Treinando o modelo com Teacher Forcing

Machine Translation with Keras

Thushan Ganegedara

Data Scientist and Author

Treinamento do modelo em detalhes

Treinamento do modelo requer:
- Uma função de perda (ex.: categorical crossentropy)
- Um otimizador (ex.: Adam)

Treinamento do modelo em detalhes

Para calcular a perda, são necessários:
- Previsões probabilísticas geradas a partir das entradas ([batch_size, seq_len, vocab_size])
  - ex.: [[0.11,...,0.81,0.04], [0.05,...,0.01, 0.93], ..., [0.78,..., 0.03,0.01]]
- Alvos reais em one-hot em francês ([batch_size, seq_len, vocab_size])
  - ex.: [[0, ..., 1, 0], [0, ..., 0, 1],..., [0, ..., 1, 0]]
- Crossentropy: diferença entre os alvos e as palavras previstas
A perda é passada a um otimizador, que ajusta os parâmetros do modelo para minimizá-la

Treinando o modelo com Teacher Forcing

n_epochs, bsize = 3, 250
for ei in range(n_epochs):
  for i in range(0,data_size,bsize):

    # Encoder inputs, decoder inputs and outputs
    en_x = sents2seqs('source', en_text[i:i+bsize], onehot=True, reverse=True)
    de_xy = sents2seqs('target', fr_text[i:i+bsize], onehot=True)

    # Separating decoder inputs and outputs
    de_x = de_xy[:,:-1,:]    
    de_y = de_xy[:,1:,:]

    # Training and evaulating on a single batch
    nmt_tf.train_on_batch([en_x,de_x], de_y)    
    res = nmt_tf.evaluate([en_x,de_x], de_y, batch_size=bsize, verbose=0)
    print("{} => Train Loss:{}, Train Acc: {}".format(ei+1,res[0], res[1]*100.0))

Fatiamento de array em detalhes

    de_x = de_xy[:,:-1,:]    
    de_y = de_xy[:,1:,:]

Fatiamento de array

Criando dados de treino e validação

train_size, valid_size = 800, 200
# Creating data indices
inds = np.arange(len(en_text))
np.random.shuffle(inds)


# Separating train and valid indices
train_inds = inds[:train_size]
valid_inds = inds[train_size:train_size+valid_size]


# Extracting train and valid data
tr_en = [en_text[ti] for ti in train_inds]
tr_fr = [fr_text[ti] for ti in train_inds]

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

print('Training (EN):\n', tr_en[:2], '\nTraining (FR):\n', tr_fr[:2])
print('\nValid (EN):\n', tr_en[:2], '\nValid (FR):\n', tr_fr[:2])

Valid (EN): ['she saw that rusty blue truck .', 'france is usually snowy during autumn , ...'] 
Valid (FR): ['elle a vu ce camion bleu rouillé .', 'la france est généralement enneigée en automne ...']

Treinando com validação

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, reverse=True)
    de_xy = sents2seqs('target', tr_fr[i:i+bsize], onehot=True)
    de_x, de_y = de_xy[:,:-1,:], de_xy[:,1:,:]
    nmt_tf.train_on_batch([en_x, de_x], de_y)

  v_en_x = sents2seqs('source', v_en, onehot=True, reverse=True)
  v_de_xy = sents2seqs('target', v_fr, onehot=True)
  v_de_x, v_de_y = v_de_xy[:,-1,:], v_de_xy[:,1:,:]

  res = nmt_tf.evaluate([v_en_x, v_de_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 => Loss:4.784221172332764, Val Acc: 1.4999999664723873
Epoch 2 => Loss:4.716882228851318, Val Acc: 44.458332657814026
Epoch 3 => Loss:4.63267183303833, Val Acc: 47.333332896232605

Vamos treinar!

Machine Translation with Keras