jbetker commited on
Commit
b78ae92
1 Parent(s): c66954b

Upgrade CLIP model and add eval_multiple

Browse files
api.py ADDED
@@ -0,0 +1,214 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import os
3
+ import random
4
+ from urllib import request
5
+
6
+ import torch
7
+ import torch.nn.functional as F
8
+ import torchaudio
9
+ import progressbar
10
+ import ocotillo
11
+
12
+ from models.diffusion_decoder import DiffusionTts
13
+ from models.autoregressive import UnifiedVoice
14
+ from tqdm import tqdm
15
+
16
+ from models.arch_util import TorchMelSpectrogram
17
+ from models.text_voice_clip import VoiceCLIP
18
+ from models.vocoder import UnivNetGenerator
19
+ from utils.audio import load_audio, wav_to_univnet_mel, denormalize_tacotron_mel
20
+ from utils.diffusion import SpacedDiffusion, space_timesteps, get_named_beta_schedule
21
+ from utils.tokenizer import VoiceBpeTokenizer, lev_distance
22
+
23
+
24
+ pbar = None
25
+ def download_models():
26
+ MODELS = {
27
+ 'clip.pth': 'https://huggingface.co/jbetker/tortoise-tts-clip/resolve/main/pytorch-model.bin',
28
+ 'diffusion.pth': 'https://huggingface.co/jbetker/tortoise-tts-diffusion-v1/resolve/main/pytorch-model.bin',
29
+ 'autoregressive.pth': 'https://huggingface.co/jbetker/tortoise-tts-autoregressive/resolve/main/pytorch-model.bin'
30
+ }
31
+ os.makedirs('.models', exist_ok=True)
32
+ def show_progress(block_num, block_size, total_size):
33
+ global pbar
34
+ if pbar is None:
35
+ pbar = progressbar.ProgressBar(maxval=total_size)
36
+ pbar.start()
37
+
38
+ downloaded = block_num * block_size
39
+ if downloaded < total_size:
40
+ pbar.update(downloaded)
41
+ else:
42
+ pbar.finish()
43
+ pbar = None
44
+ for model_name, url in MODELS.items():
45
+ if os.path.exists(f'.models/{model_name}'):
46
+ continue
47
+ print(f'Downloading {model_name} from {url}...')
48
+ request.urlretrieve(url, f'.models/{model_name}', show_progress)
49
+ print('Done.')
50
+
51
+
52
+ def load_discrete_vocoder_diffuser(trained_diffusion_steps=4000, desired_diffusion_steps=200, cond_free=True):
53
+ """
54
+ Helper function to load a GaussianDiffusion instance configured for use as a vocoder.
55
+ """
56
+ return SpacedDiffusion(use_timesteps=space_timesteps(trained_diffusion_steps, [desired_diffusion_steps]), model_mean_type='epsilon',
57
+ model_var_type='learned_range', loss_type='mse', betas=get_named_beta_schedule('linear', trained_diffusion_steps),
58
+ conditioning_free=cond_free, conditioning_free_k=1)
59
+
60
+
61
+ def load_conditioning(clip, cond_length=132300):
62
+ gap = clip.shape[-1] - cond_length
63
+ if gap < 0:
64
+ clip = F.pad(clip, pad=(0, abs(gap)))
65
+ elif gap > 0:
66
+ rand_start = random.randint(0, gap)
67
+ clip = clip[:, rand_start:rand_start + cond_length]
68
+ mel_clip = TorchMelSpectrogram()(clip.unsqueeze(0)).squeeze(0)
69
+ return mel_clip.unsqueeze(0).cuda()
70
+
71
+
72
+ def fix_autoregressive_output(codes, stop_token):
73
+ """
74
+ This function performs some padding on coded audio that fixes a mismatch issue between what the diffusion model was
75
+ trained on and what the autoregressive code generator creates (which has no padding or end).
76
+ This is highly specific to the DVAE being used, so this particular coding will not necessarily work if used with
77
+ a different DVAE. This can be inferred by feeding a audio clip padded with lots of zeros on the end through the DVAE
78
+ and copying out the last few codes.
79
+
80
+ Failing to do this padding will produce speech with a harsh end that sounds like "BLAH" or similar.
81
+ """
82
+ # Strip off the autoregressive stop token and add padding.
83
+ stop_token_indices = (codes == stop_token).nonzero()
84
+ if len(stop_token_indices) == 0:
85
+ print("No stop tokens found, enjoy that output of yours!")
86
+ return codes
87
+ else:
88
+ codes[stop_token_indices] = 83
89
+ stm = stop_token_indices.min().item()
90
+ codes[stm:] = 83
91
+ if stm - 3 < codes.shape[0]:
92
+ codes[-3] = 45
93
+ codes[-2] = 45
94
+ codes[-1] = 248
95
+
96
+ return codes
97
+
98
+
99
+ def do_spectrogram_diffusion(diffusion_model, diffuser, mel_codes, conditioning_input, mean=False):
100
+ """
101
+ Uses the specified diffusion model and DVAE model to convert the provided MEL & conditioning inputs into an audio clip.
102
+ """
103
+ with torch.no_grad():
104
+ cond_mel = wav_to_univnet_mel(conditioning_input.squeeze(1), do_normalization=False)
105
+ # Pad MEL to multiples of 32
106
+ msl = mel_codes.shape[-1]
107
+ dsl = 32
108
+ gap = dsl - (msl % dsl)
109
+ if gap > 0:
110
+ mel = torch.nn.functional.pad(mel_codes, (0, gap))
111
+
112
+ output_shape = (mel.shape[0], 100, mel.shape[-1]*4)
113
+ precomputed_embeddings = diffusion_model.timestep_independent(mel_codes, cond_mel)
114
+ if mean:
115
+ mel = diffuser.p_sample_loop(diffusion_model, output_shape, noise=torch.zeros(output_shape, device=mel_codes.device),
116
+ model_kwargs={'precomputed_aligned_embeddings': precomputed_embeddings})
117
+ else:
118
+ mel = diffuser.p_sample_loop(diffusion_model, output_shape, model_kwargs={'precomputed_aligned_embeddings': precomputed_embeddings})
119
+ return denormalize_tacotron_mel(mel)[:,:,:msl*4]
120
+
121
+
122
+ class TextToSpeech:
123
+ def __init__(self, autoregressive_batch_size=32):
124
+ self.autoregressive_batch_size = autoregressive_batch_size
