keras/backend/theano_backend.py [2228:2258]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    image_shape = _preprocess_conv2d_image_shape(int_shape(x), data_format)
    depthwise_kernel_shape = int_shape(depthwise_kernel)
    if depthwise_kernel_shape is None:
        depthwise_kernel_shape = depthwise_kernel.eval().shape  # in case of a shared variable
    depthwise_kernel_shape = _preprocess_conv2d_depthwise_filter_shape(depthwise_kernel_shape, data_format)
    pointwise_kernel_shape = int_shape(pointwise_kernel)
    if pointwise_kernel_shape is None:
        pointwise_kernel_shape = pointwise_kernel.eval().shape  # in case of a shared variable
    pointwise_kernel_shape = _preprocess_conv2d_filter_shape(pointwise_kernel_shape, data_format)

    x = _preprocess_conv2d_input(x, data_format)
    depthwise_kernel = _preprocess_conv2d_depthwise_kernel(depthwise_kernel, depthwise_kernel_shape, data_format)
    pointwise_kernel = _preprocess_conv2d_kernel(pointwise_kernel, data_format)
    th_padding = _preprocess_padding(padding)

    conv_out = T.nnet.conv2d(x, depthwise_kernel,
                             border_mode=th_padding,
                             subsample=strides,
                             input_shape=image_shape,
                             filter_shape=depthwise_kernel_shape,
                             filter_dilation=dilation_rate,
                             num_groups=image_shape[1])
    conv_out = T.nnet.conv2d(conv_out, pointwise_kernel,
                             border_mode=th_padding,
                             subsample=(1, 1),
                             input_shape=None,
                             filter_shape=pointwise_kernel_shape,
                             filter_dilation=dilation_rate)
    conv_out = _postprocess_conv2d_output(conv_out, x, padding,
                                          pointwise_kernel_shape,
                                          strides, data_format)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



keras/backend/theano_backend.py [2285:2315]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    image_shape = _preprocess_conv2d_image_shape(int_shape(x), data_format)
    depthwise_kernel_shape = int_shape(depthwise_kernel)
    if depthwise_kernel_shape is None:
        depthwise_kernel_shape = depthwise_kernel.eval().shape  # in case of a shared variable
    depthwise_kernel_shape = _preprocess_conv2d_depthwise_filter_shape(depthwise_kernel_shape, data_format)
    pointwise_kernel_shape = int_shape(pointwise_kernel)
    if pointwise_kernel_shape is None:
        pointwise_kernel_shape = pointwise_kernel.eval().shape  # in case of a shared variable
    pointwise_kernel_shape = _preprocess_conv2d_filter_shape(pointwise_kernel_shape, data_format)

    x = _preprocess_conv2d_input(x, data_format)
    depthwise_kernel = _preprocess_conv2d_depthwise_kernel(depthwise_kernel, depthwise_kernel_shape, data_format)
    pointwise_kernel = _preprocess_conv2d_kernel(pointwise_kernel, data_format)
    th_padding = _preprocess_padding(padding)

    conv_out = T.nnet.conv2d(x, depthwise_kernel,
                             border_mode=th_padding,
                             subsample=strides,
                             input_shape=image_shape,
                             filter_shape=depthwise_kernel_shape,
                             filter_dilation=dilation_rate,
                             num_groups=image_shape[1])
    conv_out = T.nnet.conv2d(conv_out, pointwise_kernel,
                             border_mode=th_padding,
                             subsample=(1, 1),
                             input_shape=None,
                             filter_shape=pointwise_kernel_shape,
                             filter_dilation=dilation_rate)
    conv_out = _postprocess_conv2d_output(conv_out, x, padding,
                                          pointwise_kernel_shape,
                                          strides, data_format)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -