Deep learning convolution3dLayer output structure

I want to add a custom layer after application of a convolution3dLayer, and need to know the order of the planes generated. For example if I have a 3x3x3 input and apply a 3x3x2 3dconv with depth of 2 and 4 features output, does the depth change take order precedence over the feature order or feature before depth. I would think the former with output like (f1d1, f1d2, f2d1, f2d2, f3d1, f3d2, f4d1, f4d2) and not (f1d1, f2d1, f3d1, f4d1, f1d2, f2d2, f3d2, f4d2) where f = feature and d = depth of the weight.

[Sai Bhargav Avula]

Can give me more detailed about the last line you mentioned ?

[Pete Gural]

While probing the deep learning toolbox with various architectures and reviewing some of Mathworks examples this week, I determined that the answer is the depth (channel) planes and feature planes that are output from a 3dconv are stored in different array indices, and thus depths/features are not concatenated together in one index - which was the source of my question on the ordering. The DL documentation could use a pictoral diagram of the 3d convolution process and description of the array structure output, especially the fact that the sliding 3d volume over height/width/depth also sums across the feature space, which is consistent with the way 2dconv works. It is just not mentioned explicitly. The multi-dimensional process was critical for understanding how I could get a complex input to process with complex arithmetic using just the real DL toolbox functionality on a first layer and then collapse that by taking the magnitude of the output's real and imaginary components. Another documentation problem concerns building custom layers, as insufficient examples are given. Can we see the ReLU class, maxpooling class, or conv2d class implementation as additional guides. Example of how to use the layer concatenation capability or how the grouped convolution works pictorially.

[Sai Bhargav Avula]

The Deep Learning toolbox is updated continiously and I have escalated all the above mentioned concerns to the respective team.