125
+ self.tokenizer = VoiceBpeTokenizer()
126
+ download_models()
127
+
128
+ self.autoregressive = UnifiedVoice(max_mel_tokens=300, max_text_tokens=200, max_conditioning_inputs=2, layers=30,
129
+ model_dim=1024,
130
+ heads=16, number_text_tokens=256, start_text_token=255, checkpointing=False,
131
+ train_solo_embeddings=False,
132
+ average_conditioning_embeddings=True).cpu().eval()
133
+ self.autoregressive.load_state_dict(torch.load('.models/autoregressive.pth'))
134
+
135
+ self.clip = VoiceCLIP(dim_text=512, dim_speech=512, dim_latent=512, num_text_tokens=256, text_enc_depth=12,
136
+ text_seq_len=350, text_heads=8,
137
+ num_speech_tokens=8192, speech_enc_depth=12, speech_heads=8, speech_seq_len=430,
138
+ use_xformers=True).cpu().eval()
139
+ self.clip.load_state_dict(torch.load('.models/clip.pth'))
140
+
141
+ self.diffusion = DiffusionTts(model_channels=512, in_channels=100, out_channels=200, in_latent_channels=1024,
142
+ channel_mult=[1, 2, 3, 4], num_res_blocks=[3, 3, 3, 3],
143
+ token_conditioning_resolutions=[1, 4, 8],
144
+ dropout=0, attention_resolutions=[4, 8], num_heads=8, kernel_size=3, scale_factor=2,
145
+ time_embed_dim_multiplier=4, unconditioned_percentage=0, conditioning_dim_factor=2,
146
+ conditioning_expansion=1).cpu().eval()
147
+ self.diffusion.load_state_dict(torch.load('.models/diffusion.pth'))
148
+
149
+ self.vocoder = UnivNetGenerator().cpu()
150
+ self.vocoder.load_state_dict(torch.load('.models/vocoder.pth')['model_g'])
151
+ self.vocoder.eval(inference=True)
152
+
153
+ def tts(self, text, voice_samples, num_autoregressive_samples=512, k=1, diffusion_iterations=100, cond_free=True):
154
+ text = torch.IntTensor(self.tokenizer.encode(text)).unsqueeze(0).cuda()
155
+ text = F.pad(text, (0, 1)) # This may not be necessary.
156
+
157
+ conds = []
158
+ if not isinstance(voice_samples, list):
159
+ voice_samples = [voice_samples]
160
+ for vs in voice_samples:
161
+ conds.append(load_conditioning(vs))
162
+ conds = torch.stack(conds, dim=1)
163
+ cond_diffusion = voice_samples[0].cuda()
164
+ # The diffusion model expects = 88200 conditioning samples.
165
+ if cond_diffusion.shape[-1] < 88200:
166
+ cond_diffusion = F.pad(cond_diffusion, (0, 88200-cond_diffusion.shape[-1]))
167
+ else:
168
+ cond_diffusion = cond_diffusion[:, :88200]
169
+
170
+ diffuser = load_discrete_vocoder_diffuser(desired_diffusion_steps=diffusion_iterations, cond_free=cond_free)
171
+
172
+ with torch.no_grad():
173
+ samples = []
174
+ num_batches = num_autoregressive_samples // self.autoregressive_batch_size
175
+ stop_mel_token = self.autoregressive.stop_mel_token
176
+ self.autoregressive = self.autoregressive.cuda()
177
+ for b in tqdm(range(num_batches)):
178
+ codes = self.autoregressive.inference_speech(conds, text, num_beams=1, repetition_penalty=1.0, do_sample=True,
179
+ top_k=50, top_p=.95,
180
+ temperature=.9,
181
+ num_return_sequences=self.autoregressive_batch_size,
182
+ length_penalty=1)
183
+ padding_needed = 250 - codes.shape[1]
184
+ codes = F.pad(codes, (0, padding_needed), value=stop_mel_token)
185
+ samples.append(codes)
186
+ self.autoregressive = self.autoregressive.cpu()
187
+
188
+ clip_results = []
189
+ self.clip = self.clip.cuda()
190
+ for batch in samples:
191
+ for i in range(batch.shape[0]):
192
+ batch[i] = fix_autoregressive_output(batch[i], stop_mel_token)
193
+ clip_results.append(self.clip(text.repeat(batch.shape[0], 1), batch, return_loss=False))
194
+ clip_results = torch.cat(clip_results, dim=0)
195
+ samples = torch.cat(samples, dim=0)
196
+ best_results = samples[torch.topk(clip_results, k=k).indices]
197
+ self.clip = self.clip.cpu()
198
+ del samples
199
+
200
+ print("Performing vocoding..")
201
+ wav_candidates = []
202
+ self.diffusion = self.diffusion.cuda()
203
+ self.vocoder = self.vocoder.cuda()
204
+ for b in range(best_results.shape[0]):
205
+ code = best_results[b].unsqueeze(0)
206
+ mel = do_spectrogram_diffusion(self.diffusion, diffuser, code, cond_diffusion, mean=False)
207
+ wav = self.vocoder.inference(mel)
208
+ wav_candidates.append(wav.cpu())
209
+ self.diffusion = self.diffusion.cpu()
210
+ self.vocoder = self.vocoder.cpu()
211
+
212
+ if len(wav_candidates) > 1:
213
+ return wav_candidates
214
+ return wav_candidates[0]
do_tts.py CHANGED
@@ -138,8 +138,8 @@ if __name__ == '__main__':
138
  parser = argparse.ArgumentParser()
139
  parser.add_argument('-text', type=str, help='Text to speak.', default="I am a language model that has learned to speak.")
140
  parser.add_argument('-voice', type=str, help='Use a preset conditioning voice (defined above). Overrides cond_path.', default='dotrice,harris,lescault,otto,atkins,grace,kennard,mol')
141
- parser.add_argument('-num_samples', type=int, help='How many total outputs the autoregressive transformer should produce.', default=1024)
142
- parser.add_argument('-num_batches', type=int, help='How many batches those samples should be produced over.', default=32)
143
  parser.add_argument('-num_diffusion_samples', type=int, help='Number of outputs that progress to the diffusion stage.', default=16)
144
  parser.add_argument('-output_path', type=str, help='Where to store outputs.', default='results/')
145
  args = parser.parse_args()
@@ -179,19 +179,15 @@ if __name__ == '__main__':
179
  del autoregressive
180
 
