File size: 9,092 Bytes
daf0288
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import wandb
from pathlib import Path
from typing import Tuple, List, Union, Dict
from omegaconf import DictConfig
from hydra.utils import instantiate
import logging
import torch
import time
from torch import nn, Tensor, autograd
from torch.utils.data import DataLoader
from torch.nn.parallel import DistributedDataParallel as DDP

from ..utils import printer, compute_grad_norm
from ..trainer.utils import configure_optimizer_weight_decay

SNAPSHOT_KEYS = set(["EPOCH", "STEP", "OPTIMIZER", "LR_SCHEDULER", "MODEL", "LOSS"])


class BeitTrainer:
    def __init__(
        self,
        device: int,
        model: nn.Module,
        model_vqvae: nn.Module,
        log: logging.Logger,
        exp_dir: Path,
        snapshot: Path = None,
        model_weights: Path = None,  # only for testing
    ) -> None:
        self.device = device
        self.log = log
        self.exp_dir = exp_dir
        self.criterion = nn.CrossEntropyLoss()
        assert (
            snapshot is None or model_weights is None
        ), "Snapshot and model weights cannot be set at the same time."

        self.model = model
        if snapshot is not None and snapshot.is_file():
            self.snapshot = self.load_snapshot(snapshot)
            self.model.load_state_dict(self.snapshot["MODEL"])
            self.start_epoch = self.snapshot["EPOCH"]
            self.global_step = self.snapshot["STEP"]
        elif model_weights is not None and model_weights.is_file():
            self.load_model(model_weights)
        else:
            self.snapshot = None
            self.start_epoch = 0
            self.global_step = 0

        self.model = self.model.to(device)
        self.model = DDP(self.model, device_ids=[device])
        self.model_vqvae = model_vqvae.to(device)

        # https://discuss.pytorch.org/t/extra-10gb-memory-on-gpu-0-in-ddp-tutorial/118113
        torch.cuda.set_device(device)  # master gpu takes up extra memory
        torch.cuda.empty_cache()

    def train_epoch(self, epoch: int, grad_clip: float = None):
        start = time.time()
        total_loss = 0.0
        total_samples = 0

        for i, obj in enumerate(self.train_dataloader):
            (trans_image, vqvae_image), bool_mask_pos = obj
            trans_image, vqvae_image, bool_mask_pos = (
                trans_image.to(self.device),
                vqvae_image.to(self.device),
                bool_mask_pos.to(self.device),
            )

            with torch.no_grad():
                input_ids = self.model_vqvae.get_codebook_indices(vqvae_image).flatten(
                    1
                )
                bool_mask_pos = bool_mask_pos.flatten(1).to(torch.bool)
                labels = input_ids[bool_mask_pos]

            with autograd.detect_anomaly():
                outputs = self.model(
                    trans_image, bool_mask_pos, return_all_tokens=False
                )
                loss = self.criterion(outputs, labels)

                self.optimizer.zero_grad()
                loss.backward()
                if grad_clip:
                    nn.utils.clip_grad_norm_(
                        self.model.parameters(), max_norm=grad_clip
                    )
                self.optimizer.step()

            loss = loss.detach().cpu().data
            total_loss += loss * trans_image.shape[0]
            total_samples += trans_image.shape[0]

            self.lr_scheduler.step()
            self.global_step += 1

            if i % 10 == 0:
                grad_norm = compute_grad_norm(self.model)
                lr = self.optimizer.param_groups[0]["lr"]
                elapsed = time.time() - start
                self.log.info(
                    printer(
                        self.device,
                        f"Epoch {epoch} Step {i + 1}/{len(self.train_dataloader)} | Loss {loss:.4f} ({total_loss / total_samples:.4f}) | Grad norm {grad_norm:.3f} | {total_samples / elapsed:4.1f} images/s | lr {lr:5.1e}",
                    )
                )

            if i % 100 == 0 and self.device == 0:
                lr = self.optimizer.param_groups[0]["lr"]
                log_info = {
                    "epoch": epoch,
                    "train_loss": loss,
                    "learning rate": lr,
                    "grad_norm": grad_norm,
                }

                wandb.log(
                    log_info,
                    step=self.global_step,
                )

        return total_loss / total_samples

    def train(
        self,
        train_dataloader: DataLoader,
        valid_dataloader: DataLoader,
        train_cfg: DictConfig,
        valid_cfg: DictConfig,
    ):
        self.train_dataloader = train_dataloader
        self.valid_dataloader = valid_dataloader

