model.py

import torch
import torch.nn as nn
import torchvision.models as models


class EncoderCNN(nn.Module):
    def __init__(self, embed_size):
        super(EncoderCNN, self).__init__()
        resnet = models.resnet50(pretrained=True)
        for param in resnet.parameters():
            param.requires_grad_(False)
        
        modules = list(resnet.children())[:-1]
        self.resnet = nn.Sequential(*modules)
        self.embed = nn.Linear(resnet.fc.in_features, embed_size)
        self.bn = nn.BatchNorm1d(embed_size, momentum=0.01)       # added Batchnorm

    def forward(self, images):
        features = self.resnet(images)
        features = features.view(features.size(0), -1)
        features = self.embed(features)
        features = self.bn(features)                              # added Batchnorm
        return features
    

class DecoderRNN(nn.Module):
    def __init__(self, embed_size, hidden_size, vocab_size, num_layers=1):
        super(DecoderRNN, self).__init__()
        
        ## Decoder parameters
        self.embed_size = embed_size
        self.hidden_size = hidden_size
        self.vocab_size = vocab_size
        self.num_layers = num_layers
        
        ## Decoder layers definition
        # Embedding each word as a vector in a embedding space of dim embed_size
        self.embedding = nn.Embedding(vocab_size, embed_size)
        
        # we set batch_first True which requires first dimension to be batch_size in forward
        # add drop out in case of multiple LSTM layers
        if num_layers > 1:
            self.lstm = nn.LSTM(embed_size, hidden_size, num_layers,
                            dropout=0.5, batch_first=True)
        else:
            self.lstm = nn.LSTM(embed_size, hidden_size, num_layers, batch_first=True)
        
        # drop out layer
        self.dropout = nn.Dropout(0.5)
        
        # Final layer to output prediction vector accross the vocab space
        self.fc = nn.Linear(hidden_size, vocab_size)
        # Softmax function is included in CrossEntropyLoss function which we use => no softmax included here
        
        # Initialize decoder layers using Xavier's initialization
        self.model_init()
        
        # To provide lisibility, print tensor dimensions when instantiating the model
        self.print_tensor_dimensions = True
    
    def forward(self, features, captions):
        
        ## Decoder forward useful parameters
        self.batch_size = features.shape[0]
        self.seq_length = captions.shape[1]
        
        # Initializing the hidden and cell states and flushing out previous hidden states
        # We don't want the previous batch to influence the output of next image-caption input.
        self.hidden = self.init_hidden(self.batch_size) 
        

        # Remove ending token since we will not require a next word from <end>
        if self.print_tensor_dimensions: print('Captions dimensions before removing <END> token: ', captions.shape)
        captions = captions[:,:-1]  # shape (batch_size, caption_length - 1)
        if self.print_tensor_dimensions: print('Captions dimensions after removing <END> token: ', captions.shape)
        
        # we embed all words into a embed_size space. Each word gets represented by a vector of embed_size
        # Shape of captions - from : batch_size x seq_length (of words)
        # to ==> batch_size x seq_length x word_embedding_dimension (embed_size)
        embeddings = self.embedding(captions)
        if self.print_tensor_dimensions: print('Captions dimensions after embedding : ', embeddings.shape)
        
        # we prepare features tensor into appropriate format: (batchsize x embed_size) ==> (batchsize x 1 x embed_size)
        if self.print_tensor_dimensions: print('features dimensions in: ', features.shape)
        features = features.view(self.batch_size, 1, self.embed_size)    
        if self.print_tensor_dimensions: print('features dim after expanding: ', features.shape)
        
        # We use similar embedding size for the features and the captions so we can concatenate both
        # The feature vector will be sumitted first to the RNN followed by the sequence of words
        # Tensor shape: batchsize x caption_length x embed_size
        embeddings = torch.cat((features, embeddings), dim=1)  
        if self.print_tensor_dimensions: print('Dimensions of combined (features, captions) inputs to LSTM : ', embeddings.shape)
        
        # Input is passed on the the LSTM
        lstm_out, hidden = self.lstm(embeddings, self.hidden)    
        lstm_out = self.dropout(lstm_out)
        
        # To chain multiple LSTM layers
        lstm_out = lstm_out.contiguous()
        
        # We ensure appropriate shape before input to the fully connected and final layer
        # Tensor shape: batchsize x caption_length x embed_size
        lstm_out = lstm_out.view(self.batch_size,self.seq_length, self.hidden_size)  
        if self.print_tensor_dimensions: print('lstm output dimensions: ', lstm_out.shape)
        
        # We submit the output of the LSTM to the fully connected layer for predictions
        outputs = self.fc(lstm_out)
        if self.print_tensor_dimensions: print('Prediction output dimensions: ', outputs.shape)
        
        # Turn off printing dimensions after instantiation
        self.print_tensor_dimensions = False
        
        return outputs
    
    
    def init_hidden(self, batch_size):
        ''' Initializes hidden state '''
        # Create two new tensors with sizes n_layers x batch_size x n_hidden,
        # initialized randomly, for hidden state and cell state of LSTM
                
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        hidden = (torch.randn(self.num_layers, batch_size, self.hidden_size).to(device),
                  torch.randn(self.num_layers, batch_size, self.hidden_size).to(device))
    
        return hidden
    
    
    def model_init(self):
        # We initialize the decoder parameters using Xavier's approach
        torch.nn.init.xavier_uniform_(self.fc.weight)
        torch.nn.init.xavier_uniform_(self.embedding.weight)
        pass


    def sample(self, inputs, states=None, max_len=20):
        " accepts pre-processed image tensor (inputs) and returns predicted sentence (list of tensor ids of length max_len) "
        
        # Make sure input feature is of the expected shape by the LSTM
        inputs = inputs.view(1,1,self.embed_size)
        
        # Initialize the caption output list
        caption=[]
        
        # Initialize the hidden state
        if states is None:
            states = self.init_hidden(1)
        
        # Loop over the max length sequence to produce the caption
        for word in range(max_len):
            
            # Pass the input to the LSTM (first the extracted features then the previous word generated and previous hidden state)
            lstm_out, states = self.lstm(inputs, states)
            lstm_out = lstm_out.contiguous()
            
            # Obtain the prediction from the fully-connected prediction layer (shape 1 x 1 x vocab_size)
            word_pred = self.fc(lstm_out)
            # Extract the index determining the position of the predicted word in vocab
            prediction = word_pred.argmax(dim=2)
            # Append the word index to the caption list
            caption.append(prediction.item())
            # prepare next input to the LSTM : Embed the last word to predict the next
            inputs = self.embedding(prediction.view(1,-1)).view(1,1,-1)
            
        return list(caption)