其次，Adam 中 没有使用 bias_correct
"""
optimizer = extend_with_piecewise_linear_lr(Adam)
optimizer = extend_with_wight_decay(optimizer)
params = {
'lr_schedule': lr_schedule,
'learning_rate': learning_rate,
'weight_decay_rate': weight_decay_rate,
'exclude_from_weight_decay': exclude_from_weight_decay,
'bias_correct': False
}
# grad acc
if grad_accum_steps > 1:
optimizer = extend_with_gradient_accumulation(optimizer)
params['grad_accum_steps'] = grad_accum_steps
optimizer = optimizer(**params)
# compile
train_model.compile(optimizer=optimizer, loss=loss)
# load init ckpt
if ckpt is not None:
bert.load_weights_from_checkpoint(ckpt)
# callback
class ModelCheckpoint(keras.callbacks.Callback):
def on_epoch_end(self, epoch, logs=None):
trans = K.eval(CRF.trans)
wordseg.trans = trans
print(trans)
acc = evaluate(val_data)
if acc > self.best_acc:
self.best_acc = acc
model.save_weights('./best_model.weights')
print('acc is: {:.3f}, best acc is :{:.4f}'.format(acc, self.best_acc))
def on_train_end(self, logs=None):
model.load_weights('./best_model.weights')
public_evaluate(test_path, test_result_path, test_score_path)
opt = extend_with_gradient_accumulation(Adam)
opt = opt(learning_rate=lr)
model.compile(loss=CRF.sparse_loss,
optimizer=opt,
metrics=[CRF.sparse_accuracy])
if __name__ == '__main__':
evaluator = Evaluator()
train_genarator = data_generator(train_data, batch_size)
model.fit_generator(train_genarator.generator(),
steps_per_epoch=len(train_genarator),
epochs=epochs,
callbacks=[evaluator])
else:
model.load_weights('./best_model.weights')