|
78 |
dim = transformer.dim
|
@@ -83,7 +78,6 @@ class CFM(nn.Module):
|
|
83 |
|
84 |
# sampling related
|
85 |
self.odeint_kwargs = odeint_kwargs
|
86 |
-
# print(f"ODE SOLVER: {self.odeint_kwargs['method']}")
|
87 |
|
88 |
self.odeint_options = odeint_options
|
89 |
|
@@ -120,6 +114,7 @@ class CFM(nn.Module):
|
|
120 |
start_time=None,
|
121 |
latent_pred_start_frame=0,
|
122 |
latent_pred_end_frame=2048,
|
|
|
123 |
):
|
124 |
self.eval()
|
125 |
|
@@ -151,10 +146,9 @@ class CFM(nn.Module):
|
|
151 |
|
152 |
if exists(text):
|
153 |
text_lens = (text != -1).sum(dim=-1)
|
154 |
-
|
155 |
|
156 |
# duration
|
157 |
-
# import pdb; pdb.set_trace()
|
158 |
cond_mask = lens_to_mask(lens)
|
159 |
if edit_mask is not None:
|
160 |
cond_mask = cond_mask & edit_mask
|
@@ -170,7 +164,7 @@ class CFM(nn.Module):
|
|
170 |
if isinstance(duration, int):
|
171 |
duration = torch.full((batch,), duration, device=device, dtype=torch.long)
|
172 |
|
173 |
-
|
174 |
duration = duration.clamp(max=max_duration)
|
175 |
max_duration = duration.amax()
|
176 |
|
@@ -178,12 +172,6 @@ class CFM(nn.Module):
|
|
178 |
if duplicate_test:
|
179 |
test_cond = F.pad(cond, (0, 0, cond_seq_len, max_duration - 2 * cond_seq_len), value=0.0)
|
180 |
|
181 |
-
# cond = F.pad(cond, (0, 0, 0, max_duration - cond_seq_len), value=0.0) # [b, t, d]
|
182 |
-
# cond_mask = F.pad(cond_mask, (0, max_duration - cond_mask.shape[-1]), value=False) # [b, max_duration]
|
183 |
-
# cond_mask = cond_mask.unsqueeze(-1) #[b, t, d]
|
184 |
-
# step_cond = torch.where(
|
185 |
-
# cond_mask, cond, torch.zeros_like(cond)
|
186 |
-
# ) # allow direct control (cut cond audio) with lens passed in
|
187 |
|
188 |
if batch > 1:
|
189 |
mask = lens_to_mask(duration)
|
@@ -197,6 +185,10 @@ class CFM(nn.Module):
|
|
197 |
start_time_embed, positive_text_embed, positive_text_residuals = self.transformer.forward_timestep_invariant(text, step_cond.shape[1], drop_text=False, start_time=start_time)
|
198 |
_, negative_text_embed, negative_text_residuals = self.transformer.forward_timestep_invariant(text, step_cond.shape[1], drop_text=True, start_time=start_time)
|
199 |
|
|
|
|
|
|
|
|
|
200 |
text_embed = torch.cat([positive_text_embed, negative_text_embed], 0)
|
201 |
text_residuals = [torch.cat([a, b], 0) for a, b in zip(positive_text_residuals, negative_text_residuals)]
|
202 |
step_cond = torch.cat([step_cond, step_cond], 0)
|
@@ -242,7 +234,6 @@ class CFM(nn.Module):
|
|
242 |
|
243 |
sampled = trajectory[-1]
|
244 |
out = sampled
|
245 |
-
# out = torch.where(cond_mask, cond, out)
|
246 |
out = torch.where(fixed_span_mask, out, cond)
|
247 |
|
248 |
if exists(vocoder):
|
@@ -286,7 +277,6 @@ class CFM(nn.Module):
|
|
286 |
x0 = torch.randn_like(x1)
|
287 |
|
288 |
# time step
|
289 |
-
# time = torch.rand((batch,), dtype=dtype, device=self.device)
|
290 |
time = torch.normal(mean=0, std=1, size=(batch,), device=self.device)
|
291 |
time = torch.nn.functional.sigmoid(time)
|
292 |
# TODO. noise_scheduler
|
|
|
42 |
transformer: nn.Module,
|
43 |
sigma=0.0,
|
44 |
odeint_kwargs: dict = dict(
|
|
|
|
|
45 |
method="euler" # 'midpoint'
|
|
|
46 |
),
|
47 |
odeint_options: dict = dict(
|
48 |
min_step=0.05
|
|
|
68 |
self.style_drop_prob = style_drop_prob
|
69 |
self.lrc_drop_prob = lrc_drop_prob
|
70 |
|
|
|
|
|
71 |
# transformer
|
72 |
self.transformer = transformer
|
73 |
dim = transformer.dim
|
|
|
78 |
|
79 |
# sampling related
|
80 |
self.odeint_kwargs = odeint_kwargs
|
|
|
81 |
|
82 |
self.odeint_options = odeint_options
|
83 |
|
|
|
114 |
start_time=None,
|
115 |
latent_pred_start_frame=0,
|
116 |
latent_pred_end_frame=2048,
|
117 |
+
vocal_flag=False
|
118 |
):
|
119 |