181
  print("Loading CLIP..")
182
- clip = VoiceCLIP(dim_text=512, dim_speech=512, dim_latent=512, num_text_tokens=256, text_enc_depth=8, text_seq_len=120, text_heads=8,
183
- num_speech_tokens=8192, speech_enc_depth=10, speech_heads=8, speech_seq_len=250).cuda().eval()
184
  clip.load_state_dict(torch.load('.models/clip.pth'))
185
  print("Performing CLIP filtering..")
186
  clip_results = []
187
  for batch in samples:
188
  for i in range(batch.shape[0]):
189
  batch[i] = fix_autoregressive_output(batch[i], stop_mel_token)
190
- text = text[:, :120] # Ugly hack to fix the fact that I didn't train CLIP to handle long enough text.
191
- clip_results.append(clip(text.repeat(batch.shape[0], 1),
192
- torch.full((batch.shape[0],), fill_value=text.shape[1]-1, dtype=torch.long, device='cuda'),
193
- batch, torch.full((batch.shape[0],), fill_value=batch.shape[1]*1024, dtype=torch.long, device='cuda'),
194
- return_loss=False))
195
  clip_results = torch.cat(clip_results, dim=0)
196
  samples = torch.cat(samples, dim=0)
197
  best_results = samples[torch.topk(clip_results, k=args.num_diffusion_samples).indices]
 
