def main():
r"""
Main function.
"""
model = QANet()
# initial parameters
logging.info('Initial paramters...')
if opt.load_trained_model:
model.load_parameters(opt.trained_model_name, ctx=CTX)
else:
logging.info('Initial model parameters...')
initial_model_parameters(model)
print(model)
if opt.is_train:
loss_function = MySoftmaxCrossEntropy()
ema = ExponentialMovingAverage(decay=opt.ema_decay)
# initial trainer
trainer = gluon.Trainer(
model.collect_params(), 'adam', {
'learning_rate': opt.init_learning_rate,
'beta1': opt.beta1,
'beta2': opt.beta2,
'epsilon': opt.epsilon
})
if opt.load_trained_model:
trainer.load_states(opt.trained_trainer_name)
# initial dataloader
train_data_loader = DataLoader(batch_size=opt.train_batch_size,
dev_set=False)
# train
logging.info('Train')
train(model, train_data_loader, trainer, loss_function, ema)
else:
logging.info('Evaluating dev set...')
f1_score, em_score = evaluate(model, dataset_type='dev', ema=None)
logging.debug('The dev dataset F1 is:%.5f, and EM is: %.5f', f1_score,
em_score)
def main():
r"""
Main function.
"""
model = QANet()
# initial parameters
print('Initial paramters...')
if NEED_LOAD_TRAINED_MODEL:
model.load_parameters(TARGET_MODEL_FILE_NAME, ctx=CTX)
else:
print('Initial model parameters...')
initial_model_parameters(model)
print(model)
if TRAIN_FLAG is True:
loss_function = MySoftmaxCrossEntropy()
ema = ExponentialMovingAverage(decay=EXPONENTIAL_MOVING_AVERAGE_DECAY)
# initial trainer
trainer = gluon.Trainer(
model.collect_params(), 'adam', {
'learning_rate': INIT_LEARNING_RATE,
'beta1': BETA1,
'beta2': BETA2,
'epsilon': EPSILON
})
if NEED_LOAD_TRAINED_MODEL:
trainer.load_states(TARGET_TRAINER_FILE_NAME)
# initial dataloader
train_data_loader = DataLoader(batch_size=TRAIN_BATCH_SIZE,
dev_set=False)
# train
print('Train...')
train(model, train_data_loader, trainer, loss_function, ema)
else:
print('Evaluating dev set...')
f1_score, em_score = evaluate(model, dataset_type='dev', ema=None)
print('The dev dataset F1 is:%s, and EM is: %s' % (f1_score, em_score))
def train(num_units, batch_size, sentence_size, embedding_size, ctx):
net = QANet(num_units, batch_size, sentence_size, embedding_size)
net.collect_params().initialize(ctx)
loss = LogLoss()
trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': .1})