File size: 30,185 Bytes
0374441
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
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
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
# load packages
import random
import yaml
import time
from munch import Munch
import numpy as np
import torch
from torch import nn
import torch.nn.functional as F
import torchaudio
import librosa
import click
import shutil
import warnings
warnings.simplefilter('ignore')
from torch.utils.tensorboard import SummaryWriter

from meldataset import build_dataloader

from Utils.ASR.models import ASRCNN
from Utils.JDC.model import JDCNet
from Utils.PLBERT.util import load_plbert

from models import *
from losses import *
from utils import *

from Modules.slmadv import SLMAdversarialLoss
from Modules.diffusion.sampler import DiffusionSampler, ADPM2Sampler, KarrasSchedule

from optimizers import build_optimizer

from accelerate import Accelerator

accelerator = Accelerator()

# simple fix for dataparallel that allows access to class attributes
class MyDataParallel(torch.nn.DataParallel):
    def __getattr__(self, name):
        try:
            return super().__getattr__(name)
        except AttributeError:
            return getattr(self.module, name)
        
import logging
from logging import StreamHandler
logger = logging.getLogger(__name__)
logger.setLevel(logging.DEBUG)
handler = StreamHandler()
handler.setLevel(logging.DEBUG)
logger.addHandler(handler)


@click.command()
@click.option('-p', '--config_path', default='Configs/config_ft.yml', type=str)
def main(config_path):
    config = yaml.safe_load(open(config_path))
    
    log_dir = config['log_dir']
    if not osp.exists(log_dir): os.makedirs(log_dir, exist_ok=True)
    shutil.copy(config_path, osp.join(log_dir, osp.basename(config_path)))
    writer = SummaryWriter(log_dir + "/tensorboard")

    # write logs
    file_handler = logging.FileHandler(osp.join(log_dir, 'train.log'))
    file_handler.setLevel(logging.DEBUG)
    file_handler.setFormatter(logging.Formatter('%(levelname)s:%(asctime)s: %(message)s'))
    logger.addHandler(file_handler)

    
    batch_size = config.get('batch_size', 10)

    epochs = config.get('epochs', 200)
    save_freq = config.get('save_freq', 2)
    log_interval = config.get('log_interval', 10)
    saving_epoch = config.get('save_freq', 2)

    data_params = config.get('data_params', None)
    sr = config['preprocess_params'].get('sr', 24000)
    train_path = data_params['train_data']
    val_path = data_params['val_data']
    root_path = data_params['root_path']
    min_length = data_params['min_length']
    OOD_data = data_params['OOD_data']

    max_len = config.get('max_len', 200)
    
    loss_params = Munch(config['loss_params'])
    diff_epoch = loss_params.diff_epoch
    joint_epoch = loss_params.joint_epoch
    
    optimizer_params = Munch(config['optimizer_params'])
    
    train_list, val_list = get_data_path_list(train_path, val_path)
    device = accelerator.device

    train_dataloader = build_dataloader(train_list,
                                        root_path,
                                        OOD_data=OOD_data,
                                        min_length=min_length,
                                        batch_size=batch_size,
                                        num_workers=2,
                                        dataset_config={},
                                        device=device)

    val_dataloader = build_dataloader(val_list,
                                      root_path,
                                      OOD_data=OOD_data,
                                      min_length=min_length,
                                      batch_size=batch_size,
                                      validation=True,
                                      num_workers=0,
                                      device=device,
                                      dataset_config={})
    
    # load pretrained ASR model
    ASR_config = config.get('ASR_config', False)
    ASR_path = config.get('ASR_path', False)
    text_aligner = load_ASR_models(ASR_path, ASR_config)
    
    # load pretrained F0 model
    F0_path = config.get('F0_path', False)
    pitch_extractor = load_F0_models(F0_path)
    
    # load PL-BERT model
    BERT_path = config.get('PLBERT_dir', False)
    plbert = load_plbert(BERT_path)
    
    # build model
    model_params = recursive_munch(config['model_params'])
    multispeaker = model_params.multispeaker
    model = build_model(model_params, text_aligner, pitch_extractor, plbert)
    _ = [model[key].to(device) for key in model]
    
    # DP
    for key in model:
        if key != "mpd" and key != "msd" and key != "wd":
            model[key] = MyDataParallel(model[key])
            
    start_epoch = 0
    iters = 0

    load_pretrained = config.get('pretrained_model', '') != '' and config.get('second_stage_load_pretrained', False)
    
    if not load_pretrained:
        if config.get('first_stage_path', '') != '':
            first_stage_path = osp.join(log_dir, config.get('first_stage_path', 'first_stage.pth'))
            print('Loading the first stage model at %s ...' % first_stage_path)
            model, _, start_epoch, iters = load_checkpoint(model, 
                None, 
                first_stage_path,
                load_only_params=True,
                ignore_modules=['bert', 'bert_encoder', 'predictor', 'predictor_encoder', 'msd', 'mpd', 'wd', 'diffusion']) # keep starting epoch for tensorboard log

            # these epochs should be counted from the start epoch
            diff_epoch += start_epoch
            joint_epoch += start_epoch
            epochs += start_epoch
            
            model.predictor_encoder = copy.deepcopy(model.style_encoder)
        else:
            raise ValueError('You need to specify the path to the first stage model.') 

    gl = GeneratorLoss(model.mpd, model.msd).to(device)
    dl = DiscriminatorLoss(model.mpd, model.msd).to(device)
    wl = WavLMLoss(model_params.slm.model, 
                   model.wd, 
                   sr, 
                   model_params.slm.sr).to(device)

    gl = MyDataParallel(gl)
    dl = MyDataParallel(dl)
    wl = MyDataParallel(wl)
    
    sampler = DiffusionSampler(
        model.diffusion.diffusion,
        sampler=ADPM2Sampler(),
        sigma_schedule=KarrasSchedule(sigma_min=0.0001, sigma_max=3.0, rho=9.0), # empirical parameters
        clamp=False
    )
    
    scheduler_params = {
        "max_lr": optimizer_params.lr,
        "pct_start": float(0),
        "epochs": epochs,
        "steps_per_epoch": len(train_dataloader),
    }
    scheduler_params_dict= {key: scheduler_params.copy() for key in model}
    scheduler_params_dict['bert']['max_lr'] = optimizer_params.bert_lr * 2
    scheduler_params_dict['decoder']['max_lr'] = optimizer_params.ft_lr * 2
    scheduler_params_dict['style_encoder']['max_lr'] = optimizer_params.ft_lr * 2
    
    optimizer = build_optimizer({key: model[key].parameters() for key in model},
                                          scheduler_params_dict=scheduler_params_dict, lr=optimizer_params.lr)
    
    # adjust BERT learning rate
    for g in optimizer.optimizers['bert'].param_groups:
        g['betas'] = (0.9, 0.99)
        g['lr'] = optimizer_params.bert_lr
        g['initial_lr'] = optimizer_params.bert_lr
        g['min_lr'] = 0
        g['weight_decay'] = 0.01
        
    # adjust acoustic module learning rate
    for module in ["decoder", "style_encoder"]:
        for g in optimizer.optimizers[module].param_groups:
            g['betas'] = (0.0, 0.99)
            g['lr'] = optimizer_params.ft_lr
            g['initial_lr'] = optimizer_params.ft_lr
            g['min_lr'] = 0
            g['weight_decay'] = 1e-4
        
    # load models if there is a model
    if load_pretrained:
        model, optimizer, start_epoch, iters = load_checkpoint(model,  optimizer, config['pretrained_model'],
                                    load_only_params=config.get('load_only_params', True))
        
    n_down = model.text_aligner.n_down

    best_loss = float('inf')  # best test loss
    loss_train_record = list([])
    loss_test_record = list([])
    iters = 0
    
    criterion = nn.L1Loss() # F0 loss (regression)
    torch.cuda.empty_cache()
    
    stft_loss = MultiResolutionSTFTLoss().to(device)
    
    print('BERT', optimizer.optimizers['bert'])
    print('decoder', optimizer.optimizers['decoder'])

    start_ds = False
    
    running_std = []
    
    slmadv_params = Munch(config['slmadv_params'])
    slmadv = SLMAdversarialLoss(model, wl, sampler, 
                                slmadv_params.min_len, 
                                slmadv_params.max_len,
                                batch_percentage=slmadv_params.batch_percentage,
                                skip_update=slmadv_params.iter, 
                                sig=slmadv_params.sig
                               )

    model, optimizer, train_dataloader = accelerator.prepare(
        model, optimizer, train_dataloader
    )

    for epoch in range(start_epoch, epochs):
        running_loss = 0
        start_time = time.time()

