File size: 14,189 Bytes
18dd6ad
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import copy
import logging
from typing import Dict, Tuple

import pandas as pd
import pytorch_lightning as ptl
import torch
import torch.nn as nn
import torch.nn.functional as F
# from torch.utils.data import DistributedSampler

# from annotator.lama.saicinpainting.evaluation import make_evaluator
# from annotator.lama.saicinpainting.training.data.datasets import make_default_train_dataloader, make_default_val_dataloader
# from annotator.lama.saicinpainting.training.losses.adversarial import make_discrim_loss
# from annotator.lama.saicinpainting.training.losses.perceptual import PerceptualLoss, ResNetPL
from annotator.lama.saicinpainting.training.modules import make_generator  #, make_discriminator
# from annotator.lama.saicinpainting.training.visualizers import make_visualizer
from annotator.lama.saicinpainting.utils import add_prefix_to_keys, average_dicts, set_requires_grad, flatten_dict, \
    get_has_ddp_rank

LOGGER = logging.getLogger(__name__)


def make_optimizer(parameters, kind='adamw', **kwargs):
    if kind == 'adam':
        optimizer_class = torch.optim.Adam
    elif kind == 'adamw':
        optimizer_class = torch.optim.AdamW
    else:
        raise ValueError(f'Unknown optimizer kind {kind}')
    return optimizer_class(parameters, **kwargs)


def update_running_average(result: nn.Module, new_iterate_model: nn.Module, decay=0.999):
    with torch.no_grad():
        res_params = dict(result.named_parameters())
        new_params = dict(new_iterate_model.named_parameters())

        for k in res_params.keys():
            res_params[k].data.mul_(decay).add_(new_params[k].data, alpha=1 - decay)


def make_multiscale_noise(base_tensor, scales=6, scale_mode='bilinear'):
    batch_size, _, height, width = base_tensor.shape
    cur_height, cur_width = height, width
    result = []
    align_corners = False if scale_mode in ('bilinear', 'bicubic') else None
    for _ in range(scales):
        cur_sample = torch.randn(batch_size, 1, cur_height, cur_width, device=base_tensor.device)
        cur_sample_scaled = F.interpolate(cur_sample, size=(height, width), mode=scale_mode, align_corners=align_corners)
        result.append(cur_sample_scaled)
        cur_height //= 2
        cur_width //= 2
    return torch.cat(result, dim=1)


class BaseInpaintingTrainingModule(ptl.LightningModule):
    def __init__(self, config, use_ddp, *args,  predict_only=False, visualize_each_iters=100,
                 average_generator=False, generator_avg_beta=0.999, average_generator_start_step=30000,
                 average_generator_period=10, store_discr_outputs_for_vis=False,
                 **kwargs):
        super().__init__(*args, **kwargs)
        LOGGER.info('BaseInpaintingTrainingModule init called')

        self.config = config

        self.generator = make_generator(config, **self.config.generator)
        self.use_ddp = use_ddp

        if not get_has_ddp_rank():
            LOGGER.info(f'Generator\n{self.generator}')

        # if not predict_only:
        #     self.save_hyperparameters(self.config)
        #     self.discriminator = make_discriminator(**self.config.discriminator)
        #     self.adversarial_loss = make_discrim_loss(**self.config.losses.adversarial)
        #     self.visualizer = make_visualizer(**self.config.visualizer)
        #     self.val_evaluator = make_evaluator(**self.config.evaluator)
        #     self.test_evaluator = make_evaluator(**self.config.evaluator)
        #
        #     if not get_has_ddp_rank():
        #         LOGGER.info(f'Discriminator\n{self.discriminator}')
        #
        #     extra_val = self.config.data.get('extra_val', ())
        #     if extra_val:
        #         self.extra_val_titles = list(extra_val)
        #         self.extra_evaluators = nn.ModuleDict({k: make_evaluator(**self.config.evaluator)
        #                                                for k in extra_val})
        #     else:
        #         self.extra_evaluators = {}
        #
        #     self.average_generator = average_generator
        #     self.generator_avg_beta = generator_avg_beta
        #     self.average_generator_start_step = average_generator_start_step
        #     self.average_generator_period = average_generator_period
        #     self.generator_average = None
        #     self.last_generator_averaging_step = -1
        #     self.store_discr_outputs_for_vis = store_discr_outputs_for_vis
        #
        #     if self.config.losses.get("l1", {"weight_known": 0})['weight_known'] > 0:
        #         self.loss_l1 = nn.L1Loss(reduction='none')
        #
        #     if self.config.losses.get("mse", {"weight": 0})['weight'] > 0:
        #         self.loss_mse = nn.MSELoss(reduction='none')
        #
        #     if self.config.losses.perceptual.weight > 0:
        #         self.loss_pl = PerceptualLoss()
        #
        #     # if self.config.losses.get("resnet_pl", {"weight": 0})['weight'] > 0:
        #     #     self.loss_resnet_pl = ResNetPL(**self.config.losses.resnet_pl)
        #     # else:
        #     #     self.loss_resnet_pl = None
        #
        #     self.loss_resnet_pl = None

