File size: 6,006 Bytes
b72a776
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
from abc import abstractmethod
from numpy import inf
from logger import TensorboardWriter


class BaseTrainer:
    """
    Base class for all trainers
    """
    def __init__(self, model, criterion, metric_ftns, optimizer, config):
        self.config = config
        self.logger = config.get_logger('trainer', config['trainer']['verbosity'])

        self.model = model
        self.criterion = criterion
        self.metric_ftns = metric_ftns
        self.optimizer = optimizer

        cfg_trainer = config['trainer']
        self.epochs = cfg_trainer['epochs']
        self.save_period = cfg_trainer['save_period']
        self.monitor = cfg_trainer.get('monitor', 'off')

        # configuration to monitor model performance and save best
        if self.monitor == 'off':
            self.mnt_mode = 'off'
            self.mnt_best = 0
        else:
            self.mnt_mode, self.mnt_metric = self.monitor.split()
            assert self.mnt_mode in ['min', 'max']

            self.mnt_best = inf if self.mnt_mode == 'min' else -inf
            self.early_stop = cfg_trainer.get('early_stop', inf)
            if self.early_stop <= 0:
                self.early_stop = inf

        self.start_epoch = 1

        self.checkpoint_dir = config.save_dir

        # setup visualization writer instance                
        self.writer = TensorboardWriter(config.log_dir, self.logger, cfg_trainer['tensorboard'])

        if config.resume is not None:
            self._resume_checkpoint(config.resume)

    @abstractmethod
    def _train_epoch(self, epoch):
        """
        Training logic for an epoch

        :param epoch: Current epoch number
        """
        raise NotImplementedError

    def train(self):
        """
        Full training logic
        """
        not_improved_count = 0
        for epoch in range(self.start_epoch, self.epochs + 1):
            result = self._train_epoch(epoch)

            # save logged informations into log dict
            log = {'epoch': epoch}
            log.update(result)

            # print logged informations to the screen
            for key, value in log.items():
                self.logger.info('    {:15s}: {}'.format(str(key), value))

            # evaluate model performance according to configured metric, save best checkpoint as model_best
            best = False
            if self.mnt_mode != 'off':
                try:
                    # check whether model performance improved or not, according to specified metric(mnt_metric)
                    improved = (self.mnt_mode == 'min' and log[self.mnt_metric] <= self.mnt_best) or \
                               (self.mnt_mode == 'max' and log[self.mnt_metric] >= self.mnt_best)
                except KeyError:
                    self.logger.warning("Warning: Metric '{}' is not found. "
                                        "Model performance monitoring is disabled.".format(self.mnt_metric))
                    self.mnt_mode = 'off'
                    improved = False

                if improved:
                    self.mnt_best = log[self.mnt_metric]
                    not_improved_count = 0
                    best = True
                else:
                    not_improved_count += 1

                if not_improved_count > self.early_stop:
                    self.logger.info("Validation performance didn\'t improve for {} epochs. "
                                     "Training stops.".format(self.early_stop))
                    break

            if epoch % self.save_period == 0:
                self._save_checkpoint(epoch, save_best=best)

    def _save_checkpoint(self, epoch, save_best=False):
        """
        Saving checkpoints

        :param epoch: current epoch number
        :param log: logging information of the epoch
        :param save_best: if True, rename the saved checkpoint to 'model_best.pth'
        """
        arch = type(self.model).__name__
        state = {
            'arch': arch,
            'epoch': epoch,
            'state_dict': self.model.state_dict(),
            'optimizer': self.optimizer.state_dict(),
            'monitor_best': self.mnt_best,
            'config': self.config
        }
        filename = str(self.checkpoint_dir / 'checkpoint-epoch{}.pth'.format(epoch))
        torch.save(state, filename)
        self.logger.info("Saving checkpoint: {} ...".format(filename))
        if save_best:
            best_path = str(self.checkpoint_dir / 'model_best.pth')
            torch.save(state, best_path)
            self.logger.info("Saving current best: model_best.pth ...")

    def _resume_checkpoint(self, resume_path):
        """
        Resume from saved checkpoints

        :param resume_path: Checkpoint path to be resumed
        """
        resume_path = str(resume_path)
        self.logger.info("Loading checkpoint: {} ...".format(resume_path))
        checkpoint = torch.load(resume_path)
        self.start_epoch = checkpoint['epoch'] + 1
        self.mnt_best = checkpoint['monitor_best']

        # load architecture params from checkpoint.
        if checkpoint['config']['arch'] != self.config['arch']:
            self.logger.warning("Warning: Architecture configuration given in config file is different from that of "
                                "checkpoint. This may yield an exception while state_dict is being loaded.")
        self.model.load_state_dict(checkpoint['state_dict'])

        # load optimizer state from checkpoint only when optimizer type is not changed.
        if checkpoint['config']['optimizer']['type'] != self.config['optimizer']['type']:
            self.logger.warning("Warning: Optimizer type given in config file is different from that of checkpoint. "
                                "Optimizer parameters not being resumed.")
        else:
            self.optimizer.load_state_dict(checkpoint['optimizer'])

        self.logger.info("Checkpoint loaded. Resume training from epoch {}".format(self.start_epoch))