        _ = [model[key].eval() for key in model]
        
        model.text_aligner.train()
        model.text_encoder.train()
        
        model.predictor.train()
        model.bert_encoder.train()
        model.bert.train()
        model.msd.train()
        model.mpd.train()

        for i, batch in enumerate(train_dataloader):
            waves = batch[0]
            batch = [b.to(device) for b in batch[1:]]
            texts, input_lengths, ref_texts, ref_lengths, mels, mel_input_length, ref_mels = batch
            with torch.no_grad():
                mask = length_to_mask(mel_input_length // (2 ** n_down)).to(device)
                mel_mask = length_to_mask(mel_input_length).to(device)
                text_mask = length_to_mask(input_lengths).to(texts.device)

                # compute reference styles
                if multispeaker and epoch >= diff_epoch:
                    ref_ss = model.style_encoder(ref_mels.unsqueeze(1))
                    ref_sp = model.predictor_encoder(ref_mels.unsqueeze(1))
                    ref = torch.cat([ref_ss, ref_sp], dim=1)
                
            try:
                ppgs, s2s_pred, s2s_attn = model.text_aligner(mels, mask, texts)
                s2s_attn = s2s_attn.transpose(-1, -2)
                s2s_attn = s2s_attn[..., 1:]
                s2s_attn = s2s_attn.transpose(-1, -2)
            except:
                continue

            mask_ST = mask_from_lens(s2s_attn, input_lengths, mel_input_length // (2 ** n_down))
            s2s_attn_mono = maximum_path(s2s_attn, mask_ST)

            # encode
            t_en = model.text_encoder(texts, input_lengths, text_mask)
            
            # 50% of chance of using monotonic version
            if bool(random.getrandbits(1)):
                asr = (t_en @ s2s_attn)
            else:
                asr = (t_en @ s2s_attn_mono)

            d_gt = s2s_attn_mono.sum(axis=-1).detach()

            # compute the style of the entire utterance
            # this operation cannot be done in batch because of the avgpool layer (may need to work on masked avgpool)
            ss = []
            gs = []
            for bib in range(len(mel_input_length)):
                mel_length = int(mel_input_length[bib].item())
                mel = mels[bib, :, :mel_input_length[bib]]
                s = model.predictor_encoder(mel.unsqueeze(0).unsqueeze(1))
                ss.append(s)
                s = model.style_encoder(mel.unsqueeze(0).unsqueeze(1))
                gs.append(s)

            s_dur = torch.stack(ss).squeeze()  # global prosodic styles
            gs = torch.stack(gs).squeeze() # global acoustic styles
            s_trg = torch.cat([gs, s_dur], dim=-1).detach() # ground truth for denoiser

            bert_dur = model.bert(texts, attention_mask=(~text_mask).int())
            d_en = model.bert_encoder(bert_dur).transpose(-1, -2) 
            
            # denoiser training
            if epoch >= diff_epoch:
                num_steps = np.random.randint(3, 5)
                
                if model_params.diffusion.dist.estimate_sigma_data:
                    model.diffusion.module.diffusion.sigma_data = s_trg.std(axis=-1).mean().item() # batch-wise std estimation
                    running_std.append(model.diffusion.module.diffusion.sigma_data)
                    
