As best I can tell, trainnet vectorizes its computations across the batch dimension of every minibatch. I.e., during training, it feeds minibatches of data into the network as a complete SxSxSxCxB array where B is the number of observations in the minibatch. This can be quite memory intensive depending on S^3*C, and limits the minibatch sizes that can be selected for training.
Is there no way to force trainnet to compute the loss function and its gradients incrementally, by feeding one member of the minibatch (of size SxSxSxCx1) to the network at a time?As best I can tell, trainnet vectorizes its computations across the batch dimension of every minibatch. I.e., during training, it feeds minibatches of data into the network as a complete SxSxSxCxB array where B is the number of observations in the minibatch. This can be quite memory intensive depending on S^3*C, and limits the minibatch sizes that can be selected for training.
Is there no way to force trainnet to compute the loss function and its gradients incrementally, by feeding one member of the minibatch (of size SxSxSxCx1) to the network at a time? As best I can tell, trainnet vectorizes its computations across the batch dimension of every minibatch. I.e., during training, it feeds minibatches of data into the network as a complete SxSxSxCxB array where B is the number of observations in the minibatch. This can be quite memory intensive depending on S^3*C, and limits the minibatch sizes that can be selected for training.
Is there no way to force trainnet to compute the loss function and its gradients incrementally, by feeding one member of the minibatch (of size SxSxSxCx1) to the network at a time? trainnet, memory, minibatch, vectorization MATLAB Answers — New Questions

​