Training the NMT model

Machine Translation with Keras

Thushan Ganegedara

Data Scientist and Author

Revisiting the 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

Optimizing the parameters

GRU layer and Dense layer have parameters
Often represented by W (weights) and b (bias) (Initialized with random values)
Responsible for transforming a given input to an useful output
Changed over time to minimize a given loss using an optimizer
- Loss: Computed as the difference between:
  - The predictions (i.e. French words generated with the model)
  - The actual outputs (i.e. actual French words).
Informed the model during model compilation

nmt.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['acc'])

Training the model

Training iterations

for ei in range(n_epochs): # Single traverse through the dataset
  for i in range(0,data_size,bsize): # Processing a single batch

Obtaining a batch of training data

  en_x = sents2seqs('source', en_text[i:i+bsize], onehot=True, reverse=True)
  de_y = sents2seqs('target', en_text[i:i+bsize], onehot=True)

Training on a single batch of data
```
  nmt.train_on_batch(en_x, de_y)
```

Evaluating the model

  res = nmt.evaluate(en_x, de_y, batch_size=bsize, verbose=0)

Training the model

Getting the training loss and the accuracy

  res = nmt.evaluate(en_x, de_y, batch_size=bsize, verbose=0)
  print("Epoch {} => Train Loss:{}, Train Acc: {}".format(
    ei+1,res[0], res[1]*100.0))

Epoch 1 => Train Loss:4.8036723136901855, Train Acc: 5.215999856591225
...
Epoch 1 => Train Loss:4.718592643737793, Train Acc: 47.0880001783371
...
Epoch 5 => Train Loss:2.8161656856536865, Train Acc: 56.40000104904175
Epoch 5 => Train Loss:2.527724266052246, Train Acc: 54.368001222610474
Epoch 5 => Train Loss:2.2689621448516846, Train Acc: 54.57599759101868
Epoch 5 => Train Loss:1.9934935569763184, Train Acc: 56.51199817657471
Epoch 5 => Train Loss:1.7581449747085571, Train Acc: 55.184000730514526
Epoch 5 => Train Loss:1.5613118410110474, Train Acc: 55.11999726295471

Avoiding overfitting

Break the dataset to two parts
- Training set - The model will be trained on
- Validation set - The model's accuracy will be monitored on
When the validation accuracy stops increasing, stop the training.

Overfitting point

Splitting the dataset

Define a train dataset size and validation dataset size
```
train_size, valid_size = 800, 200
```

Shuffle the data indices randomly

inds = np.arange(len(en_text))
np.random.shuffle(inds)

Get the train and valid indices

train_inds = inds[:train_size]
valid_inds = inds[train_size:train_size+valid_size]

Splitting the dataset

Split the dataset by separating,
- Data having train indices to a train set
- Data having valid indices to a valid set

tr_en = [en_text[ti] for ti in train_inds]
tr_fr = [fr_text[ti] for ti in train_inds]

v_en = [en_text[ti] for ti in valid_inds]
v_fr = [fr_text[ti] for ti in valid_inds]

Training the model with validation

n_epochs, bsize = 5, 250
for ei in range(n_epochs):

  for i in range(0,train_size,bsize):
    en_x = sents2seqs('source', tr_en[i:i+bsize], onehot=True, pad_type='pre')
    de_y = sents2seqs('target', tr_fr[i:i+bsize], onehot=True)
    nmt.train_on_batch(en_x, de_y)

  v_en_x = sents2seqs('source', v_en, onehot=True, pad_type='pre')
  v_de_y = sents2seqs('target', v_fr, onehot=True)

  res = nmt.evaluate(v_en_x, v_de_y, batch_size=valid_size, verbose=0)
  print("Epoch: {} => Loss:{}, Val Acc: {}".format(ei+1,res[0], res[1]*100.0))

Epoch 1 => Train Loss:4.8036723136901855, Train Acc: 5.215999856591225

Let's practice!

Machine Translation with Keras