File size: 6,469 Bytes
95456ed
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import numpy as np
import transformers

from typing import Optional


class EarlyStopping:
    def __init__(

        self,

        patience: int,

        path: str,

        delta: Optional[int] = 0,

    ):
        self.patience = patience
        self.delta = delta
        self.best_score = None
        self.early_stop = False
        self.counter = 0
        self.best_loss = np.inf
        self.path = path

    def __call__(

        self,

        val_loss,

        model,

    ):
        score = -val_loss

        if self.best_score is None:
            self.best_score = score
            self.save_checkpoint(val_loss, model)
        elif score < self.best_score + self.delta:
            self.counter += 1
            print(f"EarlyStopping counter: {self.counter} out of {self.patience}")
            if self.counter >= self.patience:
                self.early_stop = True
        else:
            self.best_score = score
            self.save_checkpoint(val_loss, model)
            self.counter = 0

    def save_checkpoint(

        self,

        val_loss,

        model,

    ):
        """Saves model when validation loss decreases."""
        print(
            f"Validation loss decreased ({self.best_loss:.6f} --> {val_loss:.6f}). Saving model..."
        )
        torch.save(model.state_dict(), self.path)
        self.best_loss = val_loss


def train(

    model,

    num_epochs: int,

    train_loader: torch.utils.data.DataLoader,

    val_loader: torch.utils.data.DataLoader,

    criterion: nn.MSELoss,

    optimizer: transformers.AdamW,

    early_stopping: EarlyStopping,

    scheduler: transformers.get_linear_schedule_with_warmup,

    device: torch.device,

    fix_dur_colname: str,

    output_attentions: Optional[bool] = None,

    use_attention_mask: Optional[bool] = None,

):
    """

    Train loop to train the Seq2Seq model.

    :param model: the model to train

    :param num_epochs: number of epochs to train

    :param train_loader: the training data loader

    :param val_loader: the validation data loader

    :param criterion: the loss function (MSE Loss)

    :param optimizer: the optimizer (AdamW)

    :param early_stopping: the early stopping object

    :param scheduler: the learning rate scheduler

    :param device: the device to train on

    :param fix_dur_colname: the name of the column containing the fixations durations

    """
    for epoch in range(num_epochs):

        model.train()

        for batch_idx, train_batch in enumerate(train_loader):

            optimizer.zero_grad()

            sp_embeddings = train_batch["sp_embeddings"].to(device)
            attention_mask = train_batch["attention_masks"].to(device)
            fix_durs = train_batch[fix_dur_colname].to(device)

            # forward pass
            if use_attention_mask:

                if output_attentions:

                    out, _ = model(
                        sp_embeddings=sp_embeddings,
                        attention_mask=attention_mask,
                        output_attentions=output_attentions,
                    )
                else:

                    out = model(
                        sp_embeddings=sp_embeddings,
                        attention_mask=attention_mask,
                        output_attentions=output_attentions,
                    )
            else:

                if output_attentions:

                    out, _ = model(
                        sp_embeddings=sp_embeddings,
                        output_attentions=output_attentions,
                    )
                else:
                    out = model(
                        sp_embeddings=sp_embeddings,
                        output_attentions=output_attentions,
                    )

            # train_loss = criterion(out, fix_durs)
            # mask the padding in the loss computation
            loss_mask = (fix_durs != 0).float()
            # train_loss = criterion(out * loss_mask, fix_durs * loss_mask)
            train_loss = criterion(out, fix_durs)

            train_loss.backward()
            optimizer.step()
            scheduler.step()

            print(f"\t  epoch {epoch+1}, batch {batch_idx+1}, loss: {train_loss.item():.4f}")

        # validation

        model.eval()
        val_loss = 0.0

        with torch.no_grad():

            for val_batch in val_loader:

                sp_embeddings = val_batch["sp_embeddings"].to(device)
                attention_mask = val_batch["attention_masks"].to(device)
                fix_durs = val_batch["fix_durs"].to(device)

                if use_attention_mask:

                    if output_attentions:
                        out, attentions = model(
                            sp_embeddings=sp_embeddings,
                            attention_mask=attention_mask,
                            output_attentions=output_attentions,
                        )
                    else:
                        out = model(
                            sp_embeddings=sp_embeddings,
                            attention_mask=attention_mask,
                            output_attentions=output_attentions,
                        )
                else:

                    # forward pass
                    if output_attentions:
                        out, attentions = model(
                            sp_embeddings=sp_embeddings,
                            output_attentions=output_attentions,
                        )

                    else:
                        out = model(
                            sp_embeddings=sp_embeddings,
                            output_attentions=output_attentions,
                        )

                val_loss_mask = (fix_durs != 0).float()
                # val_loss += criterion(out * val_loss_mask, fix_durs * val_loss_mask).item()
                val_loss += criterion(out, fix_durs).item()

        # average the losses
        val_loss /= len(val_loader)
        train_loss /= len(train_loader)

        print(f"Epoch {epoch+1}, Train Loss: {train_loss:.4f}, Val Loss: {val_loss:.4f}")

        # check for early stopping
        early_stopping(val_loss, model)
        if early_stopping.early_stop:
            print("Early stopping")
            break