We will now briefly look at the graph and loss of the network during training in TensorBoard:
Graph and loss of network during training
We can see that the encoding and decoding layer is fed with the corresponding word embeddings of the source and target text. The attention mechanism couples the encoder output and the decoder through the memory layer weights. You can click on the respective components in TensorBoard to understand the details of the connections and tensor dimensions. We will now take a look at the plot of the cost function during training:
Cost functioning plot