138
  parser = argparse.ArgumentParser()
139
  parser.add_argument('-text', type=str, help='Text to speak.', default="I am a language model that has learned to speak.")
140
  parser.add_argument('-voice', type=str, help='Use a preset conditioning voice (defined above). Overrides cond_path.', default='dotrice,harris,lescault,otto,atkins,grace,kennard,mol')
141
+ parser.add_argument('-num_samples', type=int, help='How many total outputs the autoregressive transformer should produce.', default=512)
142
+ parser.add_argument('-num_batches', type=int, help='How many batches those samples should be produced over.', default=16)
143
  parser.add_argument('-num_diffusion_samples', type=int, help='Number of outputs that progress to the diffusion stage.', default=16)
144
  parser.add_argument('-output_path', type=str, help='Where to store outputs.', default='results/')
145
  args = parser.parse_args()
 
179
  del autoregressive
180
 
181
  print("Loading CLIP..")
182
+ clip = VoiceCLIP(dim_text=512, dim_speech=512, dim_latent=512, num_text_tokens=256, text_enc_depth=12, text_seq_len=350, text_heads=8,
183
+ num_speech_tokens=8192, speech_enc_depth=12, speech_heads=8, speech_seq_len=430, use_xformers=True).cuda().eval()
184
  clip.load_state_dict(torch.load('.models/clip.pth'))
185
  print("Performing CLIP filtering..")
186
  clip_results = []
187
  for batch in samples:
188
  for i in range(batch.shape[0]):
189
  batch[i] = fix_autoregressive_output(batch[i], stop_mel_token)
190
+ clip_results.append(clip(text.repeat(batch.shape[0], 1), batch, return_loss=False))
 
