File size: 7,060 Bytes
ce252ec
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import time
import numpy as np
import torch
import librosa
from diffusion.logger.saver import Saver
from diffusion.logger import utils
from torch import autocast
from torch.cuda.amp import GradScaler

def test(args, model, vocoder, loader_test, saver):
    print(' [*] testing...')
    model.eval()

    # losses
    test_loss = 0.
    
    # intialization
    num_batches = len(loader_test)
    rtf_all = []
    
    # run
    with torch.no_grad():
        for bidx, data in enumerate(loader_test):
            fn = data['name'][0].split("/")[-1]
            speaker = data['name'][0].split("/")[-2]
            print('--------')
            print('{}/{} - {}'.format(bidx, num_batches, fn))

            # unpack data
            for k in data.keys():
                if not k.startswith('name'):
                    data[k] = data[k].to(args.device)
            print('>>', data['name'][0])

            # forward
            st_time = time.time()
            mel = model(
                    data['units'], 
                    data['f0'], 
                    data['volume'], 
                    data['spk_id'],
                    gt_spec=None,
                    infer=True, 
                    infer_speedup=args.infer.speedup, 
                    method=args.infer.method)
            signal = vocoder.infer(mel, data['f0'])
            ed_time = time.time()
                        
            # RTF
            run_time = ed_time - st_time
            song_time = signal.shape[-1] / args.data.sampling_rate
            rtf = run_time / song_time
            print('RTF: {}  | {} / {}'.format(rtf, run_time, song_time))
            rtf_all.append(rtf)
           
            # loss
            for i in range(args.train.batch_size):
                loss = model(
                    data['units'], 
                    data['f0'], 
                    data['volume'], 
                    data['spk_id'], 
                    gt_spec=data['mel'],
                    infer=False)
                test_loss += loss.item()
            
            # log mel
            saver.log_spec(f"{speaker}_{fn}.wav", data['mel'], mel)
            
            # log audi
            path_audio = data['name_ext'][0]
            audio, sr = librosa.load(path_audio, sr=args.data.sampling_rate)
            if len(audio.shape) > 1:
                audio = librosa.to_mono(audio)
            audio = torch.from_numpy(audio).unsqueeze(0).to(signal)
            saver.log_audio({f"{speaker}_{fn}_gt.wav": audio,f"{speaker}_{fn}_pred.wav": signal})
    # report
    test_loss /= args.train.batch_size
    test_loss /= num_batches 
    
    # check
    print(' [test_loss] test_loss:', test_loss)
    print(' Real Time Factor', np.mean(rtf_all))
    return test_loss


def train(args, initial_global_step, model, optimizer, scheduler, vocoder, loader_train, loader_test):
    # saver
    saver = Saver(args, initial_global_step=initial_global_step)

    # model size
    params_count = utils.get_network_paras_amount({'model': model})
    saver.log_info('--- model size ---')
    saver.log_info(params_count)
    
    # run
    num_batches = len(loader_train)
    model.train()
    saver.log_info('======= start training =======')
    scaler = GradScaler()
    if args.train.amp_dtype == 'fp32':
        dtype = torch.float32
    elif args.train.amp_dtype == 'fp16':
        dtype = torch.float16
    elif args.train.amp_dtype == 'bf16':
        dtype = torch.bfloat16
    else:
        raise ValueError(' [x] Unknown amp_dtype: ' + args.train.amp_dtype)
    saver.log_info("epoch|batch_idx/num_batches|output_dir|batch/s|lr|time|step")
    for epoch in range(args.train.epochs):
        for batch_idx, data in enumerate(loader_train):
            saver.global_step_increment()
            optimizer.zero_grad()

            # unpack data
            for k in data.keys():
                if not k.startswith('name'):
                    data[k] = data[k].to(args.device)
            
            # forward
            if dtype == torch.float32:
                loss = model(data['units'].float(), data['f0'], data['volume'], data['spk_id'], 
                                aug_shift = data['aug_shift'], gt_spec=data['mel'].float(), infer=False)
            else:
                with autocast(device_type=args.device, dtype=dtype):
                    loss = model(data['units'], data['f0'], data['volume'], data['spk_id'], 
                                    aug_shift = data['aug_shift'], gt_spec=data['mel'], infer=False)
            
            # handle nan loss
            if torch.isnan(loss):
                raise ValueError(' [x] nan loss ')
            else:
                # backpropagate
                if dtype == torch.float32:
                    loss.backward()
                    optimizer.step()
                else:
                    scaler.scale(loss).backward()
                    scaler.step(optimizer)
                    scaler.update()
                scheduler.step()
                
            # log loss
            if saver.global_step % args.train.interval_log == 0:
                current_lr =  optimizer.param_groups[0]['lr']
                saver.log_info(
                    'epoch: {} | {:3d}/{:3d} | {} | batch/s: {:.2f} | lr: {:.6} | loss: {:.3f} | time: {} | step: {}'.format(
                        epoch,
                        batch_idx,
                        num_batches,
                        args.env.expdir,
                        args.train.interval_log/saver.get_interval_time(),
                        current_lr,
                        loss.item(),
                        saver.get_total_time(),
                        saver.global_step
                    )
                )
                
                saver.log_value({
                    'train/loss': loss.item()
                })
                
                saver.log_value({
                    'train/lr': current_lr
                })
            
            # validation
            if saver.global_step % args.train.interval_val == 0:
                optimizer_save = optimizer if args.train.save_opt else None
                
                # save latest
                saver.save_model(model, optimizer_save, postfix=f'{saver.global_step}')
                last_val_step = saver.global_step - args.train.interval_val
                if last_val_step % args.train.interval_force_save != 0:
                    saver.delete_model(postfix=f'{last_val_step}')
                
                # run testing set
                test_loss = test(args, model, vocoder, loader_test, saver)
                
                # log loss
                saver.log_info(
                    ' --- <validation> --- \nloss: {:.3f}. '.format(
                        test_loss,
                    )
                )
                
                saver.log_value({
                    'validation/loss': test_loss
                })
                
                model.train()