        self.visualize_each_iters = visualize_each_iters
        LOGGER.info('BaseInpaintingTrainingModule init done')

    def configure_optimizers(self):
        discriminator_params = list(self.discriminator.parameters())
        return [
            dict(optimizer=make_optimizer(self.generator.parameters(), **self.config.optimizers.generator)),
            dict(optimizer=make_optimizer(discriminator_params, **self.config.optimizers.discriminator)),
        ]

    def train_dataloader(self):
        kwargs = dict(self.config.data.train)
        if self.use_ddp:
            kwargs['ddp_kwargs'] = dict(num_replicas=self.trainer.num_nodes * self.trainer.num_processes,
                                        rank=self.trainer.global_rank,
                                        shuffle=True)
        dataloader = make_default_train_dataloader(**self.config.data.train)
        return dataloader

    def val_dataloader(self):
        res = [make_default_val_dataloader(**self.config.data.val)]

        if self.config.data.visual_test is not None:
            res = res + [make_default_val_dataloader(**self.config.data.visual_test)]
        else:
            res = res + res

        extra_val = self.config.data.get('extra_val', ())
        if extra_val:
            res += [make_default_val_dataloader(**extra_val[k]) for k in self.extra_val_titles]

        return res

    def training_step(self, batch, batch_idx, optimizer_idx=None):
        self._is_training_step = True
        return self._do_step(batch, batch_idx, mode='train', optimizer_idx=optimizer_idx)

    def validation_step(self, batch, batch_idx, dataloader_idx):
        extra_val_key = None
        if dataloader_idx == 0:
            mode = 'val'
        elif dataloader_idx == 1:
            mode = 'test'
        else:
            mode = 'extra_val'
            extra_val_key = self.extra_val_titles[dataloader_idx - 2]
        self._is_training_step = False
        return self._do_step(batch, batch_idx, mode=mode, extra_val_key=extra_val_key)

    def training_step_end(self, batch_parts_outputs):
        if self.training and self.average_generator \
                and self.global_step >= self.average_generator_start_step \
                and self.global_step >= self.last_generator_averaging_step + self.average_generator_period:
            if self.generator_average is None:
                self.generator_average = copy.deepcopy(self.generator)
            else:
                update_running_average(self.generator_average, self.generator, decay=self.generator_avg_beta)
            self.last_generator_averaging_step = self.global_step

        full_loss = (batch_parts_outputs['loss'].mean()
                     if torch.is_tensor(batch_parts_outputs['loss'])  # loss is not tensor when no discriminator used
                     else torch.tensor(batch_parts_outputs['loss']).float().requires_grad_(True))
        log_info = {k: v.mean() for k, v in batch_parts_outputs['log_info'].items()}
        self.log_dict(log_info, on_step=True, on_epoch=False)
        return full_loss

    def validation_epoch_end(self, outputs):
        outputs = [step_out for out_group in outputs for step_out in out_group]
        averaged_logs = average_dicts(step_out['log_info'] for step_out in outputs)
        self.log_dict({k: v.mean() for k, v in averaged_logs.items()})

        pd.set_option('display.max_columns', 500)
        pd.set_option('display.width', 1000)

        # standard validation
        val_evaluator_states = [s['val_evaluator_state'] for s in outputs if 'val_evaluator_state' in s]
        val_evaluator_res = self.val_evaluator.evaluation_end(states=val_evaluator_states)
        val_evaluator_res_df = pd.DataFrame(val_evaluator_res).stack(1).unstack(0)
        val_evaluator_res_df.dropna(axis=1, how='all', inplace=True)
        LOGGER.info(f'Validation metrics after epoch #{self.current_epoch}, '
                    f'total {self.global_step} iterations:\n{val_evaluator_res_df}')

        for k, v in flatten_dict(val_evaluator_res).items():
            self.log(f'val_{k}', v)

