Locally connected 2D layer without summation over colors

yuval_meir · September 8, 2021, 11:40am

I want to create a neural network with a locally connected layer but without summation over the 3rd dimension (colors) of the inputs.

I saw in the docs of LocallyConnected2D that there is no “group” argument as in the conv2d.

Is there a way to do that?

Bhack · September 8, 2021, 12:17pm

Do you mean something like this?

github.com/keras-team/keras

Feature request: In-plane locally-connected 2D convolutions

opened 08:08PM - 28 Mar 18 UTC

closed 10:27PM - 24 Jun 21 UTC

tsoernes

I'm requesting a convolution that works on each channel separately using the sam…e filter (as https://www.tensorflow.org/versions/r1.2/api_docs/python/tf/contrib/layers/conv2d_in_plane), and but uses different filters spatially (like LocallyConnected2D). One approach would be to modify `LocallyConnected2D`: ``` individual_channels = tf.split(inputs, inputs.shape[3], -1) convs = [] for channel in individual_channels: conv = K.local_conv2d(channel, self.kernel, self.kernel_size, self.strides, (self.output_row, self.output_col), self.data_format) convs.append(conv) outputs = tf.concat(convs, -1) ``` where ``` self.kernel_shape = (output_row * output_col, self.kernel_size[0] * self.kernel_size[1], 1) ``` But the above approach is very slow.

yuval_meir · September 8, 2021, 12:30pm

Yes, I don’t see the solution there - you know how exactly tackle this problem?

Bhack · September 8, 2021, 12:35pm

I suggest you to comment on that ticket as It was just closed in june.