forked from dgaddy/silent_speech
-
Notifications
You must be signed in to change notification settings - Fork 0
/
transduction_model.py
executable file
·480 lines (375 loc) · 17.8 KB
/
transduction_model.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
# Additional imports
import neptune.new as neptune
import warnings
import random
import matplotlib.pyplot as plt
from matplotlib import cm
# NOTE: This ignores the librosa logs
warnings.simplefilter(action = "ignore", category = FutureWarning)
import os
import sys
import numpy as np
import logging
import subprocess
import torch
from torch import nn
import torch.nn.functional as F
from torch.cuda.amp.grad_scaler import GradScaler
from read_emg import EMGDataset, SizeAwareSampler
# from wavenet_model import WavenetModel, save_output as save_wavenet_output
from align import align_from_distances
from asr import evaluate
from transformer import TransformerEncoderLayer
from data_utils import phoneme_inventory, decollate_tensor
from absl import flags
from absl import app
FLAGS = flags.FLAGS
flags.DEFINE_integer('model_size', 768, 'number of hidden dimensions')
flags.DEFINE_integer('num_layers', 6, 'number of layers')
flags.DEFINE_integer('batch_size', 32, 'training batch size')
flags.DEFINE_float('learning_rate', 1e-3, 'learning rate')
flags.DEFINE_integer('learning_rate_patience', 5, 'learning rate decay patience')
flags.DEFINE_integer('learning_rate_warmup', 500, 'steps of linear warmup')
flags.DEFINE_string('start_training_from', None, 'start training from this model')
flags.DEFINE_float('data_size_fraction', 1.0, 'fraction of training data to use')
flags.DEFINE_boolean('no_session_embed', False, "don't use a session embedding")
flags.DEFINE_float('phoneme_loss_weight', 0.1, 'weight of auxiliary phoneme prediction loss')
flags.DEFINE_float('l2', 1e-7, 'weight decay')
# NOTE: FLAGS from non-imported files
flags.DEFINE_string('pretrained_wavenet_model', None, '')
flags.DEFINE_string('output_directory', None, '')
flags.DEFINE_bool('debug', False, '')
flags.DEFINE_bool('amp', False, '')
flags.DEFINE_string("neptune_token", None, "(Optional) Neptune.ai logging token")
flags.DEFINE_string("neptune_project", None, "(Optional) Neptune.ai project name")
flags.DEFINE_integer("n_epochs", 80, "")
flags.DEFINE_integer("random_seed", 1, "")
flags.DEFINE_float('recon_loss_weight', 0.1, 'weight of Ev reconstruction prediction loss')
flags.DEFINE_integer('transformer_nheads', 8, '')
"""START - ADDITIONAL UTILITY FUNCTIONS"""
def plot_mel_spectrograms(pred, y, epoch_idx=None):
fig, ax = plt.subplots(2)
epoch_txt = "" if epoch_idx==None else f", Epoch: {epoch_idx}"
ax[0].set_title(f"Mel Spectogram (Predicted){epoch_txt}")
pred = np.swapaxes(pred, 0, 1)
cax = ax[0].imshow(pred, interpolation='nearest', cmap=cm.coolwarm, origin='lower')
ax[1].set_title(f"Mel Spectogram (Actual){epoch_txt}")
y = np.swapaxes(y, 0, 1)
cax = ax[1].imshow(y, interpolation='nearest', cmap=cm.coolwarm, origin='lower')
return fig
def stack_mel_spectrogram(data):
data = data.cpu().float().detach().numpy()
# Loop over each second of `audio_features`
new_data = data[0]
for i in range(1, data.shape[0]):
new_data = np.vstack((new_data, data[i]))
return new_data
"""END - ADDITIONAL UTILITY FUNCTIONS"""
class ResBlock(nn.Module):
def __init__(self, num_ins, num_outs, stride=1):
super().__init__()
self.conv1 = nn.Conv1d(num_ins, num_outs, 3, padding=1, stride=stride)
self.bn1 = nn.BatchNorm1d(num_outs)
self.conv2 = nn.Conv1d(num_outs, num_outs, 3, padding=1)
self.bn2 = nn.BatchNorm1d(num_outs)
if stride != 1 or num_ins != num_outs:
self.residual_path = nn.Conv1d(num_ins, num_outs, 1, stride=stride)
self.res_norm = nn.BatchNorm1d(num_outs)
else:
self.residual_path = None
def forward(self, x):
input_value = x
x = F.relu(self.bn1(self.conv1(x)))