 
 
 
191
  clip_results = torch.cat(clip_results, dim=0)
192
  samples = torch.cat(samples, dim=0)
193
  best_results = samples[torch.topk(clip_results, k=args.num_diffusion_samples).indices]
eval_multiple.py ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+
3
+ import torchaudio
4
+
5
+ from api import TextToSpeech
6
+ from utils.audio import load_audio
7
+
8
+ if __name__ == '__main__':
9
+ fname = 'Y:\\libritts\\test-clean\\transcribed-brief-w2v.tsv'
10
+ outpath = 'D:\\tmp\\tortoise-tts-eval\\baseline'
11
+ outpath_real = 'D:\\tmp\\tortoise-tts-eval\\real'
12
+
13
+ os.makedirs(outpath, exist_ok=True)
14
+ os.makedirs(outpath_real, exist_ok=True)
15
+ with open(fname, 'r', encoding='utf-8') as f:
16
+ lines = [l.strip().split('\t') for l in f.readlines()]
17
+
18
+ recorder = open(os.path.join(outpath, 'transcript.tsv'), 'w', encoding='utf-8')
19
+ tts = TextToSpeech()
20
+ for e, line in enumerate(lines):
21
+ transcript = line[0]
22
+ if len(transcript) > 120:
23
+ continue # We need to support this, but cannot yet.
24
+ path = os.path.join(os.path.dirname(fname), line[1])
25
+ cond_audio = load_audio(path, 22050)
26
+ torchaudio.save(os.path.join(outpath_real, os.path.basename(line[1])), cond_audio, 22050)
27
+ sample = tts.tts(transcript, [cond_audio, cond_audio], num_autoregressive_samples=512, k=1, diffusion_iterations=200, cond_free=True)
28
+ down = torchaudio.functional.resample(sample, 24000, 22050)
29
+ fout_path = os.path.join(outpath, os.path.basename(line[1]))
30
+ torchaudio.save(fout_path, down.squeeze(0), 22050)
31
+ recorder.write(f'{transcript}\t{fout_path}\n')
32
+ recorder.flush()
33
+ recorder.close()
models/arch_util.py CHANGED
@@ -1,9 +1,11 @@
 
1
  import math
2
 
3
  import torch
4
  import torch.nn as nn
5
  import torch.nn.functional as F
6
  import torchaudio
 
7
 
8
 
9
  def zero_module(module):
@@ -316,4 +318,46 @@ class TorchMelSpectrogram(nn.Module):
316
  if self.mel_norms is not None:
317
  self.mel_norms = self.mel_norms.to(mel.device)
318
  mel = mel / self.mel_norms.unsqueeze(0).unsqueeze(-1)
319
- return mel
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import functools
2
  import math
3
 
4
  import torch
5
  import torch.nn as nn
6
  import torch.nn.functional as F
7
  import torchaudio
8
+ from x_transformers import ContinuousTransformerWrapper
9
 
10
 
11
  def zero_module(module):
 
