File size: 6,052 Bytes
0faaa54
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
197
198
199
import re
import torch
import config
import torch.nn as nn
import torch.nn.functional as F
import torchvision.models as models

from collections import OrderedDict


class DenseNet121(nn.Module):
    def __init__(self, out_size=14, checkpoint=None):
        super(DenseNet121, self).__init__()

        self.densenet121 = models.densenet121(weights='DEFAULT')
        num_classes = self.densenet121.classifier.in_features

        self.densenet121.classifier = nn.Sequential(
            nn.Linear(num_classes, out_size),
            nn.Sigmoid()
        )

        if checkpoint is not None:
            device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
            checkpoint = torch.load(checkpoint, map_location=device)

            state_dict = checkpoint['state_dict']
            new_state_dict = OrderedDict()

            for k, v in state_dict.items():
                if 'module' not in k:
                    k = f'module.{k}'
                else:
                    k = k.replace('module.densenet121.features', 'features')
                    k = k.replace('module.densenet121.classifier', 'classifier')
                    k = k.replace('.norm.1', '.norm1')
                    k = k.replace('.conv.1', '.conv1')
                    k = k.replace('.norm.2', '.norm2')
                    k = k.replace('.conv.2', '.conv2')

                new_state_dict[k] = v

            self.densenet121.load_state_dict(new_state_dict)


    def forward(self, x):
        return self.densenet121(x)


class EncoderCNN(nn.Module):
    def __init__(self, checkpoint=None):
        super(EncoderCNN, self).__init__()

        self.model = DenseNet121(
            checkpoint=checkpoint
        )

        for param in self.model.densenet121.parameters():
            param.requires_grad_(False)

    def forward(self, images):
        features = self.model.densenet121.features(images)

        batch, maps, size_1, size_2 = features.size()

        features = features.permute(0, 2, 3, 1)
        features = features.view(batch, size_1 * size_2, maps)

        return features


class Attention(nn.Module):
    def __init__(self, features_size, hidden_size, output_size=1):
        super(Attention, self).__init__()

        self.W = nn.Linear(features_size, hidden_size)
        self.U = nn.Linear(hidden_size, hidden_size)
        self.v = nn.Linear(hidden_size, output_size)

    def forward(self, features, decoder_output):
        decoder_output = decoder_output.unsqueeze(1)

        w = self.W(features)
        u = self.U(decoder_output)

        scores = self.v(torch.tanh(w + u))
        weights = F.softmax(scores, dim=1)
        context = torch.sum(weights * features, dim=1)

        weights = weights.squeeze(2)

        return context, weights


class DecoderRNN(nn.Module):
    def __init__(self, features_size, embed_size, hidden_size, vocab_size):
        super(DecoderRNN, self).__init__()

        self.vocab_size = vocab_size

        self.embedding = nn.Embedding(vocab_size, embed_size)
        self.lstm = nn.LSTMCell(embed_size + features_size, hidden_size)

        self.fc = nn.Linear(hidden_size, vocab_size)

        self.attention = Attention(features_size, hidden_size)

        self.init_h = nn.Linear(features_size, hidden_size)
        self.init_c = nn.Linear(features_size, hidden_size)

    def forward(self, features, captions):
        embeddings = self.embedding(captions)

        h, c = self.init_hidden(features)

        seq_len = len(captions[0]) - 1
        features_size = features.size(1)
        batch_size = captions.size(0)

        outputs = torch.zeros(batch_size, seq_len, self.vocab_size).to(config.DEVICE)
        atten_weights = torch.zeros(batch_size, seq_len, features_size).to(config.DEVICE)

        for i in range(seq_len):
            context, attention = self.attention(features, h)

            inputs = torch.cat((embeddings[:, i, :], context), dim=1)

            h, c = self.lstm(inputs, (h, c))
            h = F.dropout(h, p=0.5)

            output = self.fc(h)

            outputs[:, i, :] = output
            atten_weights[:, i, :] = attention

        return outputs, atten_weights

    def init_hidden(self, features):
        features = torch.mean(features, dim=1)

        h = self.init_h(features)
        c = self.init_c(features)

        return h, c


class EncoderDecoderNet(nn.Module):
    def __init__(self, features_size, embed_size, hidden_size, vocabulary, encoder_checkpoint=None):
        super(EncoderDecoderNet, self).__init__()

        self.vocabulary = vocabulary

        self.encoder = EncoderCNN(
            checkpoint=encoder_checkpoint
        )
        self.decoder = DecoderRNN(
            features_size=features_size,
            embed_size=embed_size,
            hidden_size=hidden_size,
            vocab_size=len(self.vocabulary)
        )

    def forward(self, images, captions):
        features = self.encoder(images)
        outputs, _ = self.decoder(features, captions)

        return outputs

    def generate_caption(self, image, max_length=25):
        caption = []

        with torch.no_grad():
            features = self.encoder(image)
            h, c = self.decoder.init_hidden(features)

            word = torch.tensor(self.vocabulary.stoi['<SOS>']).view(1, -1).to(config.DEVICE)
            embeddings = self.decoder.embedding(word).squeeze(0)

            for _ in range(max_length):
                context, _ = self.decoder.attention(features, h)

                inputs = torch.cat((embeddings, context), dim=1)

                h, c  = self.decoder.lstm(inputs, (h, c))

                output = self.decoder.fc(F.dropout(h, p=0.5))
                output = output.view(1, -1)

                predicted = output.argmax(1)

                if self.vocabulary.itos[predicted.item()] == '<EOS>':
                    break

                caption.append(predicted.item())

                embeddings = self.decoder.embedding(predicted)

        return [self.vocabulary.itos[idx] for idx in caption]