convolution.py 文件源码

def forward_conv_batch(self, x):
        """
        :param x: (batch, length, dim)
        :return:  (batch, length - kernel + 2*padding_size + 1, hidden_dim)
        """
        # T.nn.conv2d (batch size, input channels, input rows, input columns)
        # dl4nlp      (batch size, 1,              length,     in_dim)
        x = x.dimshuffle([0, 'x', 1, 2])
        # T.nn.conv2d (output channels, input channels, filter rows, filter columns)
        # dl4nlp      (hidden_dim,      1,              kernel_size, in_dim)
        filter_w = self.W.dimshuffle([1, 'x', 0, 2])
        # T.nn.conv2d (batch size, output channels, output rows,     output columns)
        # dl4nlp      (batch size, hidden_dim,      length+kernel-1, 1)
        conv_result = T.nnet.conv2d(x, filter_w,
                                    border_mode='valid',)
        # from theano.printing import Print
        # conv_result = Print()(conv_result)
        # (batch size, hidden_dim, length - kernel + 2*padding_size + 1, 1)
        #   -> (batch, length - kernel + 2*padding_size + 1, hidden_dim)
        conv_result = T.transpose(conv_result[:, :, :, 0], (0, 2, 1))
        return conv_result