batch_size=64
)
cost = rnnmodel.get_cost(X_v, L_v, Y_v)
out = rnnmodel.get_output(X_v, L_v)
logger.info('Initial validation cost: %s', np.mean(cost))
min_cost = cost
last_save = 0
for i in range(100000):
tick = time.clock()
X, L, Y = iterator.gen_batch(
split='train',
batch_size=64
)
tock = time.clock()
_ = rnnmodel.train_on_batch(X, L, Y)
if (i+1) % 100 == 0:
cost = rnnmodel.get_cost(X_v, L_v, Y_v)
logger.info('Validation cost after %s batches: %s', i, cost)
if cost < min_cost:
min_cost = cost
rnnmodel.save('/new_models/SoSmodel')
last_save = 0
logger.info('Saving model at validation cost %s', cost)
else:
last_save += 100
if last_save > 1000:
logger.info('No improvement after 1000 iterations. Stopping.')
break