layers.py 文件源码

def set_output(self):
        if sum(self._padding) > 0:
            padded_input = tensor.alloc(0.0,  # Value to fill the tensor
                                        self._input_shape[0],
                                        self._input_shape[1],
                                        self._input_shape[2] + 2 * self._padding[2],
                                        self._input_shape[3] + 2 * self._padding[3])

            padded_input = tensor.set_subtensor(
                padded_input[:, :, self._padding[2]:self._padding[2] + self._input_shape[2],
                             self._padding[3]:self._padding[3] + self._input_shape[3]],
                self._prev_layer.output)

            padded_input_shape = [self._input_shape[0], self._input_shape[1],
                                  self._input_shape[2] + 2 * self._padding[2],
                                  self._input_shape[3] + 2 * self._padding[3]]
        else:
            padded_input = self._prev_layer.output
            padded_input_shape = self._input_shape

        conv_out = conv.conv2d(
            input=padded_input,
            filters=self.W.val,
            filter_shape=self._filter_shape,
            image_shape=np.asarray(
                padded_input_shape, dtype=np.int16),
            border_mode='valid')

        # add the bias term. Since the bias is a vector (1D array), we first
        # reshape it to a tensor of shape (1, n_filters, 1, 1). Each bias will
        # thus be broadcasted across mini-batches and feature map
        # width & height
        self._output = conv_out + self.b.val.dimshuffle('x', 0, 'x', 'x')