Introduction to Teacher Forcing

Machine Translation with Keras

Thushan Ganegedara

Data Scientist and Author

The previous machine translator model

The previous model

Encoder Decoder full model

Encoder GRU
- Consumes English words
- Outputs a context vector
Decoder GRU
- Consumes the context vector
- Outputs a sequence of GRU outputs
Decoder Prediction layer
- Consumes the sequence of GRU outputs
- Outputs prediction probabilities for French words

Analogy: Training without Teacher Forcing

No teacher forcing 1

Analogy: Training without Teacher Forcing

No teacher forcing 2

Analogy: Training without Teacher Forcing

No teacher forcing 3

Analogy: Training with Teacher Forcing

Teacher forcing 1

Analogy: Training with Teacher Forcing

Teacher forcing 2

Analogy: Training with Teacher Forcing

Teacher forcing 3

Analogy: Training with Teacher Forcing

Teacher forcing 4

The previous machine translator model

The previous model

Encoder decoder full model

Teacher-forced model

Encoder decoder model with teacher forcing

Implementing the model with Teacher Forcing

Encoder

en_inputs = layers.Input(shape=(en_len, en_vocab))
en_gru = layers.GRU(hsize, return_state=True)
en_out, en_state = en_gru(en_inputs)

Decoder GRU

de_inputs = layers.Input(shape=(fr_len-1, fr_vocab))
de_gru = layers.GRU(hsize, return_sequences=True)
de_out = de_gru(de_inputs, initial_state=en_state)

Inputs and outputs

Encoder input - e.g. I, like, dogs
Decoder input - e.g. J'aime, les
Decoder output - e.g. les, chiens

Teacher forcing inputs and outputs

Implementing the model with Teacher Forcing

Encoder

en_inputs = layers.Input(shape=(en_len, en_vocab))
en_gru = layers.GRU(hsize, return_state=True)
en_out, en_state = en_gru(en_inputs)

Decoder GRU

de_inputs = layers.Input(shape=(fr_len-1, fr_vocab))
de_gru = layers.GRU(hsize, return_sequences=True)
de_out = de_gru(de_inputs, initial_state=en_state)

Decoder Prediction

de_dense = layers.TimeDistributed(layers.Dense(fr_vocab, activation='softmax'))
de_pred = de_dense(de_out)

Compiling the model

nmt_tf = Model(inputs=[en_inputs, de_inputs], outputs=de_pred)
nmt_tf.compile(optimizer='adam', loss="categorical_crossentropy", metrics=["acc"])

Preprocessing data

Encoder
- Inputs - All English words (onehot encoded)
  - en_x = sents2seqs('source', en_text, onehot=True, reverse=True)
Decoder
```
de_xy = sents2seqs('target', fr_text, onehot=True)
```
- Inputs - All French words except the last word (onehot encoded)
  - de_x = de_xy[:,:-1,:]
- Outputs/Targets - All French words except the first word (onehot encoded)
  - de_y = de_xy[:,1:,:]

Let's practice!

Machine Translation with Keras