So we can concentrate on the models, we will use the same dataset we are familiar with. Even though we are working with static images, we can treat these as a time series by unrolling each 28 pixel input size over 28 time steps, enabling the network to make a computation on the complete image:
In the preceding model, we use the nn.RNN class to create a model with two recurrent layers. The nn.RNN class has the following default signature:
nn.RNN(input_size, hidden_size, num_layers, batch_first=False, nonlinearity = 'tanh'
The input is our 28 x 28 MNIST images. This model takes 28 pixels of each image, unrolling ...
