File size: 11,599 Bytes
ed1cdd1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
from datetime import datetime
import shutil

import matplotlib

matplotlib.use('Agg')

from utils.hparams import hparams, set_hparams
import random
import sys
import numpy as np
import torch.distributed as dist
from pytorch_lightning.loggers import TensorBoardLogger
from utils.pl_utils import LatestModelCheckpoint, BaseTrainer, data_loader, DDP
from torch import nn
import torch.utils.data
import utils
import logging
import os

torch.multiprocessing.set_sharing_strategy(os.getenv('TORCH_SHARE_STRATEGY', 'file_system'))

log_format = '%(asctime)s %(message)s'
logging.basicConfig(stream=sys.stdout, level=logging.INFO,
                    format=log_format, datefmt='%m/%d %I:%M:%S %p')

class BaseTask(nn.Module):
    '''
        Base class for training tasks.
        1. *load_ckpt*:
            load checkpoint;
        2. *training_step*:
            record and log the loss;
        3. *optimizer_step*:
            run backwards step;
        4. *start*:
            load training configs, backup code, log to tensorboard, start training;
        5. *configure_ddp* and *init_ddp_connection*:
            start parallel training.

        Subclasses should define:
        1. *build_model*, *build_optimizer*, *build_scheduler*:
            how to build the model, the optimizer and the training scheduler;
        2. *_training_step*:
            one training step of the model;
        3. *validation_end* and *_validation_end*:
            postprocess the validation output.
    '''
    def __init__(self, *args, **kwargs):
        # dataset configs
        super(BaseTask, self).__init__(*args, **kwargs)
        self.current_epoch = 0
        self.global_step = 0
        self.loaded_optimizer_states_dict = {}
        self.trainer = None
        self.logger = None
        self.on_gpu = False
        self.use_dp = False
        self.use_ddp = False
        self.example_input_array = None

        self.max_tokens = hparams['max_tokens']
        self.max_sentences = hparams['max_sentences']
        self.max_eval_tokens = hparams['max_eval_tokens']
        if self.max_eval_tokens == -1:
            hparams['max_eval_tokens'] = self.max_eval_tokens = self.max_tokens
        self.max_eval_sentences = hparams['max_eval_sentences']
        if self.max_eval_sentences == -1:
            hparams['max_eval_sentences'] = self.max_eval_sentences = self.max_sentences

        self.model = None
        self.training_losses_meter = None

    ###########
    # Training, validation and testing
    ###########
    def build_model(self):
        raise NotImplementedError

    def load_ckpt(self, ckpt_base_dir, current_model_name=None, model_name='model', force=True, strict=True):
        # This function is updated on 2021.12.13
        if current_model_name is None:
            current_model_name = model_name
        utils.load_ckpt(self.__getattr__(current_model_name), ckpt_base_dir, current_model_name, force, strict)

    def on_epoch_start(self):
        self.training_losses_meter = {'total_loss': utils.AvgrageMeter()}

    def _training_step(self, sample, batch_idx, optimizer_idx):
        """

        :param sample:
        :param batch_idx:
        :return: total loss: torch.Tensor, loss_log: dict
        """
        raise NotImplementedError

    def training_step(self, sample, batch_idx, optimizer_idx=-1):
        loss_ret = self._training_step(sample, batch_idx, optimizer_idx)
        self.opt_idx = optimizer_idx
        if loss_ret is None:
            return {'loss': None}
        total_loss, log_outputs = loss_ret
        log_outputs = utils.tensors_to_scalars(log_outputs)
        for k, v in log_outputs.items():
            if k not in self.training_losses_meter:
                self.training_losses_meter[k] = utils.AvgrageMeter()
            if not np.isnan(v):
                self.training_losses_meter[k].update(v)
        self.training_losses_meter['total_loss'].update(total_loss.item())

        try:
            log_outputs['lr'] = self.scheduler.get_lr()
            if isinstance(log_outputs['lr'], list):
                log_outputs['lr'] = log_outputs['lr'][0]
        except:
            pass

        # log_outputs['all_loss'] = total_loss.item()
        progress_bar_log = log_outputs
        tb_log = {f'tr/{k}': v for k, v in log_outputs.items()}
        return {
            'loss': total_loss,
            'progress_bar': progress_bar_log,
            'log': tb_log
        }

    def optimizer_step(self, epoch, batch_idx, optimizer, optimizer_idx):
        optimizer.step()
        optimizer.zero_grad()
        if self.scheduler is not None:
            self.scheduler.step(self.global_step // hparams['accumulate_grad_batches'])

    def on_epoch_end(self):
        loss_outputs = {k: round(v.avg, 4) for k, v in self.training_losses_meter.items()}
        print(f"\n==============\n "
              f"Epoch {self.current_epoch} ended. Steps: {self.global_step}. {loss_outputs}"
              f"\n==============\n")

    def validation_step(self, sample, batch_idx):
        """

        :param sample:
        :param batch_idx:
        :return: output: dict
        """
        raise NotImplementedError

    def _validation_end(self, outputs):
        """

