Unfortunately, as of MATLAB R2020a, the requested workflow is not available with a built-in solution.
However, below are two possible workarounds:
1) "trainNetwork" approach with a custom layer
This approach would be based on using a layer property "RegressionWeights". The workflow is similar to the following example of a weighted classification output layer but needs to be implemented for a custom regression output layer.
To access the specific documentation for creating a custom weighted cross-entropy classification layer in MATLAB R2020a, please run the following command in the command window:
>> web(fullfile(docroot, 'deeplearning/ug/create-custom-weighted-cross-entropy-classification-layer.html'))
However, this approach might be quite cumbersome since it requires implementing the backward function and keeping track of the indexing of the batches as well. This approach is not recommended.
2) "dlnetwork" approach with custom training loop
This approach involves implementing a custom training loop using "dlnetwork". In this case, no output layers are required, but you can specify the loss function that you want to use (as explained in the "Model Predictions Function" section in the following example):
To access the specific documentation for training a network using a custom training loop in MATLAB R2020a, please run the following command in the command window:
>> web(fullfile(docroot, 'deeplearning/ug/train-network-using-custom-training-loop.html'))
The above example uses the "crossentropy" loss. However, you will need to implement your own version of the weighted mean squared error (WMSE) loss function and call it instead of the “crossentropy” function. Note that if you want to apply the weights on the mini-batches, you would need to extract the indexes of the data that you are using (in the above example, this is done by indexing with the variable "idx" in every epoch) and use them as input to the WMSE function.
Please also note that this would NOT require implementing a backward method, as the computation will be performed using Automatic Differentiation.
The current mean squared error (MSE) function is implemented as:
function X = mse(X, T, observationDim) % Half Mean Squared Error N = size(X, observationDim); X = sum((X-T).^2, 'all') / (2*N); end
Please modify it for WMSE. It will be straightforward to add the input W (indexed to have the same size as mini-batch) and perform the desired calculation.
The only drawback of this approach is that you will not be able to use the "trainNetwork" function. Instead, you will need to implement a custom training loop, which might be slightly cumbersome.
The "custom training loop" feature was introduced recently as a tool to provide users with more flexibility, and you should be able to achieve the desired workflow using the above approach.
Please follow the link below to search for the required information regarding the latest release:
