The mission of the model solver is to find the best set of parameters, that minimize the train/accuracy errors. On this chapter we will give a UML description with some piece of python/matlab code that allows you implement it yourself.

From the UML description we can infer some information about the Solver class:

It uses the training set, and has a reference to your model

Uses different type of optimizers(ex: SGD, ADAM, SGD with momentum)

Keep tracks of all the loss, accuracy during the training phase

Keep the set of parameters, that achieved best validation performance

This is the method called when you actually want to start a model training, the methods Step, Check_Accuracy are called inside the Train method:

Calculate number of iterations per epoch, based on number of epochs, train size, and batch size

Call step, for each iteration

Decay the learning rate

Calculate the validation accuracy

Cache the best parameters based on validation accuracy

Basically during the step operation the following operations are done:

Extract a batch from the training set.

Get the model loss and gradients

Perform a parameter update with one of the optimizers.

This method basically is called at the end of each epoch. Basically it uses the current set of parameters, and predict the whole validation set. The objective is at the end get the accuracy. $accuracy=mean(y_{predicted}==y_{validation})$â€‹

We mentioned during the "Step" operation that we get the model loss and gradients. This operation is implemented by the "getLoss" method. Consider the following basic model.

â€‹ Bellow we have the "getLoss" function for the previous simple model.

Also bellow we have the "softmax*loss" function including "dout", $$\frac{\partial L}{\partial X*{scores}}$$