To practice RNNs, we will use the dataset previously used to construct the CNN. I refer to the MNIST dataset, a large database of handwritten digits. It has a set of 70,000 examples of data. It is a subset of NIST's larger dataset. Images of 28 x 28 pixel resolution are stored in a matrix of 70,000 rows and 785 columns; each pixel value from the 28 x 28 matrix and one value is the actual digit. In a fixed-size image, the digits have been size-normalized.
In this case, we will implement an RNN (LSTM) using the TensorFlow library to classify images. We will consider every image row as a sequence of pixels. Because the MNIST image shape is 28 x 28, we will handle 28 sequences of 28 time steps for every sample.
To start, we will analyze the code line by line; then we will see how to process it with the tools made available by...