self.eval()
|
120 |
|
|
|
146 |
|
147 |
if exists(text):
|
148 |
text_lens = (text != -1).sum(dim=-1)
|
149 |
+
|
150 |
|
151 |
# duration
|
|
|
152 |
cond_mask = lens_to_mask(lens)
|
153 |
if edit_mask is not None:
|
154 |
cond_mask = cond_mask & edit_mask
|
|
|
164 |
if isinstance(duration, int):
|
165 |
duration = torch.full((batch,), duration, device=device, dtype=torch.long)
|
166 |
|
167 |
+
|
168 |
duration = duration.clamp(max=max_duration)
|
169 |
max_duration = duration.amax()
|
170 |
|
|
|
172 |
if duplicate_test:
|
173 |
test_cond = F.pad(cond, (0, 0, cond_seq_len, max_duration - 2 * cond_seq_len), value=0.0)
|
174 |
|
|
|
|
|
|
|
|
|
|
|
|
|
175 |
|
176 |
if batch > 1:
|
177 |
mask = lens_to_mask(duration)
|
|
|
185 |
start_time_embed, positive_text_embed, positive_text_residuals = self.transformer.forward_timestep_invariant(text, step_cond.shape[1], drop_text=False, start_time=start_time)
|
186 |
_, negative_text_embed, negative_text_residuals = self.transformer.forward_timestep_invariant(text, step_cond.shape[1], drop_text=True, start_time=start_time)
|
187 |
|
188 |
+
if vocal_flag:
|
189 |
+
style_prompt = negative_style_prompt
|
190 |
+
negative_style_prompt = torch.zeros_like(style_prompt)
|
191 |
+
|
192 |
text_embed = torch.cat([positive_text_embed, negative_text_embed], 0)
|
193 |
text_residuals = [torch.cat([a, b], 0) for a, b in zip(positive_text_residuals, negative_text_residuals)]
|
194 |
step_cond = torch.cat([step_cond, step_cond], 0)
|
|
|
234 |
|
235 |
sampled = trajectory[-1]
|
236 |
out = sampled
|
|
|
237 |
out = torch.where(fixed_span_mask, out, cond)
|
238 |
|
239 |
if exists(vocoder):
|
|
|
277 |
x0 = torch.randn_like(x1)
|
278 |
|
279 |
# time step
|
|
|
280 |
time = torch.normal(mean=0, std=1, size=(batch,), device=self.device)
|
281 |
time = torch.nn.functional.sigmoid(time)
|
282 |
# TODO. noise_scheduler
|
diffrhythm/model/dit.py
CHANGED
@@ -13,8 +13,6 @@ import torch
|
|
13 |
from torch import nn
|
14 |
import torch
|
15 |
import torch.nn.functional as F
|
16 |
-
|
17 |
-
from x_transformers.x_transformers import RotaryEmbedding
|
18 |
from transformers.models.llama.modeling_llama import LlamaDecoderLayer, LlamaRotaryEmbedding
|
19 |
from transformers.models.llama import LlamaConfig
|
20 |
from torch.utils.checkpoint import checkpoint
|
@@ -32,8 +30,6 @@ from diffrhythm.model.modules import (
|
|
32 |
# apply_liger_kernel_to_llama()
|
33 |
|
34 |
# Text embedding
|
35 |
-
|
36 |
-
|
37 |
class TextEmbedding(nn.Module):
|
38 |
def __init__(self, text_num_embeds, text_dim, conv_layers=0, conv_mult=2):
|
39 |
super().__init__()
|
@@ -50,10 +46,7 @@ class TextEmbedding(nn.Module):
|
|
50 |
self.extra_modeling = False
|
51 |
|
52 |
def forward(self, text: int["b nt"], seq_len, drop_text=False): # noqa: F722
|
53 |
-
#text = text + 1 # use 0 as filler token. preprocess of batch pad -1, see list_str_to_idx()
|
54 |
-
#text = text[:, :seq_len] # curtail if character tokens are more than the mel spec tokens
|
55 |
batch, text_len = text.shape[0], text.shape[1]
|
56 |
-
#text = F.pad(text, (0, seq_len - text_len), value=0)
|
57 |
|
58 |
if drop_text: # cfg for text
|
59 |
text = torch.zeros_like(text)
|
@@ -75,8 +68,6 @@ class TextEmbedding(nn.Module):
|
|
75 |
|
76 |
|
77 |
# noised input audio and context mixing embedding
|
78 |
-
|
79 |
-
|
80 |
class InputEmbedding(nn.Module):
|
81 |
def __init__(self, mel_dim, text_dim, out_dim, cond_dim):
|
82 |
super().__init__()
|
@@ -89,7 +80,6 @@ class InputEmbedding(nn.Module):
|
|
89 |
|
90 |
style_emb = style_emb.unsqueeze(1).repeat(1, x.shape[1], 1)
|
91 |
time_emb = time_emb.unsqueeze(1).repeat(1, x.shape[1], 1)
|
92 |
-
