dense = ScaleDense(lr_multiplier=5,
units=num_classes,
activation='softmax',
weights=predecessor_model.layers[-1].get_weights())
output = dense(output)
predecessor_3_model = Model(predecessor_model.inputs, output)
predecessor_3_model.compile(
loss='sparse_categorical_crossentropy',
optimizer=Adam(1e-5), # 用足够小的学习率
metrics=['sparse_categorical_accuracy'],
)
predecessor_3_model.summary()
if __name__ == '__main__':
epochs = 5
# 训练predecessor
predecessor_evaluator = Evaluator('best_predecessor.weights')
predecessor_model.fit_generator(train_generator.generator(),
steps_per_epoch=len(train_generator),
epochs=epochs,
callbacks=[predecessor_evaluator])
# 训练predecessor_3_model
predecessor_model.load_weights('best_predecessor.weights')
predecessor_3_evaluator = Evaluator('best_predecessor_3.weights')
predecessor_3_model.fit_generator(train_generator.generator(),
steps_per_epoch=len(train_generator),
epochs=10,
callbacks=[predecessor_3_evaluator])
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')
total += len(y_true)
right += (y_true == y_pred).sum()
print(total, right)
return right / total
class Evaluator(keras.callbacks.Callback):
def __init__(self, save_path='best_model.weights'):
self.best_val_acc = 0.
self.save_path = save_path
def on_epoch_end(self, epoch, logs=None):
val_acc = evaluate()
if val_acc > self.best_val_acc:
self.best_val_acc = val_acc
self.model.save_weights(self.save_path)
print('current acc :{}, best val acc: {}'.format(
val_acc, self.best_val_acc))
if __name__ == '__main__':
evaluator = Evaluator()
train_model.fit_generator(train_generator.generator(),
steps_per_epoch=len(train_generator),
epochs=5,
callbacks=[evaluator])
else:
classifier.load_weights('best_model.weights')
total, right = 0., 0.
for x_true, y_true in tqdm(data):
y_pred = model.predict(x_true).argmax(axis=1)
y_true = y_true[:, 0]
total += len(y_true)
right += (y_true == y_pred).sum()
return right / total
class Evaluator(keras.callbacks.Callback):
def __init__(self):
self.best_acc = 0.
def on_epoch_end(self, epoch, logs=None):
acc = evaluate(val_generator)
if acc > self.best_acc:
self.best_acc = acc
self.model.save_weights('best_baseline.weights')
print('acc: {}, best acc: {}'.format(acc, self.best_acc))
if __name__ == '__main__':
evaluator = Evaluator()
model.fit_generator(train_generator.generator(),
steps_per_epoch=len(train_generator),
epochs=epochs,
callbacks=[evaluator])
else:
model.load_weights('best_baseline.weights')
right += (y_true == y_pred).sum()
return right / total
class Evaluator(keras.callbacks.Callback):
def __init__(self, savename):
self.best_val_acc = 0.
self.savename = savename
def on_epoch_end(self, epoch, logs=None):
val_acc = evaluate(valid_generator, self.model)
if val_acc > self.best_val_acc:
self.best_val_acc = val_acc
self.model.save_weights(self.savename)
print(
u'val_acc: %.5f, best_val_acc: %.5f\n' %
(val_acc, self.best_val_acc)
)
if __name__ == '__main__':
evaluator = Evaluator('best_clf.weights')
model.compile(loss='sparse_categorical_crossentropy', optimizer=Adam(1e-5), metrics=['acc'])
model.fit_generator(train_sim_generator.generator(),
steps_per_epoch=len(train_sim_generator) * 2,
epochs=5,
callbacks=[evaluator]
)
else:
model.load_weights('best_clf.weights')
'answers': [{
'text': new_answer
}]
})
paragraphs = []
for context, qas in paras.items():
paragraphs.append({'context': context, 'qas': qas})
data = {'data': [{'paragraphs': paragraphs}]}
with open(file_name, 'w') as f:
json.dump(data, f)
if __name__ == '__main__':
train_generator = data_generator(train_data + val_data, batch_size)
evaluator = Evaluator()
model.fit_generator(train_generator.generator(),
steps_per_epoch=len(train_generator),
epochs=epochs,
callbacks=[evaluator])
# generate question and answer
model.load_weights('question_answer_generation.weights')
file_name = '/home/mingming.xu/datasets/NLP/qa/dureader_robust-data/train_qa_generator.json'
generate_new_data(file_name)
else:
model.load_weights('question_answer_generation.weights')
file_name = '/home/mingming.xu/datasets/NLP/qa/dureader_robust-data/train_qa_generator.json'
generate_new_data(file_name)
if acc > self.best_acc:
self.best_acc = acc
self.model.save_weights('best_pet_model.weights')
print('acc :{}, best acc:{}'.format(acc, self.best_acc))
def write_to_file(path):
preds = []
for x, _ in tqdm(test_generator):
pred = predict(x)
preds.extend(pred)
ret = []
for data, p in zip(test_data, preds):
ret.append([data[0], data[2], str(p)])
with open(path, 'w') as f:
for r in ret:
f.write('\t'.join(r) + '\n')
if __name__ == '__main__':
evaluator = Evaluator()
train_model.fit_generator(train_generator.generator(),
steps_per_epoch=len(train_generator),
epochs=10,
callbacks=[evaluator])
train_model.load_weights('best_pet_model.weights')
write_to_file('submission.tsv')
