File size: 8,407 Bytes
d61b9c7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Scene Text Recognition Model Hub
# Copyright 2022 Darwin Bautista
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import logging
import math
from typing import Any, Tuple, List, Optional

import torch
import torch.nn.functional as F
from torch import Tensor, nn
from torch.optim import AdamW
from torch.optim.lr_scheduler import OneCycleLR

from pytorch_lightning.utilities.types import STEP_OUTPUT
from timm.optim.optim_factory import param_groups_weight_decay

from strhub.models.base import CrossEntropySystem
from strhub.models.utils import init_weights
from .model_abinet_iter import ABINetIterModel as Model

log = logging.getLogger(__name__)


class ABINet(CrossEntropySystem):

    def __init__(self, charset_train: str, charset_test: str, max_label_length: int,
                 batch_size: int, lr: float, warmup_pct: float, weight_decay: float,
                 iter_size: int, d_model: int, nhead: int, d_inner: int, dropout: float, activation: str,
                 v_loss_weight: float, v_attention: str, v_attention_mode: str, v_backbone: str, v_num_layers: int,
                 l_loss_weight: float, l_num_layers: int, l_detach: bool, l_use_self_attn: bool,
                 l_lr: float, a_loss_weight: float, lm_only: bool = False, **kwargs) -> None:
        super().__init__(charset_train, charset_test, batch_size, lr, warmup_pct, weight_decay)
        self.scheduler = None
        self.save_hyperparameters()
        self.max_label_length = max_label_length
        self.num_classes = len(self.tokenizer) - 2  # We don't predict <bos> nor <pad>
        self.model = Model(max_label_length, self.eos_id, self.num_classes, iter_size, d_model, nhead, d_inner,
                           dropout, activation, v_loss_weight, v_attention, v_attention_mode, v_backbone, v_num_layers,
                           l_loss_weight, l_num_layers, l_detach, l_use_self_attn, a_loss_weight)
        self.model.apply(init_weights)
        # FIXME: doesn't support resumption from checkpoint yet
        self._reset_alignment = True
        self._reset_optimizers = True
        self.l_lr = l_lr
        self.lm_only = lm_only
        # Train LM only. Freeze other submodels.
        if lm_only:
            self.l_lr = lr  # for tuning
            self.model.vision.requires_grad_(False)
            self.model.alignment.requires_grad_(False)

    @property
    def _pretraining(self):
        # In the original work, VM was pretrained for 8 epochs while full model was trained for an additional 10 epochs.
        total_steps = self.trainer.estimated_stepping_batches * self.trainer.accumulate_grad_batches
        return self.global_step < (8 / (8 + 10)) * total_steps

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'model.language.proj.weight'}

    def _add_weight_decay(self, model: nn.Module, skip_list=()):
        if self.weight_decay:
            return param_groups_weight_decay(model, self.weight_decay, skip_list)
        else:
            return [{'params': model.parameters()}]

    def configure_optimizers(self):
        agb = self.trainer.accumulate_grad_batches
        # Linear scaling so that the effective learning rate is constant regardless of the number of GPUs used with DDP.
        lr_scale = agb * math.sqrt(self.trainer.num_devices) * self.batch_size / 256.
        lr = lr_scale * self.lr
        l_lr = lr_scale * self.l_lr
        params = []
        params.extend(self._add_weight_decay(self.model.vision))
        params.extend(self._add_weight_decay(self.model.alignment))
        # We use a different learning rate for the LM.
        for p in self._add_weight_decay(self.model.language, ('proj.weight',)):
            p['lr'] = l_lr
            params.append(p)
        max_lr = [p.get('lr', lr) for p in params]
        optim = AdamW(params, lr)
        self.scheduler = OneCycleLR(optim, max_lr, self.trainer.estimated_stepping_batches,
                                    pct_start=self.warmup_pct, cycle_momentum=False)
        return {'optimizer': optim, 'lr_scheduler': {'scheduler': self.scheduler, 'interval': 'step'}}

    def forward(self, images: Tensor, max_length: Optional[int] = None) -> Tensor:
        max_length = self.max_label_length if max_length is None else min(max_length, self.max_label_length)
        logits = self.model.forward(images)[0]['logits']
        return logits[:, :max_length + 1]  # truncate

    def calc_loss(self, targets, *res_lists) -> Tensor:
        total_loss = 0
        for res_list in res_lists:
            loss = 0
            if isinstance(res_list, dict):
                res_list = [res_list]
            for res in res_list:
                logits = res['logits'].flatten(end_dim=1)
                loss += F.cross_entropy(logits, targets.flatten(), ignore_index=self.pad_id)
            loss /= len(res_list)
            self.log('loss_' + res_list[0]['name'], loss)
            total_loss += res_list[0]['loss_weight'] * loss
        return total_loss

    def on_train_batch_start(self, batch: Any, batch_idx: int) -> None:
        if not self._pretraining and self._reset_optimizers:
            log.info('Pretraining ends. Updating base LRs.')
            self._reset_optimizers = False
            # Make base_lr the same for all groups
            base_lr = self.scheduler.base_lrs[0]  # base_lr of group 0 - VM
            self.scheduler.base_lrs = [base_lr] * len(self.scheduler.base_lrs)

    def _prepare_inputs_and_targets(self, labels):
        # Use dummy label to ensure sequence length is constant.
        dummy = ['0' * self.max_label_length]
        targets = self.tokenizer.encode(dummy + list(labels), self.device)[1:]
        targets = targets[:, 1:]  # remove <bos>. Unused here.
        # Inputs are padded with eos_id
        inputs = torch.where(targets == self.pad_id, self.eos_id, targets)
        inputs = F.one_hot(inputs, self.num_classes).float()
        lengths = torch.as_tensor(list(map(len, labels)), device=self.device) + 1  # +1 for eos
        return inputs, lengths, targets

    def training_step(self, batch, batch_idx) -> STEP_OUTPUT:
        images, labels = batch
        inputs, lengths, targets = self._prepare_inputs_and_targets(labels)
        if self.lm_only:
            l_res = self.model.language(inputs, lengths)
            loss = self.calc_loss(targets, l_res)
        # Pretrain submodels independently first
        elif self._pretraining:
            # Vision
            v_res = self.model.vision(images)
            # Language
            l_res = self.model.language(inputs, lengths)
            # We also train the alignment model to 'satisfy' DDP requirements (all parameters should be used).
            # We'll reset its parameters prior to joint training.
            a_res = self.model.alignment(l_res['feature'].detach(), v_res['feature'].detach())
            loss = self.calc_loss(targets, v_res, l_res, a_res)
        else:
            # Reset alignment model's parameters once prior to full model training.
            if self._reset_alignment:
                log.info('Pretraining ends. Resetting alignment model.')
                self._reset_alignment = False
                self.model.alignment.apply(init_weights)
            all_a_res, all_l_res, v_res = self.model.forward(images)
            loss = self.calc_loss(targets, v_res, all_l_res, all_a_res)
        self.log('loss', loss)
        return loss

    def forward_logits_loss(self, images: Tensor, labels: List[str]) -> Tuple[Tensor, Tensor, int]:
        if self.lm_only:
            inputs, lengths, targets = self._prepare_inputs_and_targets(labels)
            l_res = self.model.language(inputs, lengths)
            loss = self.calc_loss(targets, l_res)
            loss_numel = (targets != self.pad_id).sum()
            return l_res['logits'], loss, loss_numel
        else:
            return super().forward_logits_loss(images, labels)