318
  if self.mel_norms is not None:
319
  self.mel_norms = self.mel_norms.to(mel.device)
320
  mel = mel / self.mel_norms.unsqueeze(0).unsqueeze(-1)
321
+ return mel
322
+
323
+
324
+ class CheckpointedLayer(nn.Module):
325
+ """
326
+ Wraps a module. When forward() is called, passes kwargs that require_grad through torch.checkpoint() and bypasses
327
+ checkpoint for all other args.
328
+ """
329
+ def __init__(self, wrap):
330
+ super().__init__()
331
+ self.wrap = wrap
332
+
333
+ def forward(self, x, *args, **kwargs):
334
+ for k, v in kwargs.items():
335
+ assert not (isinstance(v, torch.Tensor) and v.requires_grad) # This would screw up checkpointing.
336
+ partial = functools.partial(self.wrap, **kwargs)
337
+ return torch.utils.checkpoint.checkpoint(partial, x, *args)
338
+
339
+
340
+ class CheckpointedXTransformerEncoder(nn.Module):
341
+ """
342
+ Wraps a ContinuousTransformerWrapper and applies CheckpointedLayer to each layer and permutes from channels-mid
343
+ to channels-last that XTransformer expects.
344
+ """
345
+ def __init__(self, needs_permute=True, exit_permute=True, checkpoint=True, **xtransformer_kwargs):
346
+ super().__init__()
347
+ self.transformer = ContinuousTransformerWrapper(**xtransformer_kwargs)
348
+ self.needs_permute = needs_permute
349
+ self.exit_permute = exit_permute
350
+
351
+ if not checkpoint:
352
+ return
353
+ for i in range(len(self.transformer.attn_layers.layers)):
354
+ n, b, r = self.transformer.attn_layers.layers[i]
355
+ self.transformer.attn_layers.layers[i] = nn.ModuleList([n, CheckpointedLayer(b), r])
356
+
357
+ def forward(self, x, **kwargs):
358
+ if self.needs_permute:
359
+ x = x.permute(0,2,1)
360
+ h = self.transformer(x, **kwargs)
361
+ if self.exit_permute:
362
+ h = h.permute(0,2,1)
363
+ return h
models/diffusion_decoder.py CHANGED
@@ -15,7 +15,8 @@ from torch.nn import Linear
15
  from torch.utils.checkpoint import checkpoint
16
  from x_transformers import ContinuousTransformerWrapper, Encoder
17
 
18
- from models.arch_util import normalization, zero_module, Downsample, Upsample, AudioMiniEncoder, AttentionBlock
 
19
 
20
 
21
  def is_latent(t):
@@ -157,43 +158,6 @@ class ResBlock(TimestepBlock):
157
  return self.skip_connection(x) + h
158
 
159
 
160
- class CheckpointedLayer(nn.Module):
161
- """
162
- Wraps a module. When forward() is called, passes kwargs that require_grad through torch.checkpoint() and bypasses
163
- checkpoint for all other args.
164
- """
165
- def __init__(self, wrap):
166
- super().__init__()
167
- self.wrap = wrap
168
-
169
- def forward(self, x, *args, **kwargs):
170
- for k, v in kwargs.items():
171
- assert not (isinstance(v, torch.Tensor) and v.requires_grad) # This would screw up checkpointing.
172
- partial = functools.partial(self.wrap, **kwargs)
173
- return torch.utils.checkpoint.checkpoint(partial, x, *args)
174
-
175
-
176
- class CheckpointedXTransformerEncoder(nn.Module):
177
- """
178
- Wraps a ContinuousTransformerWrapper and applies CheckpointedLayer to each layer and permutes from channels-mid
179
- to channels-last that XTransformer expects.
180
- """
181
- def __init__(self, needs_permute=True, **xtransformer_kwargs):
182
- super().__init__()
183
- self.transformer = ContinuousTransformerWrapper(**xtransformer_kwargs)
184
- self.needs_permute = needs_permute
185
-
186
- for i in range(len(self.transformer.attn_layers.layers)):
187
- n, b, r = self.transformer.attn_layers.layers[i]
188
- self.transformer.attn_layers.layers[i] = nn.ModuleList([n, CheckpointedLayer(b), r])
189
-
190
- def forward(self, x, **kwargs):
191
- if self.needs_permute:
192
- x = x.permute(0,2,1)
193
- h = self.transformer(x, **kwargs)
194
- return h.permute(0,2,1)
195
-
196
-
197
  class DiffusionTts(nn.Module):
198
  """
199
  The full UNet model with attention and timestep embedding.
 
