def generator_fn(x, y, shuffle=True, looping=True):
return nmt_train_generator(x,
y,
de_vocab_size,
batch_size,
shuffle=shuffle,
looping=looping)
def __init__(self, data, eval_every, batch_size, tar_vocab_size, encoder,
decoder):
self.src, self.tar = data
self.generator = nmt_train_generator(self.src, self.tar,
tar_vocab_size, batch_size)
self.eval_every = eval_every
self.batch_size = batch_size
self.encoder = encoder
self.decoder = decoder
return Member(model,
param_names=['lr', 'dropout'],
tune_lr=True,
use_eval_metric='bleu',
custom_metrics={
'bleu':
lambda x, y, _: bleu_score_enc_dec(
encoder_model, decoder_model, x, y, batch_size)
})
steps = en_train.shape[0] // batch_size
steps_ready = 1000
eval_every = 100
generator_fn = lambda x, y, shuffle=True, looping=True: nmt_train_generator(
x, y, de_vocab_size, batch_size, shuffle=shuffle, looping=looping)
pbt = PbtPsoOptimizer(build_member,
population_size,
parameters,
steps_ready=steps_ready,
omega=0.5,
phi1=0.5,
phi2=1.0)
model, results = pbt.train(en_train_t,
de_train_t,
en_train_v,
de_train_v,
steps=steps,
eval_every=eval_every,
generator_fn=generator_fn)
# model parameters
hidden_size = 96
embedding_size = 100
timesteps = 30
# hyperparameters
lr = 0.001
dropout = 0.2
model, encoder_model, decoder_model = define_nmt(hidden_size,
embedding_size, timesteps,
en_vocab_size,
de_vocab_size, dropout,
lr)
train_generator = nmt_train_generator(en_train_t, de_train_t,
de_vocab_size, batch_size)
steps = en_train.shape[0] // batch_size
eval_every = 100
bleu_logger = BleuLogger((en_train_v, de_train_v), eval_every, batch_size,
de_vocab_size, encoder_model, decoder_model)
model.fit_generator(train_generator,
steps_per_epoch=steps,
callbacks=[bleu_logger])
model.save_weights('baseline.h5')
results = pd.DataFrame({
'loss': bleu_logger.losses,
'bleu': bleu_logger.scores