x = self.bn2(self.conv2(x))
if self.residual_path is not None:
res = self.res_norm(self.residual_path(input_value))
else:
res = input_value
return F.relu(x + res)
class Model(nn.Module):
def __init__(self, num_ins, num_outs, num_aux_outs, num_recon_outs, num_sessions, reconstruction_loss=False):
super().__init__()
self.conv_blocks = nn.Sequential(
ResBlock(8, FLAGS.model_size, 2),
ResBlock(FLAGS.model_size, FLAGS.model_size, 2),
ResBlock(FLAGS.model_size, FLAGS.model_size, 2),
)
self.w_raw_in = nn.Linear(FLAGS.model_size, FLAGS.model_size)
if not FLAGS.no_session_embed:
emb_size = 32
self.session_emb = nn.Embedding(num_sessions, emb_size)
self.w_emb = nn.Linear(emb_size, FLAGS.model_size)
encoder_layer = TransformerEncoderLayer(
d_model=FLAGS.model_size,
nhead=FLAGS.transformer_nheads,
relative_positional=True,
relative_positional_distance=100,
dim_feedforward=3072)
self.transformer = nn.TransformerEncoder(encoder_layer, FLAGS.num_layers)
self.w_out = nn.Linear(FLAGS.model_size, num_outs)
self.w_aux = nn.Linear(FLAGS.model_size, num_aux_outs)
self.w_recon = nn.Linear(FLAGS.model_size, num_recon_outs)
self.reconstruction_loss = reconstruction_loss
def forward(self, x_feat, x_raw, session_ids):
# x shape is (batch, time, electrode)
x_raw = x_raw.transpose(1,2) # put channel before time for conv
x_raw = self.conv_blocks(x_raw)
x_raw = x_raw.transpose(1,2)
x_raw = self.w_raw_in(x_raw)
if FLAGS.no_session_embed:
x = x_raw
else:
emb = self.session_emb(session_ids)
x = x_raw + self.w_emb(emb)
x = x.transpose(0,1) # put time first
x = self.transformer(x)
x = x.transpose(0,1)
return self.w_out(x), self.w_aux(x), self.w_recon(x)
def test(model, testset, device, epoch, run):
model.eval()
dataloader = torch.utils.data.DataLoader(testset, batch_size=32, collate_fn=testset.collate_fixed_length)
losses = []
accuracies = []
phoneme_confusion = np.zeros((len(phoneme_inventory),len(phoneme_inventory)))
recon_losses = []
logged = False
with torch.no_grad():
for example in dataloader:
X = example['emg'].to(device)
X_raw = example['raw_emg'].to(device)
sess = example['session_ids'].to(device)
with torch.autocast(
enabled=FLAGS.amp,
dtype=torch.bfloat16,
device_type="cuda"):
pred, phoneme_pred, X_recon = model(X, X_raw, sess)
if not logged:
audio_features = example['audio_features']
plot_pred = decollate_tensor(pred, example['lengths'])
plot_y = decollate_tensor(audio_features, example['audio_feature_lengths'])
print("plot_pred SHAPE:", plot_pred[0].shape)
# Log predicted mel_spectrogram
fig = plot_mel_spectrograms(
stack_mel_spectrogram(plot_pred[0]),
stack_mel_spectrogram(plot_y[0]),
epoch_idx=epoch)
fname = f"cur_epoch_{epoch}.png"
fig.savefig(fname)
if run:
run["model/visualisation"].upload(fname)
run[f"model/visualisation_epoch_{epoch}"].upload(fname)
logged = True
loss, phon_acc, recon_loss = dtw_loss(pred, phoneme_pred, X_recon, example, True, phoneme_confusion)
losses.append(loss.item())
recon_losses.append(recon_loss.item())
accuracies.append(phon_acc)
model.train()
return np.mean(losses), np.mean(accuracies), phoneme_confusion, np.mean(recon_losses) #TODO size-weight average
"""
def save_output(model, datapoint, filename, device, gold_mfcc=False):
model.eval()
if gold_mfcc:
y = datapoint['audio_features']
else:
with torch.no_grad():
sess = torch.tensor(datapoint['session_ids'], device=device).unsqueeze(0)
X = torch.tensor(datapoint['emg'], dtype=torch.float32, device=device).unsqueeze(0)
X_raw = torch.tensor(datapoint['raw_emg'], dtype=torch.float32, device=device).unsqueeze(0)
pred, _ = model(X, X_raw, sess)
pred = pred.squeeze(0)
y = pred.cpu().detach().numpy()
wavenet_model = WavenetModel(y.shape[1]).to(device)
assert FLAGS.pretrained_wavenet_model is not None
wavenet_model.load_state_dict(torch.load(FLAGS.pretrained_wavenet_model))
save_wavenet_output(wavenet_model, y, filename, device)
model.train()
"""
def dtw_loss(predictions, phoneme_predictions, X_recon, example, phoneme_eval=False, phoneme_confusion=None):
# device = predictions.device()
device = predictions.device
inputs = decollate_tensor(example['raw_emg'], example['lengths'])
predictions = decollate_tensor(predictions, example['lengths'])
phoneme_predictions = decollate_tensor(phoneme_predictions, example['lengths'])
recon_predictions = decollate_tensor(X_recon, example['lengths'])
audio_features = example['audio_features'].to(device)
phoneme_targets = example['phonemes']
audio_features = decollate_tensor(audio_features, example['audio_feature_lengths'])
losses = []
recon_losses = []
correct_phones = 0
total_length = 0
for pred, y, pred_phone, y_phone, recon, y_recon, silent in zip(predictions, audio_features, phoneme_predictions, phoneme_targets, recon_predictions, inputs, example['silent']):
assert len(pred.size()) == 2 and len(y.size()) == 2
y_phone = y_phone.to(device)
if silent:
dists = torch.cdist(pred.unsqueeze(0), y.unsqueeze(0))
costs = dists.squeeze(0)
# pred_phone (seq1_len, 48), y_phone (seq2_len)
# phone_probs (seq1_len, seq2_len)
pred_phone = F.log_softmax(pred_phone, -1)
phone_lprobs = pred_phone[:,y_phone]
recon_loss = F.mse_loss(recon.to(device), y_recon.to(device))
costs = costs + \
FLAGS.phoneme_loss_weight * -phone_lprobs + \
FLAGS.recon_loss_weight * recon_loss
alignment = align_from_distances(costs.T.cpu().detach().numpy())
loss = costs[alignment,range(len(alignment))].sum()
if phoneme_eval:
alignment = align_from_distances(costs.T.cpu().detach().numpy())
pred_phone = pred_phone.argmax(-1)
correct_phones += (pred_phone[alignment] == y_phone).sum().item()
for p, t in zip(pred_phone[alignment].tolist(), y_phone.tolist()):
phoneme_confusion[p, t] += 1
else:
assert y.size(0) == pred.size(0)
dists = F.pairwise_distance(y, pred)
assert len(pred_phone.size()) == 2 and len(y_phone.size()) == 1
phoneme_loss = F.cross_entropy(pred_phone, y_phone, reduction='sum')
recon_loss = F.mse_loss(recon.to(device), y_recon.to(device))
loss = dists.cpu().sum() + \
FLAGS.phoneme_loss_weight * phoneme_loss.cpu() + \
FLAGS.recon_loss_weight * recon_loss.cpu()
if phoneme_eval:
pred_phone = pred_phone.argmax(-1)
correct_phones += (pred_phone == y_phone).sum().item()
for p, t in zip(pred_phone.tolist(), y_phone.tolist()):
phoneme_confusion[p, t] += 1
losses.append(loss)
recon_losses.append(recon_loss)
total_length += y.size(0)
return sum(losses)/total_length, correct_phones/total_length, sum(recon_losses)/total_length
def train_model(trainset, devset, device, save_sound_outputs=True, n_epochs=80, run=None):
if FLAGS.data_size_fraction >= 1:
training_subset = trainset
else:
training_subset = \
torch.utils.data.Subset(
trainset,
list(range(int(len(trainset)*FLAGS.data_size_fraction))))
dataloader = torch.utils.data.DataLoader(
training_subset, pin_memory=(device=='cuda'),
collate_fn=devset.collate_fixed_length,
num_workers=8,
batch_sampler=SizeAwareSampler(trainset, 256000))
n_phones = len(phoneme_inventory)
model = Model(num_ins=devset.num_features,
num_outs=devset.num_speech_features,
num_aux_outs=n_phones,
num_recon_outs=8,
num_sessions=devset.num_sessions,
reconstruction_loss=(True if FLAGS.recon_loss_weight > 0.0 else False)).to(device)
if FLAGS.start_training_from is not None:
state_dict = torch.load(FLAGS.start_training_from)
del state_dict['session_emb.weight']
model.load_state_dict(state_dict, strict=False)
optim = torch.optim.AdamW(model.parameters(), weight_decay=FLAGS.l2)
lr_sched = torch.optim.lr_scheduler.ReduceLROnPlateau(optim, 'min', 0.5, patience=FLAGS.learning_rate_patience)
def set_lr(new_lr):
for param_group in optim.param_groups:
param_group['lr'] = new_lr