        :param outputs:
        :return: loss_output: dict
        """
        raise NotImplementedError

    def validation_end(self, outputs):
        loss_output = self._validation_end(outputs)
        print(f"\n==============\n "
              f"valid results: {loss_output}"
              f"\n==============\n")
        return {
            'log': {f'val/{k}': v for k, v in loss_output.items()},
            'val_loss': loss_output['total_loss']
        }

    def build_scheduler(self, optimizer):
        raise NotImplementedError

    def build_optimizer(self, model):
        raise NotImplementedError

    def configure_optimizers(self):
        optm = self.build_optimizer(self.model)
        self.scheduler = self.build_scheduler(optm)
        return [optm]

    def test_start(self):
        pass

    def test_step(self, sample, batch_idx):
        return self.validation_step(sample, batch_idx)

    def test_end(self, outputs):
        return self.validation_end(outputs)

    ###########
    # Running configuration
    ###########

    @classmethod
    def start(cls):
        set_hparams()
        os.environ['MASTER_PORT'] = str(random.randint(15000, 30000))
        random.seed(hparams['seed'])
        np.random.seed(hparams['seed'])
        task = cls()
        work_dir = hparams['work_dir']
        trainer = BaseTrainer(checkpoint_callback=LatestModelCheckpoint(
                                  filepath=work_dir,
                                  verbose=True,
                                  monitor='val_loss',
                                  mode='min',
                                  num_ckpt_keep=hparams['num_ckpt_keep'],
                                  save_best=hparams['save_best'],
                                  period=1 if hparams['save_ckpt'] else 100000
                              ),
                              logger=TensorBoardLogger(
                                  save_dir=work_dir,
                                  name='lightning_logs',
                                  version='lastest'
                              ),
                              gradient_clip_val=hparams['clip_grad_norm'],
                              val_check_interval=hparams['val_check_interval'],
                              row_log_interval=hparams['log_interval'],
                              max_updates=hparams['max_updates'],
                              num_sanity_val_steps=hparams['num_sanity_val_steps'] if not hparams[
                                  'validate'] else 10000,
                              accumulate_grad_batches=hparams['accumulate_grad_batches'])
        if not hparams['infer']:  # train
            # copy_code = input(f'{hparams["save_codes"]} code backup? y/n: ') == 'y'
            # copy_code = True # backup code every time
            # if copy_code:
            #     t = datetime.now().strftime('%Y%m%d%H%M%S')
            #     code_dir = f'{work_dir}/codes/{t}'
            #     # TODO: test filesystem calls
            #     os.makedirs(code_dir, exist_ok=True)
            #     # subprocess.check_call(f'mkdir "{code_dir}"', shell=True)
            #     for c in hparams['save_codes']:
            #         shutil.copytree(c, code_dir, dirs_exist_ok=True)
            #         # subprocess.check_call(f'xcopy "{c}" "{code_dir}/" /s /e /y', shell=True)
            #     print(f"| Copied codes to {code_dir}.")
            trainer.checkpoint_callback.task = task
            trainer.fit(task)
        else:
            trainer.test(task)

    def configure_ddp(self, model, device_ids):
        model = DDP(
            model,
            device_ids=device_ids,
            find_unused_parameters=True
        )
        if dist.get_rank() != 0 and not hparams['debug']:
            sys.stdout = open(os.devnull, "w")
            sys.stderr = open(os.devnull, "w")
        random.seed(hparams['seed'])
        np.random.seed(hparams['seed'])
        return model

    def training_end(self, *args, **kwargs):
        return None

    def init_ddp_connection(self, proc_rank, world_size):
        set_hparams(print_hparams=False)
        # guarantees unique ports across jobs from same grid search
        default_port = 12910
        # if user gave a port number, use that one instead
        try:
            default_port = os.environ['MASTER_PORT']
        except Exception:
            os.environ['MASTER_PORT'] = str(default_port)

        # figure out the root node addr
        root_node = '127.0.0.2'
        root_node = self.trainer.resolve_root_node_address(root_node)
        os.environ['MASTER_ADDR'] = root_node
        dist.init_process_group('nccl', rank=proc_rank, world_size=world_size)

    @data_loader
    def train_dataloader(self):
        return None

    @data_loader
    def test_dataloader(self):
        return None

    @data_loader
    def val_dataloader(self):
        return None

    def on_load_checkpoint(self, checkpoint):
        pass

    def on_save_checkpoint(self, checkpoint):
        pass

    def on_sanity_check_start(self):
        pass

    def on_train_start(self):
        pass

    def on_train_end(self):
        pass

    def on_batch_start(self, batch):
        pass

    def on_batch_end(self):
        pass

    def on_pre_performance_check(self):
        pass

    def on_post_performance_check(self):
        pass

    def on_before_zero_grad(self, optimizer):
        pass

    def on_after_backward(self):
        pass

    def backward(self, loss, optimizer):
        loss.backward()

    def grad_norm(self, norm_type):
        results = {}
        total_norm = 0
        for name, p in self.named_parameters():
            if p.requires_grad:
                try:
                    param_norm = p.grad.data.norm(norm_type)
                    total_norm += param_norm ** norm_type
                    norm = param_norm ** (1 / norm_type)

                    grad = round(norm.data.cpu().numpy().flatten()[0], 3)
                    results['grad_{}_norm_{}'.format(norm_type, name)] = grad
                except Exception:
                    # this param had no grad
                    pass

        total_norm = total_norm ** (1. / norm_type)
        grad = round(total_norm.data.cpu().numpy().flatten()[0], 3)
        results['grad_{}_norm_total'.format(norm_type)] = grad
        return results