        # standard visual test
        test_evaluator_states = [s['test_evaluator_state'] for s in outputs
                                 if 'test_evaluator_state' in s]
        test_evaluator_res = self.test_evaluator.evaluation_end(states=test_evaluator_states)
        test_evaluator_res_df = pd.DataFrame(test_evaluator_res).stack(1).unstack(0)
        test_evaluator_res_df.dropna(axis=1, how='all', inplace=True)
        LOGGER.info(f'Test metrics after epoch #{self.current_epoch}, '
                    f'total {self.global_step} iterations:\n{test_evaluator_res_df}')

        for k, v in flatten_dict(test_evaluator_res).items():
            self.log(f'test_{k}', v)

        # extra validations
        if self.extra_evaluators:
            for cur_eval_title, cur_evaluator in self.extra_evaluators.items():
                cur_state_key = f'extra_val_{cur_eval_title}_evaluator_state'
                cur_states = [s[cur_state_key] for s in outputs if cur_state_key in s]
                cur_evaluator_res = cur_evaluator.evaluation_end(states=cur_states)
                cur_evaluator_res_df = pd.DataFrame(cur_evaluator_res).stack(1).unstack(0)
                cur_evaluator_res_df.dropna(axis=1, how='all', inplace=True)
                LOGGER.info(f'Extra val {cur_eval_title} metrics after epoch #{self.current_epoch}, '
                            f'total {self.global_step} iterations:\n{cur_evaluator_res_df}')
                for k, v in flatten_dict(cur_evaluator_res).items():
                    self.log(f'extra_val_{cur_eval_title}_{k}', v)

    def _do_step(self, batch, batch_idx, mode='train', optimizer_idx=None, extra_val_key=None):
        if optimizer_idx == 0:  # step for generator
            set_requires_grad(self.generator, True)
            set_requires_grad(self.discriminator, False)
        elif optimizer_idx == 1:  # step for discriminator
            set_requires_grad(self.generator, False)
            set_requires_grad(self.discriminator, True)

        batch = self(batch)

        total_loss = 0
        metrics = {}

        if optimizer_idx is None or optimizer_idx == 0:  # step for generator
            total_loss, metrics = self.generator_loss(batch)

        elif optimizer_idx is None or optimizer_idx == 1:  # step for discriminator
            if self.config.losses.adversarial.weight > 0:
                total_loss, metrics = self.discriminator_loss(batch)

        if self.get_ddp_rank() in (None, 0) and (batch_idx % self.visualize_each_iters == 0 or mode == 'test'):
            if self.config.losses.adversarial.weight > 0:
                if self.store_discr_outputs_for_vis:
                    with torch.no_grad():
                        self.store_discr_outputs(batch)
            vis_suffix = f'_{mode}'
            if mode == 'extra_val':
                vis_suffix += f'_{extra_val_key}'
            self.visualizer(self.current_epoch, batch_idx, batch, suffix=vis_suffix)

        metrics_prefix = f'{mode}_'
        if mode == 'extra_val':
            metrics_prefix += f'{extra_val_key}_'
        result = dict(loss=total_loss, log_info=add_prefix_to_keys(metrics, metrics_prefix))
        if mode == 'val':
            result['val_evaluator_state'] = self.val_evaluator.process_batch(batch)
        elif mode == 'test':
            result['test_evaluator_state'] = self.test_evaluator.process_batch(batch)
        elif mode == 'extra_val':
            result[f'extra_val_{extra_val_key}_evaluator_state'] = self.extra_evaluators[extra_val_key].process_batch(batch)

        return result

    def get_current_generator(self, no_average=False):
        if not no_average and not self.training and self.average_generator and self.generator_average is not None:
            return self.generator_average
        return self.generator

    def forward(self, batch: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
        """Pass data through generator and obtain at leas 'predicted_image' and 'inpainted' keys"""
        raise NotImplementedError()

    def generator_loss(self, batch) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
        raise NotImplementedError()

    def discriminator_loss(self, batch) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
        raise NotImplementedError()

    def store_discr_outputs(self, batch):
        out_size = batch['image'].shape[2:]
        discr_real_out, _ = self.discriminator(batch['image'])
        discr_fake_out, _ = self.discriminator(batch['predicted_image'])
        batch['discr_output_real'] = F.interpolate(discr_real_out, size=out_size, mode='nearest')
        batch['discr_output_fake'] = F.interpolate(discr_fake_out, size=out_size, mode='nearest')
        batch['discr_output_diff'] = batch['discr_output_real'] - batch['discr_output_fake']

    def get_ddp_rank(self):
        return self.trainer.global_rank if (self.trainer.num_nodes * self.trainer.num_processes) > 1 else None