File size: 11,148 Bytes
1646c30
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import os
import gc
from tqdm import tqdm
import wandb

import torch
from torch.optim import AdamW
from torch.utils.data import DataLoader, Dataset, SequentialSampler
from torch.optim.lr_scheduler import LinearLR, SequentialLR

from einops import rearrange

from accelerate import Accelerator
from accelerate.utils import DistributedDataParallelKwargs

from ema_pytorch import EMA

from model import CFM
from model.utils import exists, default
from model.dataset import DynamicBatchSampler, collate_fn


# trainer

class Trainer:
    def __init__(
        self,
        model: CFM,
        epochs,
        learning_rate,
        num_warmup_updates = 20000,
        save_per_updates = 1000, 
        checkpoint_path = None,
        batch_size = 32, 
        batch_size_type: str = "sample",
        max_samples = 32,
        grad_accumulation_steps = 1,
        max_grad_norm = 1.0,
        noise_scheduler: str | None = None,
        duration_predictor: torch.nn.Module | None = None,
        wandb_project = "test_e2-tts",
        wandb_run_name = "test_run",
        wandb_resume_id: str = None,
        last_per_steps = None,
        accelerate_kwargs: dict = dict(),
        ema_kwargs: dict = dict()
    ):
        
        ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters = True)

        self.accelerator = Accelerator(
            log_with = "wandb",
            kwargs_handlers = [ddp_kwargs],
            gradient_accumulation_steps = grad_accumulation_steps,
            **accelerate_kwargs
        )
        
        if exists(wandb_resume_id):
            init_kwargs={"wandb": {"resume": "allow", "name": wandb_run_name, 'id': wandb_resume_id}}
        else:
            init_kwargs={"wandb": {"resume": "allow", "name": wandb_run_name}}
        self.accelerator.init_trackers(
            project_name = wandb_project, 
            init_kwargs=init_kwargs,
            config={"epochs": epochs,
                    "learning_rate": learning_rate,
                    "num_warmup_updates": num_warmup_updates, 
                    "batch_size": batch_size,
                    "batch_size_type": batch_size_type,
                    "max_samples": max_samples,
                    "grad_accumulation_steps": grad_accumulation_steps,
                    "max_grad_norm": max_grad_norm,
                    "gpus": self.accelerator.num_processes,
                    "noise_scheduler": noise_scheduler}
            )

        self.model = model

        if self.is_main:
            self.ema_model = EMA(
                model,
                include_online_model = False,
                **ema_kwargs
            )

            self.ema_model.to(self.accelerator.device)

        self.epochs = epochs
        self.num_warmup_updates = num_warmup_updates
        self.save_per_updates = save_per_updates
        self.last_per_steps = default(last_per_steps, save_per_updates * grad_accumulation_steps)
        self.checkpoint_path = default(checkpoint_path, 'ckpts/test_e2-tts')

        self.batch_size = batch_size
        self.batch_size_type = batch_size_type
        self.max_samples = max_samples
        self.grad_accumulation_steps = grad_accumulation_steps
        self.max_grad_norm = max_grad_norm

        self.noise_scheduler = noise_scheduler

        self.duration_predictor = duration_predictor

        self.optimizer = AdamW(model.parameters(), lr=learning_rate)
        self.model, self.optimizer = self.accelerator.prepare(
            self.model, self.optimizer
        )

    @property
    def is_main(self):
        return self.accelerator.is_main_process

    def save_checkpoint(self, step, last=False):
        self.accelerator.wait_for_everyone()
        if self.is_main:
            checkpoint = dict(
                model_state_dict = self.accelerator.unwrap_model(self.model).state_dict(),
                optimizer_state_dict = self.accelerator.unwrap_model(self.optimizer).state_dict(),
                ema_model_state_dict = self.ema_model.state_dict(),
                scheduler_state_dict = self.scheduler.state_dict(),
                step = step
            )
            if not os.path.exists(self.checkpoint_path):
                os.makedirs(self.checkpoint_path)
            if last == True:
                self.accelerator.save(checkpoint, f"{self.checkpoint_path}/model_last.pt")
                print(f"Saved last checkpoint at step {step}")
            else:
                self.accelerator.save(checkpoint, f"{self.checkpoint_path}/model_{step}.pt")

    def load_checkpoint(self):
        if not exists(self.checkpoint_path) or not os.path.exists(self.checkpoint_path) or not os.listdir(self.checkpoint_path):
            return 0
        
        self.accelerator.wait_for_everyone()
        if "model_last.pt" in os.listdir(self.checkpoint_path):
            latest_checkpoint = "model_last.pt"
        else:
            latest_checkpoint = sorted(os.listdir(self.checkpoint_path), key=lambda x: int(''.join(filter(str.isdigit, x))))[-1]
        # checkpoint = torch.load(f"{self.checkpoint_path}/{latest_checkpoint}", map_location=self.accelerator.device)  # rather use accelerator.load_state ಥ_ಥ
        checkpoint = torch.load(f"{self.checkpoint_path}/{latest_checkpoint}", map_location="cpu")
        self.accelerator.unwrap_model(self.model).load_state_dict(checkpoint['model_state_dict'])
        self.accelerator.unwrap_model(self.optimizer).load_state_dict(checkpoint['optimizer_state_dict'])

