annotated-transformer-kr/trainer.py at master · simonjisu/annotated-transformer-kr · 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
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
import torch
from pathlib import Path
import numpy as np
from transformer.utils import get_pos
from tqdm import tqdm

class Trainer(object):
    """model trainer: torchtext"""
    def __init__(self, optimizer, train_loader, test_loader, n_step, device, save_path,
                 enc_sos_idx=None, enc_eos_idx=None, dec_sos_idx=None, dec_eos_idx=None, metrics_method="acc", verbose=0):
        if not Path("trainlog/").exists():
            Path("trainlog").mkdir()
        self.save_path = save_path
        self.record_path = Path("trainlog/")/("train-log-"+ save_path.split("/")[-1].split('.')[0] +".txt")
        if Path(self.record_path).exists():
            Path(self.record_path).unlink()

        self.n_step = n_step
        self.device = device
        self.enc_sos_idx = enc_sos_idx
        self.enc_eos_idx = enc_eos_idx
        self.dec_sos_idx = dec_sos_idx
        self.dec_eos_idx = dec_eos_idx

        self.optimizer = optimizer
        self.train_loader = train_loader
        self.test_loader = test_loader
        self.metrics_method = metrics_method
        self.verbose = verbose
        self.tocpu = lambda x: x.cpu()

    def main(self, model, loss_function, rt_losses=False):
        import time

        start_time = time.time()

        train_losses = []
        test_losses = []
        train_accs = []
        test_accs = []
        if self.metrics_method == "acc":
            lowest_metrics = 0.0
        elif self.metrics_method == "loss":
            lowest_metrics = 999
        else:
            assert False, "metrics_method = acc or loss"

        for step in range(1, self.n_step+1):
            train_loss, train_acc = self.train(model, loss_function, step)
            test_loss, test_acc = self.test(model, loss_function)

            train_losses.append(train_loss)
            train_accs.append(train_acc)
            test_losses.append(test_loss)
            test_accs.append(test_acc)

            self._print(step, train_loss, test_loss, train_acc, test_acc)
            if self.metrics_method == "acc":
                test_metrics, test_metrics_list = test_acc, test_accs
            elif self.metrics_method == "loss":
                test_metrics, test_metrics_list = test_loss, test_losses
            lowest_metrics, early_break = self.save_model(model, test_metrics, test_metrics_list, lowest_metrics)
            if early_break:
                print(" - early break!!")
                break

        # Save Log for drawing graph
        np.savetxt("trainlog/losses-accs-" + self.save_path.split("/")[-1].split('.')[0] +".txt",
                   np.array([train_losses, train_accs, test_losses, test_accs]), fmt="%.4e")
        # Time
        end_time = time.time()
        total_time = end_time-start_time
        hour = int(total_time // (60*60))
        minute = int((total_time - hour*60*60) // 60)
        second = total_time - hour*60*60 - minute*60
        txt = f"\nTraining Excution time with validation: {hour:d} h {minute:d} m {second:.4f} s"
        self._print_record(txt)

    def train(self, model, loss_function, step):
        """train model"""
        model.train()
        train_loss = 0
        n_correct = 0
        n_word = 0

        # setting iterator by verbose
        if self.verbose == 0:
            iterator = self.train_loader
        elif self.verbose == 1:
            iterator = tqdm(self.train_loader, desc=f"Training: {step}", total=len(self.train_loader))
        else:
            assert False, "set verbose 0 or 1"

        for batch in iterator:
            src, trg = batch.src, batch.trg
            batch_size = src.size(0)

            src_pos = get_pos(src, model.pad_idx, self.enc_sos_idx, self.enc_eos_idx)
            trg_pos = get_pos(trg, model.pad_idx, self.dec_sos_idx, self.dec_eos_idx)

            self.optimizer.zero_grad()
            output = model(src, src_pos, trg, trg_pos)

            real_trg = trg[:, 1:].contiguous()
            loss = loss_function(output, real_trg)
            loss.backward()
            self.optimizer.step()
            # record
            train_loss += loss.item()
            pred = self.tocpu(output).view(-1, output.size(-1))
            n_correct += (pred.argmax(-1) == self.tocpu(real_trg).view(-1)).sum().item()
            n_word += trg.ne(model.pad_idx).sum().item()
            train_acc = n_correct / n_word

        return train_loss / n_word, train_acc

    def test(self, model, loss_function):
        """test model"""
        model.eval()
        test_loss = 0
        n_correct = 0
        n_word = 0

        with torch.no_grad():
            for batch in self.test_loader:
                src, trg = batch.src, batch.trg

                src_pos = get_pos(src, model.pad_idx, self.enc_sos_idx, self.enc_eos_idx)
                trg_pos = get_pos(trg, model.pad_idx, self.dec_sos_idx, self.dec_eos_idx)

                output = model(src, src_pos, trg, trg_pos)

                real_trg = trg[:, 1:].contiguous()
                loss = loss_function(output, real_trg)

                # record
                test_loss += loss.item()
                pred = self.tocpu(output).view(-1, output.size(-1))
                n_correct += (pred.argmax(-1) == self.tocpu(real_trg).view(-1)).sum().item()
                n_word += trg.ne(model.pad_idx).sum().item()
                test_acc = n_correct / n_word


        return test_loss / n_word, test_acc

    def save_model(self, model, test_metrics, test_metrics_list, lowest_metrics):
        """early stopping"""
        early_break = False
        if len(test_metrics_list) >= 2:
            if self.metrics_method == "acc":
                if test_metrics >= lowest_metrics:
                    torch.save(model.state_dict(), self.save_path)
                    lowest_metrics = test_metrics
                    self._print_record("  - discard previous state, best model state saved!")
                if lowest_metrics == 1.0:
                    early_break = True

            elif self.metrics_method == "loss":
                if test_metrics <= lowest_metrics:
                    torch.save(model.state_dict(), self.save_path)
                    lowest_metrics = test_metrics
                    self._print_record("  - discard previous state, best model state saved!")
                if lowest_metrics == 0.0:
                    # TODO: change early break for 'loss' method
                    early_break = True

        return lowest_metrics, early_break

    def _print(self, step, train_loss, test_loss, train_acc, test_acc):
        """print log"""
        print_list = []
        print_list += [f"[{step}/{self.n_step}] Training Result"]
        print_list += [f"  - train_acc: {train_acc*100:.2f}%\t test_acc: {test_acc*100:.2f}%"]
        print_list += [f"  - train_ppl: {np.exp(min(train_loss, 100)):8.4f}\t test_ppl: {np.exp(min(test_loss, 100)):8.4f}"]
        for txt in print_list:
            self._print_record(txt)


    def _print_record(self, txt):
        print(txt)
        with open(self.record_path, "a") as config:
            config.write(txt+"\n")