vit/model.py at master · iolkhovsky/vit · 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
import numpy as np
import pytorch_lightning as pl
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision
from typing import Any, Dict

from core import VisionTransformer
from utils import visualize_prediction


def compute_accuracy(logits, targets):
    _, predicted = torch.max(logits, 1)
    correct = (predicted == targets).sum().item()
    accuracy = correct / targets.size(0)
    return accuracy


def tensor_to_numpy(tensor):
    if tensor.device != torch.device('cpu'):
        tensor = tensor.cpu()
    return tensor.numpy()


class VitClassifier(pl.LightningModule):
    def __init__(self, num_classes=10, labels_map=None, *args, **kwargs):
        super().__init__()
        self._core = VisionTransformer(num_classes=num_classes)
        self._criterion = nn.CrossEntropyLoss()
        self._labels_map = labels_map

    @staticmethod
    def preprocess(images_batch, resize=(72, 72)):
        assert len(images_batch.shape) == 4
        if isinstance(images_batch, np.ndarray):
            images_batch = torch.from_numpy(images_batch)
        assert torch.is_tensor(images_batch)
        if images_batch.shape[1] != 3 and images_batch.shape[3] == 3:
            images_batch = torch.permute(images_batch, (0, 3, 1, 2))
        images_batch = images_batch.float() / 255.
        images_batch = torchvision.transforms.Normalize(
            mean=[0.485, 0.456, 0.406],
            std=[0.229, 0.224, 0.225]
        )(images_batch)
        if resize is not None:
            images_batch = torchvision.transforms.Resize(resize)(images_batch)
        return images_batch

    def forward(self, images_batch, return_scores=False):
        images_batch = VitClassifier.preprocess(images_batch)
        attn_scores = None
        if return_scores:
            logits, attn_scores = self._core(images_batch, return_scores=True)
        else:
            logits = self._core(images_batch)

        if not self.training:
            scores = torch.softmax(logits, dim=-1)
            max_scores, max_indices = torch.max(scores, dim=1)
            if return_scores:
                return max_scores, max_indices, attn_scores
            else:
                return max_scores, max_indices
        else:
            if return_scores:
                return logits, attn_scores
            else:
                return logits

    def configure_optimizers(self):
        optimizer = optim.Adam(self.parameters(), lr=0.001)
        scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.95)
        return [[optimizer], [scheduler]]

    def training_step(self, batch: Dict[str, Any], batch_idx: int) -> torch.Tensor:
        images, labels = batch
        logits = self.forward(images)
        loss = self._criterion(logits, labels)
        acc = compute_accuracy(logits, labels)
        self.log_dict(
            {
                f"loss/train": loss.detach(),
                f"accuracy/train": acc,
            }
        )
        return loss

    def validation_step(self, batch: Dict[str, Any], batch_idx: int) -> None:
        if batch_idx == 0:
            images, labels = batch
            cls_scores, classes, attn_scores = self.forward(images, return_scores=True)
            accuracy = (classes == labels).float().mean()
            self.log_dict(
                {
                    f"accuracy/val": accuracy,
                }
            )
            self.visualize(
                images=images,
                targets=labels,
                pred_labels=classes,
                pred_scores=cls_scores,
                attn_map=attn_scores,
            )

    def visualize(self, images, targets, pred_labels, pred_scores, attn_map):
        predictions, inputs = [], []
        for image, target, pred_label, score, map in \
                zip(images, targets, pred_labels, pred_scores, attn_map):
            image = tensor_to_numpy(image).astype(np.uint8)
            score_str = '{:.3f}'.format(tensor_to_numpy(score))
            target_label = int(tensor_to_numpy(target))
            pred_label = int(tensor_to_numpy(pred_label))
            if self._labels_map is not None:
                target_label = self._labels_map[target_label]
                pred_label = self._labels_map[pred_label]
            text = f'Target: {target_label}'
            text += f'\nPrediction: {pred_label} Score: {score_str}'
            attn = tensor_to_numpy(map)
            predictions.append(
                visualize_prediction(
                    img=image,
                    text=text,
                    attention=attn,
                )
            )
            inputs.append(
                visualize_prediction(
                    img=image,
                    text=text,
                )
            )

        writer = self.logger.experiment

        predictions_tensors = [torch.permute(torch.from_numpy(x), (2, 0, 1)) for x in predictions]
        predictions_grid = torchvision.utils.make_grid(predictions_tensors)
        writer.add_image(f'Predictions', predictions_grid, self.global_step)

        inputs_tensors = [torch.permute(torch.from_numpy(x), (2, 0, 1)) for x in inputs]
        inputs_grid = torchvision.utils.make_grid(inputs_tensors)
        writer.add_image(f'Inputs', inputs_grid, self.global_step)