target_lr = FLAGS.learning_rate
def schedule_lr(iteration):
iteration = iteration + 1
if iteration <= FLAGS.learning_rate_warmup:
set_lr(iteration*target_lr/FLAGS.learning_rate_warmup)
scaler = GradScaler()
batch_idx = 0
for epoch_idx in range(n_epochs):
losses = []
recon_losses = []
for example in dataloader:
optim.zero_grad()
schedule_lr(batch_idx)
X = example['emg'].to(device)
X_raw = example['raw_emg'].to(device)
print("X_RAW SHAPE:", X_raw.shape)
sess = example['session_ids'].to(device)
with torch.autocast(
enabled=FLAGS.amp,
dtype=torch.bfloat16,
device_type="cuda"):
pred, phoneme_pred, X_recon = model(X, X_raw, sess)
loss, _, recon_loss = dtw_loss(pred, phoneme_pred, X_recon, example)
losses.append(loss.item())
recon_losses.append(recon_loss.item())
print("LOSS:", loss)
loss = loss.to(device)
scaler.scale(loss).backward()
scaler.step(optim)
scaler.update()
#loss.backward()
#optim.step()
batch_idx += 1
train_loss = np.mean(losses)
recon_loss = np.mean(recon_losses)
val, phoneme_acc, _, recon_loss = test(model, devset, device, epoch_idx, run)
lr_sched.step(val)
if run:
run["val_loss"].log(val)
run["train_loss"].log(train_loss)
run["recon_loss"].log(recon_loss)
run["phoneme_acc"].log(phoneme_acc*100)
logging.info(f'finished epoch {epoch_idx+1} - validation loss: {val:.4f} training loss: {train_loss:.4f} recon loss: {recon_loss:.4f} phoneme accuracy: {phoneme_acc*100:.2f}')
torch.save(model.state_dict(), os.path.join(FLAGS.output_directory, 'full_model.pt'))
"""
if save_sound_outputs:
save_output(model, devset[0], os.path.join(FLAGS.output_directory, f'epoch_{epoch_idx}_output.wav'), device)
"""
model.load_state_dict(torch.load(os.path.join(FLAGS.output_directory,'model.pt'))) # re-load best parameters
"""
if save_sound_outputs:
for i, datapoint in enumerate(devset):
save_output(model, datapoint, os.path.join(FLAGS.output_directory, f'example_output_{i}.wav'), device)
"""
evaluate(devset, FLAGS.output_directory)
return model
def main(unused_argv):
if FLAGS.neptune_project and FLAGS.neptune_token:
run = neptune.init(project=FLAGS.neptune_project,
api_token=FLAGS.neptune_token)
else:
run = None
if run:
run["hparams"] = {
"model_size": FLAGS.model_size,
"num_layers": FLAGS.num_layers,
"batch_size": FLAGS.batch_size,
"learning_rate": FLAGS.learning_rate,
"phoneme_loss_weight": FLAGS.phoneme_loss_weight,
"recon_loss_weight": FLAGS.recon_loss_weight,
"l2": FLAGS.l2,
"data_size_fraction": FLAGS.data_size_fraction,
"learning_rate_patience": FLAGS.learning_rate_patience,
"learning_rate_warmup": FLAGS.learning_rate_warmup,
"n_epochs": FLAGS.n_epochs,
"random_seed": FLAGS.random_seed,
"mel_spectrogram": FLAGS.mel_spectrogram,
"transformer_nheads": FLAGS.transformer_nheads,
"normalizers_file": FLAGS.normalizers_file,
}
random.seed(FLAGS.random_seed)
torch.manual_seed(FLAGS.random_seed)
np.random.seed(FLAGS.random_seed)
os.makedirs(FLAGS.output_directory, exist_ok=True)
logging.basicConfig(handlers=[
logging.FileHandler(os.path.join(FLAGS.output_directory, 'log.txt'), 'w'),
logging.StreamHandler()
], level=logging.INFO, format="%(message)s")
logging.info(subprocess.run(['git','rev-parse','HEAD'], stdout=subprocess.PIPE, universal_newlines=True).stdout)
logging.info(subprocess.run(['git','diff'], stdout=subprocess.PIPE, universal_newlines=True).stdout)
logging.info(sys.argv)
trainset = EMGDataset(dev=False,test=False)
devset = EMGDataset(dev=True)
logging.info('output example: %s', devset.example_indices[0])
logging.info('train / dev split: %d %d',len(trainset),len(devset))
device = 'cuda' if torch.cuda.is_available() and not FLAGS.debug else 'cpu'
model = train_model(trainset,
devset,
device,
save_sound_outputs=(FLAGS.pretrained_wavenet_model is not None),
n_epochs=FLAGS.n_epochs,
run=run)
if run:
run.stop()
if __name__ == '__main__':
app.run(main)