        if self.is_main:
            self.ema_model.load_state_dict(checkpoint['ema_model_state_dict'])

        if self.scheduler:
            self.scheduler.load_state_dict(checkpoint['scheduler_state_dict'])
        
        step = checkpoint['step']
        del checkpoint; gc.collect()
        return step

    def train(self, train_dataset: Dataset, num_workers=16, resumable_with_seed: int = None):
        
        if exists(resumable_with_seed):
            generator = torch.Generator()
            generator.manual_seed(resumable_with_seed)
        else: 
            generator = None

        if self.batch_size_type == "sample":
            train_dataloader = DataLoader(train_dataset, collate_fn=collate_fn, num_workers=num_workers, pin_memory=True, 
                                          batch_size=self.batch_size, shuffle=True, generator=generator)
        elif self.batch_size_type == "frame":
            self.accelerator.even_batches = False
            sampler = SequentialSampler(train_dataset)
            batch_sampler = DynamicBatchSampler(sampler, self.batch_size, max_samples=self.max_samples, random_seed=resumable_with_seed, drop_last=False)
            train_dataloader = DataLoader(train_dataset, collate_fn=collate_fn, num_workers=num_workers, pin_memory=True,
                                          batch_sampler=batch_sampler)
        else:
            raise ValueError(f"batch_size_type must be either 'sample' or 'frame', but recieved {self.batch_size_type}")
        
        #  accelerator.prepare() dispatches batches to devices;
        #  which means the length of dataloader calculated before, should consider the number of devices
        warmup_steps = self.num_warmup_updates * self.accelerator.num_processes  # consider a fixed warmup steps while using accelerate multi-gpu ddp
                                                                                 # otherwise by default with split_batches=False, warmup steps change with num_processes
        total_steps = len(train_dataloader) * self.epochs / self.grad_accumulation_steps
        decay_steps = total_steps - warmup_steps
        warmup_scheduler = LinearLR(self.optimizer, start_factor=1e-8, end_factor=1.0, total_iters=warmup_steps)
        decay_scheduler = LinearLR(self.optimizer, start_factor=1.0, end_factor=1e-8, total_iters=decay_steps)
        self.scheduler = SequentialLR(self.optimizer, 
                                      schedulers=[warmup_scheduler, decay_scheduler],
                                      milestones=[warmup_steps])
        train_dataloader, self.scheduler = self.accelerator.prepare(train_dataloader, self.scheduler)  # actual steps = 1 gpu steps / gpus
        start_step = self.load_checkpoint()
        global_step = start_step

        if exists(resumable_with_seed):
            orig_epoch_step = len(train_dataloader)
            skipped_epoch = int(start_step // orig_epoch_step)
            skipped_batch = start_step % orig_epoch_step
            skipped_dataloader = self.accelerator.skip_first_batches(train_dataloader, num_batches=skipped_batch)
        else:
            skipped_epoch = 0

        for epoch in range(skipped_epoch, self.epochs):
            self.model.train()
            if exists(resumable_with_seed) and epoch == skipped_epoch:
                progress_bar = tqdm(skipped_dataloader, desc=f"Epoch {epoch+1}/{self.epochs}", unit="step", disable=not self.accelerator.is_local_main_process, 
                                    initial=skipped_batch, total=orig_epoch_step)
            else:
                progress_bar = tqdm(train_dataloader, desc=f"Epoch {epoch+1}/{self.epochs}", unit="step", disable=not self.accelerator.is_local_main_process)

            for batch in progress_bar:
                with self.accelerator.accumulate(self.model):
                    text_inputs = batch['text']
                    mel_spec = rearrange(batch['mel'], 'b d n -> b n d')
                    mel_lengths = batch["mel_lengths"]

                    # TODO. add duration predictor training
                    if self.duration_predictor is not None and self.accelerator.is_local_main_process:
                        dur_loss = self.duration_predictor(mel_spec, lens=batch.get('durations'))
                        self.accelerator.log({"duration loss": dur_loss.item()}, step=global_step)

                    loss, cond, pred = self.model(mel_spec, text=text_inputs, lens=mel_lengths, noise_scheduler=self.noise_scheduler)
                    self.accelerator.backward(loss)

                    if self.max_grad_norm > 0 and self.accelerator.sync_gradients:
                        self.accelerator.clip_grad_norm_(self.model.parameters(), self.max_grad_norm)

                    self.optimizer.step()
                    self.scheduler.step()
                    self.optimizer.zero_grad()

                if self.is_main:
                    self.ema_model.update()

                global_step += 1

                if self.accelerator.is_local_main_process:
                    self.accelerator.log({"loss": loss.item(), "lr": self.scheduler.get_last_lr()[0]}, step=global_step)
                
                progress_bar.set_postfix(step=str(global_step), loss=loss.item())
                
                if global_step % (self.save_per_updates * self.grad_accumulation_steps) == 0:
                    self.save_checkpoint(global_step)
                
                if global_step % self.last_per_steps == 0:
                    self.save_checkpoint(global_step, last=True)
        
        self.accelerator.end_training()