SimTA/engine/base_engine.py at main · HINTLab/SimTA · GitHub

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
import os
import shutil
from datetime import datetime

import numpy as np
import torch
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm


class BaseEngine:

    def __init__(self, cfg, log_dir):
        self.cfg = cfg
        start_time = datetime.now().strftime("%Y%m%d-%H%M%S")
        self.log_path = os.path.join(log_dir, start_time)
        self.cfg_path = os.path.join(self.log_path, "config")
        self.history = {"train": {}, "val": {}}

    def compile(self, model, optimizer, loss_fn, metrics=[]):
        self.model = model
        self.optimizer = optimizer
        self.loss_fn = loss_fn
        self.metrics = metrics

    def forward(self, sample):
        raise NotImplementedError("forward method not implemented.")

    def backward(self, loss):
        loss.backward()
        self.optimizer.step()
        if self.scheduler is not None:
            self.scheduler.step()

    def _metrics_epoch_begin(self):
        for i in range(len(self.metrics)):
            self.metrics[i].on_epoch_begin()

    def _metrics_batch_end(self, output, target):
        for i in range(len(self.metrics)):
            self.metrics[i].on_batch_end(output, target)

    def _metrics_epoch_end(self):
        for i in range(len(self.metrics)):
            self.metrics[i].on_epoch_end()

    def _log_history(self, phase, loss):
        self.history[phase]["loss"] = self.history[phase].get("loss", [])
        self.history[phase]["loss"].append(loss)
        for i in range(len(self.metrics)):
            self.history[phase][self.metrics[i].name] =\
                self.history[phase].get(self.metrics[i].name, [])
            self.history[phase][self.metrics[i].name]\
                .append(self.metrics[i].value)

    def _update_tensorboard(self):
        for k in self.history["train"].keys():
            self.tb_writer.add_scalars(k, {"train":
                self.history["train"][k][-1],
                "val": self.history["val"][k][-1]},
                self.cur_epoch)

        self.tb_writer.add_scalars("loss", {"train":
            self.history["train"]["loss"][-1], "val":
            self.history["val"]["loss"][-1]}, self.cur_epoch)

        self.tb_writer.flush()

    def _train_epoch(self, data):
        self.model.train()
        epoch_loss = 0
        data_len = 0
        self._metrics_epoch_begin()
        for _, sample in enumerate(data):
            self.optimizer.zero_grad()
            output, target, loss = self.forward(sample)
            self.backward(loss)

            batch_loss = loss.detach().item()
            epoch_loss += batch_loss * output.size(0)
            data_len += output.size(0)

            self._metrics_batch_end(output, target)

        epoch_loss /= data_len
        self._metrics_epoch_end()
        self._log_history("train", epoch_loss)

    def evaluate(self, data):
        self.model.eval()
        total_loss = 0
        data_len = 0
        self._metrics_epoch_begin()

        with torch.set_grad_enabled(False):
            for _, sample in enumerate(data):
                output, target, loss = self.forward(sample)

                batch_loss = loss.detach().item()
                total_loss += batch_loss * output.size(0)
                data_len += output.size(0)
                self._metrics_batch_end(output, target)

            total_loss /= data_len

        self._metrics_epoch_end()

        return total_loss, {self.metrics[i].name:
            self.metrics[i].value for i in range(len(self.metrics))}

    def _eval_epoch(self, data):
        epoch_loss, _ = self.evaluate(data)
        self._log_history("val", epoch_loss)

    def train(self, data_train, epochs, data_val=None, scheduler=None):
        if not os.path.exists(self.log_path):
            os.mkdir(self.log_path)
        self.tb_writer = SummaryWriter(self.log_path)
        if not os.path.exists(self.cfg_path):
            os.mkdir(self.cfg_path)
        shutil.copy(self.cfg.__file__, self.cfg_path)

        self.total_epochs = epochs
        self.cur_epoch = 0
        if scheduler is not None:
            self.scheduler = scheduler

        progress = tqdm(total=self.total_epochs)
        for i in range(self.total_epochs):
            progress.set_description("Epoch {}/{}".format(i + 1,
                self.total_epochs))
            self._train_epoch(data_train)

            if data_val is not None:
                self._eval_epoch(data_val)

            self._update_tensorboard()
            progress.update(1)
            self.cur_epoch += 1

    def get_history(self):
        return self.history

    def predict(self, data):
        self.model.eval()
        y_pred = []
        y_prob = []

        with torch.set_grad_enabled(False):
            for _, sample in tqdm(enumerate(data), total=len(data)):
                output, _, _ = self.forward(sample)
                y_prob.append(output.detach().cpu().numpy())
                y_pred.append(np.argmax(output.detach().cpu().numpy(),
                    axis=-1))

        y_pred = np.concatenate(y_pred, axis=0)
        y_prob = np.concatenate(y_prob, axis=0)

        return y_pred, y_prob