# print(x.shape, cond.shape, text_embed.shape, style_emb.shape, time_emb.shape)
|
93 |
x = self.proj(torch.cat((x, cond, text_embed, style_emb, time_emb), dim=-1))
|
94 |
x = self.conv_pos_embed(x) + x
|
95 |
return x
|
@@ -125,17 +115,13 @@ class DiT(nn.Module):
|
|
125 |
self.text_embed = TextEmbedding(text_num_embeds, text_dim, conv_layers=conv_layers)
|
126 |
self.input_embed = InputEmbedding(mel_dim, text_dim, dim, cond_dim=cond_dim)
|
127 |
|
128 |
-
#self.rotary_embed = RotaryEmbedding(dim_head)
|
129 |
|
130 |
self.dim = dim
|
131 |
self.depth = depth
|
132 |
|
133 |
-
#self.transformer_blocks = nn.ModuleList(
|
134 |
-
# [DiTBlock(dim=dim, heads=heads, dim_head=dim_head, ff_mult=ff_mult, dropout=dropout, use_style_prompt=use_style_prompt) for _ in range(depth)]
|
135 |
-
#)
|
136 |
llama_config = LlamaConfig(hidden_size=dim, intermediate_size=dim * ff_mult, hidden_act='silu')
|
137 |
llama_config._attn_implementation = 'sdpa'
|
138 |
-
|
139 |
self.transformer_blocks = nn.ModuleList(
|
140 |
[LlamaDecoderLayer(llama_config, layer_idx=i) for i in range(depth)]
|
141 |
)
|
@@ -157,8 +143,6 @@ class DiT(nn.Module):
|
|
157 |
self.norm_out = AdaLayerNormZero_Final(dim, cond_dim) # final modulation
|
158 |
self.proj_out = nn.Linear(dim, mel_dim)
|
159 |
|
160 |
-
# if use_style_prompt:
|
161 |
-
# self.prompt_rnn = nn.LSTM(64, cond_dim, 1, batch_first=True)
|
162 |
|
163 |
def forward_timestep_invariant(self, text, seq_len, drop_text, start_time):
|
164 |
s_t = self.start_time_embed(start_time)
|
|
|
13 |
from torch import nn
|
14 |
import torch
|
15 |
import torch.nn.functional as F
|
|
|
|
|
16 |
from transformers.models.llama.modeling_llama import LlamaDecoderLayer, LlamaRotaryEmbedding
|
17 |
from transformers.models.llama import LlamaConfig
|
18 |
from torch.utils.checkpoint import checkpoint
|
|
|
30 |
# apply_liger_kernel_to_llama()
|
31 |
|
32 |
# Text embedding
|
|
|
|
|
33 |
class TextEmbedding(nn.Module):
|
34 |
def __init__(self, text_num_embeds, text_dim, conv_layers=0, conv_mult=2):
|
35 |
super().__init__()
|
|
|
46 |
self.extra_modeling = False
|
47 |
|
48 |
def forward(self, text: int["b nt"], seq_len, drop_text=False): # noqa: F722
|
|
|
|
|
49 |
batch, text_len = text.shape[0], text.shape[1]
|
|
|
50 |
|
51 |
if drop_text: # cfg for text
|
52 |
text = torch.zeros_like(text)
|
|
|
68 |
|
69 |
|
70 |
# noised input audio and context mixing embedding
|
|
|
|
|
71 |
class InputEmbedding(nn.Module):
|
72 |
def __init__(self, mel_dim, text_dim, out_dim, cond_dim):
|
73 |
super().__init__()
|
|
|
80 |
|
81 |
style_emb = style_emb.unsqueeze(1).repeat(1, x.shape[1], 1)
|
82 |
time_emb = time_emb.unsqueeze(1).repeat(1, x.shape[1], 1)
|
|
|
83 |
x = self.proj(torch.cat((x, cond, text_embed, style_emb, time_emb), dim=-1))
|
84 |
x = self.conv_pos_embed(x) + x
|
85 |
return x
|
|
|
115 |
self.text_embed = TextEmbedding(text_num_embeds, text_dim, conv_layers=conv_layers)
|
116 |
self.input_embed = InputEmbedding(mel_dim, text_dim, dim, cond_dim=cond_dim)
|
117 |
|
|
|
118 |
|
119 |
self.dim = dim
|
120 |
self.depth = depth
|
121 |
|
|
|
|
|
|
|
122 |
llama_config = LlamaConfig(hidden_size=dim, intermediate_size=dim * ff_mult, hidden_act='silu')
|
123 |
llama_config._attn_implementation = 'sdpa'
|
124 |
+
|
125 |
self.transformer_blocks = nn.ModuleList(
|
126 |
[LlamaDecoderLayer(llama_config, layer_idx=i) for i in range(depth)]
|
127 |
)
|
|
|
143 |
self.norm_out = AdaLayerNormZero_Final(dim, cond_dim) # final modulation
|
144 |
self.proj_out = nn.Linear(dim, mel_dim)
|
145 |
|
|
|
|
|
146 |
|
147 |
def forward_timestep_invariant(self, text, seq_len, drop_text, start_time):
|
148 |
s_t = self.start_time_embed(start_time)
|