Defining the decoder of the inference model

The inference model is the model that will be used out in the wild to perform translations when required by the user. In this exercise, you will need to implement the decoder of the inference model.

The inference model decoder is different to the decoder of the training model. We can't feed the decoder with French words because that is what we want to predict. Luckily, there is a solution. We can use the predicted French word from the previous time step to feed the inference model decoder. Therefore, when you want to generate a translation, the decoder needs to generate one word at a time, while consuming the previous output as an input.

For this exercise, the variables hsize (hidden size of the GRU layer), fr_len and fr_vocab have been imported.

Define an Input layer which accepts a batch of onehot encoded French word sequences (sequence length 1).
Define another Input layer which accepts a batch of hsize state, which you will use to feed the previous state to the decoder.
Get the output and state of the decoder GRU.
Define a model that accepts the French words Input and the previous state Input and outputs the final prediction and the new GRU state.

Exercise

Defining the decoder of the inference model

Instructions