                if multispeaker:
                    s_preds = sampler(noise = torch.randn_like(s_trg).unsqueeze(1).to(device), 
                          embedding=bert_dur,
                          embedding_scale=1,
                                   features=ref, # reference from the same speaker as the embedding
                             embedding_mask_proba=0.1,
                             num_steps=num_steps).squeeze(1)
                    loss_diff = model.diffusion(s_trg.unsqueeze(1), embedding=bert_dur, features=ref).mean() # EDM loss
                    loss_sty = F.l1_loss(s_preds, s_trg.detach()) # style reconstruction loss
                else:
                    s_preds = sampler(noise = torch.randn_like(s_trg).unsqueeze(1).to(device), 
                          embedding=bert_dur,
                          embedding_scale=1,
                             embedding_mask_proba=0.1,
                             num_steps=num_steps).squeeze(1)                    
                    loss_diff = model.diffusion.module.diffusion(s_trg.unsqueeze(1), embedding=bert_dur).mean() # EDM loss
                    loss_sty = F.l1_loss(s_preds, s_trg.detach()) # style reconstruction loss
            else:
                loss_sty = 0
                loss_diff = 0

                
            s_loss = 0
            

            d, p = model.predictor(d_en, s_dur, 
                                                    input_lengths, 
                                                    s2s_attn_mono, 
                                                    text_mask)
                
            mel_len_st = int(mel_input_length.min().item() / 2 - 1)
            mel_len = min(int(mel_input_length.min().item() / 2 - 1), max_len // 2)
            en = []
            gt = []
            p_en = []
            wav = []
            st = []
            
            for bib in range(len(mel_input_length)):
                mel_length = int(mel_input_length[bib].item() / 2)

                random_start = np.random.randint(0, mel_length - mel_len)
                en.append(asr[bib, :, random_start:random_start+mel_len])
                p_en.append(p[bib, :, random_start:random_start+mel_len])
                gt.append(mels[bib, :, (random_start * 2):((random_start+mel_len) * 2)])
                
                y = waves[bib][(random_start * 2) * 300:((random_start+mel_len) * 2) * 300]
                wav.append(torch.from_numpy(y).to(device))
                
                # style reference (better to be different from the GT)
                random_start = np.random.randint(0, mel_length - mel_len_st)
                st.append(mels[bib, :, (random_start * 2):((random_start+mel_len_st) * 2)])
                
            wav = torch.stack(wav).float().detach()

            en = torch.stack(en)
            p_en = torch.stack(p_en)
            gt = torch.stack(gt).detach()
            st = torch.stack(st).detach()
            
            
            if gt.size(-1) < 80:
                continue
            
            s = model.style_encoder(gt.unsqueeze(1))           
            s_dur = model.predictor_encoder(gt.unsqueeze(1))
                
            with torch.no_grad():
                F0_real, _, F0 = model.pitch_extractor(gt.unsqueeze(1))
                F0 = F0.reshape(F0.shape[0], F0.shape[1] * 2, F0.shape[2], 1).squeeze()

                N_real = log_norm(gt.unsqueeze(1)).squeeze(1)
                
                y_rec_gt = wav.unsqueeze(1)
                y_rec_gt_pred = model.decoder(en, F0_real, N_real, s)

                wav = y_rec_gt

            F0_fake, N_fake = model.predictor.F0Ntrain(p_en, s_dur)

            y_rec = model.decoder(en, F0_fake, N_fake, s)

            loss_F0_rec =  (F.smooth_l1_loss(F0_real, F0_fake)) / 10
            loss_norm_rec = F.smooth_l1_loss(N_real, N_fake)

            optimizer.zero_grad()
            d_loss = dl(wav.detach(), y_rec.detach()).mean()
            accelerator.backward(d_loss)
            optimizer.step('msd')
            optimizer.step('mpd')