        # ensure correct weight decay: https://github.com/karpathy/minGPT/blob/37baab71b9abea1b76ab957409a1cc2fbfba8a26/mingpt/model.py#L215
        optim_params = configure_optimizer_weight_decay(
            self.model.module, weight_decay=train_cfg.optimizer.weight_decay
        )
        self.optimizer = instantiate(train_cfg.optimizer, optim_params)

        self.lr_scheduler = instantiate(
            train_cfg.lr_scheduler, optimizer=self.optimizer
        )

        if self.snapshot is not None:
            self.optimizer.load_state_dict(self.snapshot["OPTIMIZER"])
            self.lr_scheduler.load_state_dict(self.snapshot["LR_SCHEDULER"])

        best_loss = float("inf")
        self.model.train()
        for epoch in range(self.start_epoch, train_cfg.epochs):
            train_dataloader.sampler.set_epoch(epoch)
            train_loss = self.train_epoch(epoch, grad_clip=train_cfg.grad_clip)

            torch.cuda.empty_cache()

            valid_loss = self.valid(valid_cfg)

            if self.device == 0:
                wandb.log(
                    {
                        "train loss (epoch)": train_loss,
                        "valid loss (epoch)": valid_loss,
                    },
                    step=self.global_step,
                )

                if epoch % train_cfg.save_every == 0:
                    self.save_snapshot(epoch, best_loss)
                if valid_loss < best_loss:
                    self.save_model(epoch)
                    best_loss = valid_loss

    def valid(self, cfg: DictConfig):
        total_samples = 0
        total_loss = 0.0

        self.model.eval()
        for i, obj in enumerate(self.valid_dataloader):
            (trans_image, vqvae_image), bool_mask_pos = obj
            trans_image, vqvae_image, bool_mask_pos = (
                trans_image.to(self.device),
                vqvae_image.to(self.device),
                bool_mask_pos.to(self.device),
            )

            with torch.no_grad():
                input_ids = self.model_vqvae.get_codebook_indices(vqvae_image).flatten(
                    1
                )
                bool_mask_pos = bool_mask_pos.flatten(1).to(torch.bool)
                labels = input_ids[bool_mask_pos]

                outputs = self.model(
                    trans_image, bool_mask_pos, return_all_tokens=False
                )
                loss = self.criterion(outputs, labels)

                loss = loss.detach().cpu().data
                total_loss += loss * trans_image.shape[0]
                total_samples += trans_image.shape[0]

            if i % 10 == 0:
                self.log.info(
                    printer(
                        self.device,
                        f"Valid: Step {i + 1}/{len(self.valid_dataloader)} | Loss {loss:.4f} ({total_loss / total_samples:.4f})",
                    )
                )

        return total_loss / total_samples

    def save_model(self, epoch: int):
        filename = Path(self.exp_dir) / "model" / f"epoch{epoch}_model.pt"
        torch.save(self.model.module.state_dict(), filename)
        self.log.info(printer(self.device, f"Saving model to {filename}"))
        filename = Path(self.exp_dir) / "model" / f"best.pt"
        torch.save(self.model.module.state_dict(), filename)

    def load_model(self, path: Union[str, Path]):
        self.model.load_state_dict(torch.load(path, map_location="cpu"))
        self.log.info(printer(self.device, f"Loading model from {path}"))

    def save_snapshot(self, epoch: int, best_loss: float):
        state_info = {
            "EPOCH": epoch + 1,
            "STEP": self.global_step,
            "OPTIMIZER": self.optimizer.state_dict(),
            "LR_SCHEDULER": self.lr_scheduler.state_dict(),
            "MODEL": self.model.module.state_dict(),
            "LOSS": best_loss,
        }

        snapshot_path = Path(self.exp_dir) / "snapshot" / f"epoch{epoch}_snapshot.pt"
        torch.save(state_info, snapshot_path)

        self.log.info(printer(self.device, f"Saving snapshot to {snapshot_path}"))

    def load_snapshot(self, path: Path):
        self.log.info(printer(self.device, f"Loading snapshot from {path}"))
        snapshot = torch.load(path, map_location="cpu")
        assert SNAPSHOT_KEYS.issubset(snapshot.keys())
        return snapshot