Suppose that in the stochastic gradient descent method we wish to repeatedly draw minibatches of size \(N\) from \(\tau_{n}\) where we assume that \(N \times m=n\) for some large integer \(\mathrm{m}\). Instead of repeatedly resampling from \(\tau_{n}\) an alternative is to reshuffle \(\tau_{n}\) via a random permutation \(\Pi\) and then advance sequentially through the reshuffled training set to construct \(m\) non-overlapping minibatches. A single traversal of such a reshuffled training set is called an epoch. The following pseudo-code describes the procedure.

Write Python code that implements the stochastic gradient descent with data reshuffling, and use it to train the neural net in Section 9.5.1.

Transcribed Image Text:

Algorithm 9.5.1: Stochastic Gradient Descent with Reshuffling input: Training set T, = {(x,y;)}}_, initial weight matrices and bias vectors (W, b), activation functions (S), learning rates (a1, a2,...}. output: The parameters of the trained learner. 111 and epoch - 0 2 while stopping condition is not met do 3 4 5 6 7 Draw U,..., U, du(0, 1). ...A Let II be the permutation of {1,...,n} that satisfies Un