            # generator loss
            optimizer.zero_grad()

            loss_mel = stft_loss(y_rec, wav)
            loss_gen_all = gl(wav, y_rec).mean()
            loss_lm = wl(wav.detach().squeeze(), y_rec.squeeze()).mean()

            loss_ce = 0
            loss_dur = 0
            for _s2s_pred, _text_input, _text_length in zip(d, (d_gt), input_lengths):
                _s2s_pred = _s2s_pred[:_text_length, :]
                _text_input = _text_input[:_text_length].long()
                _s2s_trg = torch.zeros_like(_s2s_pred)
                for p in range(_s2s_trg.shape[0]):
                    _s2s_trg[p, :_text_input[p]] = 1
                _dur_pred = torch.sigmoid(_s2s_pred).sum(axis=1)

                loss_dur += F.l1_loss(_dur_pred[1:_text_length-1], 
                                       _text_input[1:_text_length-1])
                loss_ce += F.binary_cross_entropy_with_logits(_s2s_pred.flatten(), _s2s_trg.flatten())

            loss_ce /= texts.size(0)
            loss_dur /= texts.size(0)
            
            loss_s2s = 0
            for _s2s_pred, _text_input, _text_length in zip(s2s_pred, texts, input_lengths):
                loss_s2s += F.cross_entropy(_s2s_pred[:_text_length], _text_input[:_text_length])
            loss_s2s /= texts.size(0)

            loss_mono = F.l1_loss(s2s_attn, s2s_attn_mono) * 10

            g_loss = loss_params.lambda_mel * loss_mel + \
                     loss_params.lambda_F0 * loss_F0_rec + \
                     loss_params.lambda_ce * loss_ce + \
                     loss_params.lambda_norm * loss_norm_rec + \
                     loss_params.lambda_dur * loss_dur + \
                     loss_params.lambda_gen * loss_gen_all + \
                     loss_params.lambda_slm * loss_lm + \
                     loss_params.lambda_sty * loss_sty + \
                     loss_params.lambda_diff * loss_diff + \
                    loss_params.lambda_mono * loss_mono + \
                    loss_params.lambda_s2s * loss_s2s
            
            running_loss += loss_mel.item()
            accelerator.backward(g_loss)
            if torch.isnan(g_loss):
                from IPython.core.debugger import set_trace
                set_trace()

            optimizer.step('bert_encoder')
            optimizer.step('bert')
            optimizer.step('predictor')
            optimizer.step('predictor_encoder')
            optimizer.step('style_encoder')
            optimizer.step('decoder')
            
            optimizer.step('text_encoder')
            optimizer.step('text_aligner')
            
            if epoch >= diff_epoch:
                optimizer.step('diffusion')

            d_loss_slm, loss_gen_lm = 0, 0
            if epoch >= joint_epoch:
                # randomly pick whether to use in-distribution text
                if np.random.rand() < 0.5:
                    use_ind = True
                else:
                    use_ind = False

                if use_ind:
                    ref_lengths = input_lengths
                    ref_texts = texts
                    
                slm_out = slmadv(i, 
                                 y_rec_gt, 
                                 y_rec_gt_pred, 
                                 waves, 
                                 mel_input_length,
                                 ref_texts, 
                                 ref_lengths, use_ind, s_trg.detach(), ref if multispeaker else None)

                if slm_out is not None:
                    d_loss_slm, loss_gen_lm, y_pred = slm_out

                    # SLM generator loss
                    optimizer.zero_grad()
                    accelerator.backward(loss_gen_lm)

                    # compute the gradient norm
                    total_norm = {}
                    for key in model.keys():
                        total_norm[key] = 0
                        parameters = [p for p in model[key].parameters() if p.grad is not None and p.requires_grad]
                        for p in parameters:
                            param_norm = p.grad.detach().data.norm(2)
                            total_norm[key] += param_norm.item() ** 2
                        total_norm[key] = total_norm[key] ** 0.5

                    # gradient scaling
                    if total_norm['predictor'] > slmadv_params.thresh:
                        for key in model.keys():
                            for p in model[key].parameters():
                                if p.grad is not None:
                                    p.grad *= (1 / total_norm['predictor'])

                    for p in model.predictor.duration_proj.parameters():
                        if p.grad is not None:
                            p.grad *= slmadv_params.scale

                    for p in model.predictor.lstm.parameters():
                        if p.grad is not None:
                            p.grad *= slmadv_params.scale

                    for p in model.diffusion.parameters():
                        if p.grad is not None:
                            p.grad *= slmadv_params.scale
                    
                    optimizer.step('bert_encoder')
                    optimizer.step('bert')
                    optimizer.step('predictor')
                    optimizer.step('diffusion')

