How to flatten the output of convolution1dLayer

Hello!

I tried Sequence Classification Using 1D Convolution example and replaced its layer structure.

As listed below, I changed the global average pooling layer to a simple flatten layer using the function layer.

layers = [ ...

sequenceInputLayer(numFeatures)

convolution1dLayer(filterSize,numFilters,Padding="causal")

reluLayer

layerNormalizationLayer

convolution1dLayer(filterSize,2*numFilters,Padding="causal")

reluLayer

layerNormalizationLayer

functionLayer(@(X) dlarray(X(:),"CB"),Formattable=true,Description="My flatten") %globalAveragePooling1dLayer

fullyConnectedLayer(numClasses)

softmaxLayer

classificationLayer];

"analyzeNetwork(layers)" had no error, but training failed.

Error "Incorrect dimensions for matrix multiplication" was poped in trainNetwork process.

I want to evaluate and compare traditional flatten like Keras flatten() to global pooling.

Is there any good way for this work?

Answer 1

[Abolfazl Chaman Motlagh]

by X(:) you are reshaping x to Nx1 vector, but you chose "CB" format for the array which is going to assume the array is 2 dimensional.( and it should be) try using size and reshape.

if the previous format is "CTB" then two first dimension should merge into one dimension. so do this:

functionLayer(@(X) dlarray(reshape(X,[size(X,1)*size(X,2),size(X,3)]),"CB"),Formattable=true,Description="My flatten") 

if it's not, format your data to be in this order because it is much harder task to merger non consecutive dimensions.

in creating costum layers and networks you will face a lot of errors like this for adaptation of new layer. so you may still have some troubles i guess:)

[bookmaster]

Thanks for the quick reply.

I tried the replied method, but the same error still occurred.

As mentioned in your reply, the conv1D output, i.e., each feature map (CBT) for sequence input should be replaced with the feature vector (CB) for the fully connected layer connected by Softmax. Indeed, the global pooling layer performs that kind of thing in this example.

Anyway, it is simple in the Keras environment but these additional troubles occur in the MATLAB environment.

Is that the wrong approach to design 'flatten' using a function layer?