15
  from torch.utils.checkpoint import checkpoint
16
  from x_transformers import ContinuousTransformerWrapper, Encoder
17
 
18
+ from models.arch_util import normalization, zero_module, Downsample, Upsample, AudioMiniEncoder, AttentionBlock, \
19
+ CheckpointedXTransformerEncoder
20
 
21
 
22
  def is_latent(t):
 
158
  return self.skip_connection(x) + h
159
 
160
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
161
  class DiffusionTts(nn.Module):
162
  """
163
  The full UNet model with attention and timestep embedding.
models/text_voice_clip.py CHANGED
@@ -2,6 +2,9 @@ import torch
2
  import torch.nn as nn
3
  import torch.nn.functional as F
4
  from torch import einsum
 
 
 
5
  from models.transformer import Transformer
6
 
7
 
@@ -13,7 +16,6 @@ def masked_mean(t, mask, dim = 1):
13
  t = t.masked_fill(~mask[:, :, None], 0.)
14
  return t.sum(dim = 1) / mask.sum(dim = 1)[..., None]
15
 
16
-
17
  class VoiceCLIP(nn.Module):
18
  """
19
  CLIP model retrofitted for performing contrastive evaluation between tokenized audio data and the corresponding
@@ -39,40 +41,69 @@ class VoiceCLIP(nn.Module):
39
  text_mask_percentage=0,
40
  voice_mask_percentage=0,
41
  wav_token_compression=1024,
 
42
  ):
43
  super().__init__()
44
  self.text_emb = nn.Embedding(num_text_tokens, dim_text)
45
- self.text_pos_emb = nn.Embedding(text_seq_len, dim_text)
46
- self.text_transformer = Transformer(causal=False, seq_len=text_seq_len, dim=dim_text, depth=text_enc_depth,
47
- heads=text_heads)
48
  self.to_text_latent = nn.Linear(dim_text, dim_latent, bias=False)
49
 
50
  self.speech_emb = nn.Embedding(num_speech_tokens, dim_speech)
51
- self.speech_pos_emb = nn.Embedding(num_speech_tokens, dim_speech)
52
- self.speech_transformer = Transformer(causal=False, seq_len=speech_seq_len, dim=dim_speech,
53
- depth=speech_enc_depth, heads=speech_heads)
54
  self.to_speech_latent = nn.Linear(dim_speech, dim_latent, bias=False)
55
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
56
  self.temperature = nn.Parameter(torch.tensor(1.))
57
  self.text_mask_percentage = text_mask_percentage
58
  self.voice_mask_percentage = voice_mask_percentage
59
  self.wav_token_compression = wav_token_compression
 
 
 
 
60
 
61
  def forward(
62
  self,
63
  text,
64
- text_lengths,
65
  speech_tokens,
66
- wav_lengths,
67
  return_loss=False
68
  ):
69
- # This model will receive micro-batches with a ton of padding for both the text and MELs. Ameliorate this by
70
- # chopping the inputs by the maximum actual length.
71
- max_text_len = text_lengths.max()
72
- text = text[:, :max_text_len]
73
- max_mel_len = wav_lengths.max() // self.wav_token_compression
74
- speech_tokens = speech_tokens[:, :max_mel_len]
75
-
76
  b, device = text.shape[0], text.device
77
  if self.training:
78
  text_mask = torch.rand_like(text.float()) > self.text_mask_percentage
@@ -82,10 +113,11 @@ class VoiceCLIP(nn.Module):
82
  voice_mask = torch.ones_like(speech_tokens.float()).bool()
83
 
84
  text_emb = self.text_emb(text)
85
- text_emb += self.text_pos_emb(torch.arange(text.shape[1], device=device))
86
-
87
  speech_emb = self.speech_emb(speech_tokens)
88
- speech_emb += self.speech_pos_emb(torch.arange(speech_emb.shape[1], device=device))
 
 
 
89
 
90
  enc_text = self.text_transformer(text_emb, mask=text_mask)
91
  enc_speech = self.speech_transformer(speech_emb, mask=voice_mask)
 
2
  import torch.nn as nn
3
  import torch.nn.functional as F
4
  from torch import einsum
5
+ from x_transformers import Encoder
6
+
7
+ from models.arch_util import CheckpointedXTransformerEncoder
8
  from models.transformer import Transformer
9
 
10
 
 
16
  t = t.masked_fill(~mask[:, :, None], 0.)
17
  return t.sum(dim = 1) / mask.sum(dim = 1)[..., None]
18
 
 
19
  class VoiceCLIP(nn.Module):
20
  """
21
  CLIP model retrofitted for performing contrastive evaluation between tokenized audio data and the corresponding
 
41
  text_mask_percentage=0,
42
  voice_mask_percentage=0,
43
  wav_token_compression=1024,
44
+ use_xformers=False,
45
  ):
46
  super().__init__()
47
  self.text_emb = nn.Embedding(num_text_tokens, dim_text)
 
 
 
48
  self.to_text_latent = nn.Linear(dim_text, dim_latent, bias=False)
49
 
50
  self.speech_emb = nn.Embedding(num_speech_tokens, dim_speech)
 
 
 
51
  self.to_speech_latent = nn.Linear(dim_speech, dim_latent, bias=False)
52
 
53
+ if use_xformers:
54
+ self.text_transformer = CheckpointedXTransformerEncoder(
55
+ needs_permute=False,
56
+ exit_permute=False,
57
+ max_seq_len=-1,
58
+ use_pos_emb=False,
59
+ attn_layers=Encoder(
60
+ dim=dim_text,
61
+ depth=text_enc_depth,
62
+ heads=text_heads,
63
+ ff_dropout=.1,
64
+ ff_mult=2,
65
+ attn_dropout=.1,
66
+ use_rmsnorm=True,
67
+ ff_glu=True,
68
+ rotary_pos_emb=True,
69
+ ))
70
+ self.speech_transformer = CheckpointedXTransformerEncoder(
71
+ needs_permute=False,
72
+ exit_permute=False,
73
+ max_seq_len=-1,
74
+ use_pos_emb=False,
75
+ attn_layers=Encoder(
76
+ dim=dim_speech,
77
+ depth=speech_enc_depth,
78
+ heads=speech_heads,
79
+ ff_dropout=.1,
80
+ ff_mult=2,
81
+ attn_dropout=.1,
82
+ use_rmsnorm=True,
83
+ ff_glu=True,
84
+ rotary_pos_emb=True,
85
+ ))
86
+ else:
87
+ self.text_transformer = Transformer(causal=False, seq_len=text_seq_len, dim=dim_text, depth=text_enc_depth,
88
+ heads=text_heads)
89
+ self.speech_transformer = Transformer(causal=False, seq_len=speech_seq_len, dim=dim_speech,
90
+ depth=speech_enc_depth, heads=speech_heads)
91
+
92
  self.temperature = nn.Parameter(torch.tensor(1.))
93
  self.text_mask_percentage = text_mask_percentage
94
  self.voice_mask_percentage = voice_mask_percentage
95
  self.wav_token_compression = wav_token_compression
96
+ self.xformers = use_xformers
97
+ if not use_xformers:
98
+ self.text_pos_emb = nn.Embedding(text_seq_len, dim_text)
99
+ self.speech_pos_emb = nn.Embedding(num_speech_tokens, dim_speech)
100
 
101
  def forward(
102
  self,
103
  text,
 
104
  speech_tokens,
 
105
  return_loss=False
106
  ):
 
 
 
 
 
 
 
107
  b, device = text.shape[0], text.device
108
  if self.training:
109
  text_mask = torch.rand_like(text.float()) > self.text_mask_percentage
 
113
  voice_mask = torch.ones_like(speech_tokens.float()).bool()
114
 
115
  text_emb = self.text_emb(text)
 
 
116
  speech_emb = self.speech_emb(speech_tokens)
117
+
118
+ if not self.xformers:
119
+ text_emb += self.text_pos_emb(torch.arange(text.shape[1], device=device))
120
+ speech_emb += self.speech_pos_emb(torch.arange(speech_emb.shape[1], device=device))
121
 
122
  enc_text = self.text_transformer(text_emb, mask=text_mask)
123
  enc_speech = self.speech_transformer(speech_emb, mask=voice_mask)