PixelRNNs Many-To-One

seq2vec.py

class GenModel(nn.Module):
    def _init_weights(self, module):
        if module.state_dict().get('weight') != None:
          if type(module) == nn.BatchNorm1d:
            nn.init.zeros_(module.bias)
          else:
            nn.init.kaiming_uniform_(module.weight)

    def __init__(self, input_size, hidden_size,  num_layers, bidirectional):
        super().__init__()
 
        self.convPool = nn.Sequential(
          nn.Conv1d(kernel_size=4, 
                    #dilation=0,
                    in_channels=input_size,
                    out_channels=hidden_size,  
                    stride=1), 
          nn.BatchNorm1d(hidden_size),
          nn.ReLU(), 
          nn.AvgPool1d(kernel_size=4),
        )
        self.convPool.apply(self._init_weights)

        self.rnn = nn.LSTM(input_size=hidden_size,
                            hidden_size=hidden_size,
                            num_layers=num_layers,
                            batch_first=True,
                            bidirectional=bidirectional)
        self.rnn.apply(self._init_weights)
        
        # Output of RNN is duplicated automatically if bidirectional is true.
        rnn_units = hidden_size*2 if bidirectional else hidden_size

        self.flatten = nn.Sequential(
            nn.Flatten(),
            nn.BatchNorm1d(rnn_units),
        )

        self.logits  = nn.Sequential(
            nn.Linear(rnn_units, input_size),
            nn.BatchNorm1d(input_size)
        )
        self.logits.apply(self._init_weights)
        self.predictor = nn.Sigmoid()

    def forward(self, *X):
        x = X[0]
        x = torch.permute(x, (0, 2, 1)  )

        y = self.convPool(x)
        y = torch.permute(y, (0, 2, 1)  )

        y, (h, c) = self.rnn(y, (X[1], X[2]) ) if len(X)==3 else self.rnn(y)
        y = y[:,-1,:]

        y = self.flatten(y)

        y_logits = self.logits(y)
        
        y = self.predictor(y_logits)

        return y, y_logits, (h, c)