File size: 5,255 Bytes
67f827d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.models as models

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")


class Encoder(nn.Module):
    def __init__(self):
        super(Encoder, self).__init__()
        resnet = models.resnet50(pretrained=True)
        for param in resnet.parameters():
            param.requires_grad_(False)

        modules = list(resnet.children())[:-2]  # extracting the last conv layer from the model
        self.resnet = nn.Sequential(*modules)

    def forward(self, imgs):
        features = self.resnet(imgs)
        features = features.permute(0, 2, 3, 1)  # batch x 7 x 7 x 2048
        features = features.view(features.size(0), -1, features.size(-1))  # batch x 49 x 2048
        return features


class Attention(nn.Module):
    def __init__(self, encoder_dims, decoder_dims, attention_dims):
        super(Attention, self).__init__()
        self.attention_dims = attention_dims  # size of attention network
        self.U = nn.Linear(encoder_dims, attention_dims)  # a^(t)
        self.W = nn.Linear(decoder_dims, attention_dims)  # s^(t` - 1)
        self.A = nn.Linear(attention_dims, 1)  # cvt the attention dims back to 1

    def forward(self, features, hidden):
        u_as = self.U(features)
        w_as = self.W(hidden)
        combined_state = torch.tanh(u_as + w_as.unsqueeze(1))
        attention_score = self.A(combined_state)
        attention_score = attention_score.squeeze(2)
        alpha = F.softmax(attention_score, dim=1)
        attention_weights = features * alpha.unsqueeze(2)  # batch x num_timesteps (49) x features
        attention_weights = attention_weights.sum(dim=1)
        return alpha, attention_weights


class Decoder(nn.Module):
    def __init__(self, embed_size, vocab_size, attention_dim, encoder_dim, decoder_dim, fc_dims, p=0.3,
                 embeddings=None):
        super().__init__()

        self.vocab_size = vocab_size
        self.attention_dim = attention_dim
        self.decoder_dim = decoder_dim

        self.embedding = nn.Embedding(vocab_size, embedding_dim=embed_size)
        self.attention = Attention(encoder_dim, decoder_dim, attention_dim)

        self.init_h = nn.Linear(encoder_dim, decoder_dim)
        self.init_c = nn.Linear(encoder_dim, decoder_dim)
        self.lstm = nn.LSTMCell(encoder_dim + embed_size, decoder_dim, bias=True)
        self.fcn1 = nn.Linear(decoder_dim, vocab_size)
        self.fcn2 = nn.Linear(fc_dims, vocab_size)
        self.drop = nn.Dropout(p)

        if embeddings is not None:
            self.load_pretrained_embed(embeddings)

    def forward(self, features, captions):

        seq_length = len(captions[0]) - 1  # Exclude the last one
        batch_size = captions.size(0)
        num_timesteps = features.size(1)

        embed = self.embedding(captions)
        h, c = self.init_hidden_state(features)  # initialize h and c for LSTM

        preds = torch.zeros(batch_size, seq_length, self.vocab_size).to(device)
        alphas = torch.zeros(batch_size, seq_length, num_timesteps).to(device)

        for s in range(seq_length):
            alpha, context = self.attention(features, h)
            lstm_inp = torch.cat((embed[:, s], context), dim=1)
            h, c = self.lstm(lstm_inp, (h, c))
            out = self.drop(self.fcn1(h))
            preds[:, s] = out
            alphas[:, s] = alpha

        return preds, alphas

    def gen_captions(self, features, max_len=20, vocab=None):
        h, c = self.init_hidden_state(features)
        alphas = []
        captions = []
        word = torch.tensor(vocab.stoi["<SOS>"]).view(1, -1).to(device)
        embed = self.embedding(word)
        for i in range(max_len):
            alpha, context = self.attention(features, h)
            alphas.append(alpha.cpu().detach().numpy())

            lstm_inp = torch.cat((embed[:, 0], context), dim=1)
            h, c = self.lstm(lstm_inp, (h, c))
            out = self.drop(self.fcn1(h))
            word_out_idx = torch.argmax(out, dim=1)
            captions.append(word_out_idx.item())
            if vocab.itos[word_out_idx.item()] == "<EOS>":
                break
            embed = self.embedding(word_out_idx.unsqueeze(0))

        return [vocab.itos[word] for word in captions], alphas

    def load_pretrained_embed(self, embeddings):
        self.embedding.weight = nn.Parameter(embeddings)
        for p in self.embedding.parameters():
            p.requires_grad = True

    def init_hidden_state(self, encoder_output):
        mean_encoder_out = encoder_output.mean(dim=1)
        h = self.init_h(mean_encoder_out)
        c = self.init_c(mean_encoder_out)
        return h, c


class EncoderDecoder(nn.Module):
    def __init__(self, embed_size, vocab_size, attention_dim, encoder_dim, decoder_dim, fc_dims, p=0.3,
                 embeddings=None):
        super().__init__()
        self.EncoderCNN = Encoder()
        self.DecoderLSTM = Decoder(embed_size, vocab_size, attention_dim, encoder_dim, decoder_dim, fc_dims, p,
                                   embeddings)

    def forward(self, imgs, caps):
        features = self.EncoderCNN(imgs)
        out = self.DecoderLSTM(features, caps)
        return out