                    # SLM discriminator loss
                    if d_loss_slm != 0:
                        optimizer.zero_grad()
                        accelerator.backward(d_loss_slm)
                        optimizer.step('wd')

            iters = iters + 1
            
            if (i+1)%log_interval == 0:
                logger.info ('Epoch [%d/%d], Step [%d/%d], Loss: %.5f, Disc Loss: %.5f, Dur Loss: %.5f, CE Loss: %.5f, Norm Loss: %.5f, F0 Loss: %.5f, LM Loss: %.5f, Gen Loss: %.5f, Sty Loss: %.5f, Diff Loss: %.5f, DiscLM Loss: %.5f, GenLM Loss: %.5f, SLoss: %.5f, S2S Loss: %.5f, Mono Loss: %.5f'
                    %(epoch+1, epochs, i+1, len(train_list)//batch_size, running_loss / log_interval, d_loss, loss_dur, loss_ce, loss_norm_rec, loss_F0_rec, loss_lm, loss_gen_all, loss_sty, loss_diff, d_loss_slm, loss_gen_lm, s_loss, loss_s2s, loss_mono))
                
                writer.add_scalar('train/mel_loss', running_loss / log_interval, iters)
                writer.add_scalar('train/gen_loss', loss_gen_all, iters)
                writer.add_scalar('train/d_loss', d_loss, iters)
                writer.add_scalar('train/ce_loss', loss_ce, iters)
                writer.add_scalar('train/dur_loss', loss_dur, iters)
                writer.add_scalar('train/slm_loss', loss_lm, iters)
                writer.add_scalar('train/norm_loss', loss_norm_rec, iters)
                writer.add_scalar('train/F0_loss', loss_F0_rec, iters)
                writer.add_scalar('train/sty_loss', loss_sty, iters)
                writer.add_scalar('train/diff_loss', loss_diff, iters)
                writer.add_scalar('train/d_loss_slm', d_loss_slm, iters)
                writer.add_scalar('train/gen_loss_slm', loss_gen_lm, iters)
                
                running_loss = 0
                
                print('Time elasped:', time.time()-start_time)
            
        loss_test = 0
        loss_align = 0
        loss_f = 0
        _ = [model[key].eval() for key in model]

        with torch.no_grad():
            iters_test = 0
            for batch_idx, batch in enumerate(val_dataloader):
                optimizer.zero_grad()

                try:
                    waves = batch[0]
                    batch = [b.to(device) for b in batch[1:]]
                    texts, input_lengths, ref_texts, ref_lengths, mels, mel_input_length, ref_mels = batch
                    with torch.no_grad():
                        mask = length_to_mask(mel_input_length // (2 ** n_down)).to('cuda')
                        text_mask = length_to_mask(input_lengths).to(texts.device)

                        _, _, s2s_attn = model.text_aligner(mels, mask, texts)
                        s2s_attn = s2s_attn.transpose(-1, -2)
                        s2s_attn = s2s_attn[..., 1:]
                        s2s_attn = s2s_attn.transpose(-1, -2)

                        mask_ST = mask_from_lens(s2s_attn, input_lengths, mel_input_length // (2 ** n_down))
                        s2s_attn_mono = maximum_path(s2s_attn, mask_ST)

                        # encode
                        t_en = model.text_encoder(texts, input_lengths, text_mask)
                        asr = (t_en @ s2s_attn_mono)

                        d_gt = s2s_attn_mono.sum(axis=-1).detach()

                    ss = []
                    gs = []

                    for bib in range(len(mel_input_length)):
                        mel_length = int(mel_input_length[bib].item())
                        mel = mels[bib, :, :mel_input_length[bib]]
                        s = model.predictor_encoder(mel.unsqueeze(0).unsqueeze(1))
                        ss.append(s)
                        s = model.style_encoder(mel.unsqueeze(0).unsqueeze(1))
                        gs.append(s)

                    s = torch.stack(ss).squeeze()
                    gs = torch.stack(gs).squeeze()
                    s_trg = torch.cat([s, gs], dim=-1).detach()

                    bert_dur = model.bert(texts, attention_mask=(~text_mask).int())
                    d_en = model.bert_encoder(bert_dur).transpose(-1, -2) 
                    d, p = model.predictor(d_en, s, 
                                                        input_lengths, 
                                                        s2s_attn_mono, 
                                                        text_mask)
                    # get clips
                    mel_len = int(mel_input_length.min().item() / 2 - 1)
                    en = []
                    gt = []

                    p_en = []
                    wav = []

                    for bib in range(len(mel_input_length)):
                        mel_length = int(mel_input_length[bib].item() / 2)

