def train(self, dataset):
self.training_control = utils.training_control(self.global_step, print_span=10,
evaluation_span=100,
max_step=self.max_step) # batch*evaluation_span = dataset size = one epoch
for batch in dataset.training_generator(batch_size=self.batch_size, portion= 0.5):
accuracy, loss, wm, _ = self.run([self.accuracy, self.loss,
self.warmup,
self.train_op],
feed_dict={self.input_x: batch['x'],
self.input_dev: batch['dev'],
self.input_y: batch['y'],
self.is_training: True})
step_control = self.run(self.training_control)
if step_control['time_to_print']:
print('train_loss = ' + str(loss) +
' train_acc= '+str(accuracy) +
' step ' + str(step_control['step']) +
' warmup' + str(wm))
if step_control['time_to_stop']:
break
if step_control['time_to_evaluate']:
if_stop = self.evaluate(dataset.val_set)
self.save_checkpoint()
if if_stop:
break
def train(self, dataset, lr):
self.training_control = utils.training_control(
self.global_step,
print_span=10,
evaluation_span=round(dataset.train_set.shape[0] /
self.batch_size),
max_step=100000
) # batch*evaluation_span = dataset size = one epoch
for batch in dataset.training_generator(batch_size=self.batch_size):
results, loss, _ = self.run(
[self.logits, self.loss, self.train_op],
feed_dict={
self.input_x: batch['x'],
self.input_y: batch['y'],
self.is_training: True
})
step_control = self.run(self.training_control)
if step_control['time_to_print']:
print('train_loss= ' + str(loss) + ' round' +
str(step_control['step']))
if step_control['time_to_stop']:
break
if step_control['time_to_evaluate']:
if_stop = self.evaluate(dataset.val_set)
self.save_checkpoint()
if if_stop:
break