                        random_start = np.random.randint(0, mel_length - mel_len)
                        en.append(asr[bib, :, random_start:random_start+mel_len])
                        p_en.append(p[bib, :, random_start:random_start+mel_len])

                        gt.append(mels[bib, :, (random_start * 2):((random_start+mel_len) * 2)])
                        y = waves[bib][(random_start * 2) * 300:((random_start+mel_len) * 2) * 300]
                        wav.append(torch.from_numpy(y).to(device))

                    wav = torch.stack(wav).float().detach()

                    en = torch.stack(en)
                    p_en = torch.stack(p_en)
                    gt = torch.stack(gt).detach()
                    s = model.predictor_encoder(gt.unsqueeze(1))

                    F0_fake, N_fake = model.predictor.F0Ntrain(p_en, s)

                    loss_dur = 0
                    for _s2s_pred, _text_input, _text_length in zip(d, (d_gt), input_lengths):
                        _s2s_pred = _s2s_pred[:_text_length, :]
                        _text_input = _text_input[:_text_length].long()
                        _s2s_trg = torch.zeros_like(_s2s_pred)
                        for bib in range(_s2s_trg.shape[0]):
                            _s2s_trg[bib, :_text_input[bib]] = 1
                        _dur_pred = torch.sigmoid(_s2s_pred).sum(axis=1)
                        loss_dur += F.l1_loss(_dur_pred[1:_text_length-1], 
                                               _text_input[1:_text_length-1])

                    loss_dur /= texts.size(0)

                    s = model.style_encoder(gt.unsqueeze(1))

                    y_rec = model.decoder(en, F0_fake, N_fake, s)
                    loss_mel = stft_loss(y_rec.squeeze(), wav.detach())

                    F0_real, _, F0 = model.pitch_extractor(gt.unsqueeze(1)) 

                    loss_F0 = F.l1_loss(F0_real, F0_fake) / 10

                    loss_test += (loss_mel).mean()
                    loss_align += (loss_dur).mean()
                    loss_f += (loss_F0).mean()

                    iters_test += 1
                except:
                    continue

        print('Epochs:', epoch + 1)
        logger.info('Validation loss: %.3f, Dur loss: %.3f, F0 loss: %.3f' % (loss_test / iters_test, loss_align / iters_test, loss_f / iters_test) + '\n\n\n')
        print('\n\n\n')
        writer.add_scalar('eval/mel_loss', loss_test / iters_test, epoch + 1)
        writer.add_scalar('eval/dur_loss', loss_test / iters_test, epoch + 1)
        writer.add_scalar('eval/F0_loss', loss_f / iters_test, epoch + 1)
        
        
        if (epoch + 1) % save_freq == 0 :
            if (loss_test / iters_test) < best_loss:
                best_loss = loss_test / iters_test
            print('Saving..')
            state = {
                'net':  {key: model[key].state_dict() for key in model}, 
                'optimizer': optimizer.state_dict(),
                'iters': iters,
                'val_loss': loss_test / iters_test,
                'epoch': epoch,
            }
            save_path = osp.join(log_dir, 'epoch_2nd_%05d.pth' % epoch)
            torch.save(state, save_path)

            # if estimate sigma, save the estimated simga
            if model_params.diffusion.dist.estimate_sigma_data:
                config['model_params']['diffusion']['dist']['sigma_data'] = float(np.mean(running_std))

                with open(osp.join(log_dir, osp.basename(config_path)), 'w') as outfile:
                    yaml.dump(config, outfile, default_flow_style=True)

                            
if __name__=="__main__":
    main()