def infer(self, dev_data):
hparams = self.hparams
sess = self.sess
assert len(dev_data[0]) == len(
dev_data[1]), "Size of features data must be equal to label"
preds = []
total_loss = []
for idx in range(len(dev_data[0]) // hparams.batch_size + 1):
batch=dev_data[0][idx*hparams.batch_size:\
min((idx+1)*hparams.batch_size,len(dev_data[0]))]
if len(batch) == 0:
break
batch = utils.hash_batch(batch, hparams)
label=dev_data[1][idx*hparams.batch_size:\
min((idx+1)*hparams.batch_size,len(dev_data[1]))]
"""
infer时不开dropout,BN
"""
pred=sess.run(self.prob,feed_dict=\
{self.features:batch,self.label:label,self.use_norm:False,
self.dnn_dropout:len(hparams.dnn_dropout) * [1]
})
preds.append(pred)
preds = np.concatenate(preds)
return preds
def train(self, train_data, dev_data):
hparams = self.hparams # 超参数
sess = self.sess # sess
assert len(train_data[0]) == len(
train_data[1]), "Size of features data must be equal to label"
"""
训练
"""
for epoch in range(hparams.epoch):
info = {}
info['loss'] = []
info['norm'] = []
start_time = time.time()
for idx in range(math.ceil(
len(train_data[0]) / hparams.batch_size)):
"""
train_data[0] 是数据， train_data[1]是label
"""
# 截取一个batch的数据，最后一个batch的大小可能不满batch_size
batch = train_data[0][idx * hparams.batch_size:min(
(idx + 1) * hparams.batch_size, len(train_data[0]))]
# 对每一个特征值做hash(类似于做one-hot)
batch = utils.hash_batch(batch, hparams)
# 截取一个batch的label，最后一个batch的大小可能不满batch_size
label = train_data[1][idx * hparams.batch_size:min(
(idx + 1) * hparams.batch_size, len(train_data[1]))]
loss,_,norm=sess.run([self.loss,self.update,self.grad_norm],feed_dict=\
{self.features:batch,self.label:label,self.use_norm:True,
self.second_order_dropout:hparams.second_order_dropout,
self.dnn_dropout:hparams.dnn_dropout
})
info['loss'].append(loss) # 损失
info['norm'].append(norm)
if (idx + 1) % hparams.num_display_steps == 0:
info['learning_rate'] = hparams.learning_rate
info["train_ppl"] = np.mean(info['loss'])
info["avg_grad_norm"] = np.mean(info['norm'])
utils.print_step_info(" ", epoch, idx + 1, info)
del info
info = {}
info['loss'] = []
info['norm'] = []
"""eval时如果模型效果好会保存模型"""
if (idx + 1) % hparams.num_eval_steps == 0 and dev_data:
T = (time.time() - start_time)
self.eval(T, dev_data, hparams, sess)
self.saver.restore(sess, 'model_tmp/model')
T = (time.time() - start_time)
self.eval(T, dev_data, hparams, sess)
os.system("rm -r model_tmp")
def train(self, train_data, dev_data=None):
"""
train_data: (num_train_data * field*size, num_train_data *1)
dev_data: (num_dev_data * field*size, num_dev_data * field*size)
"""
hparams = self.hparams
sess = self.sess
assert len(train_data[0]) == len(train_data[1])
for epoch in range(hparams.epoch):
info = {}
info['loss'] = []
info['norm'] = []
start_time = time.time()
for idx in range(math.ceil(
len(train_data[0]) / hparams.batch_size)):
"""分batch,并针对每个特征值做hash"""
batch=train_data[0][idx*hparams.batch_size:\
min((idx+1)*hparams.batch_size,len(train_data[0]))]
batch = utils.hash_batch(batch, hparams) # hash
label=train_data[1][idx*hparams.batch_size:\
min((idx+1)*hparams.batch_size,len(train_data[1]))]
"""训练"""
loss,_,norm=sess.run([self.loss,self.update,self.grad_norm],\
feed_dict={self.features:batch,self.label:label})
"""打印一些信息"""
info['loss'].append(loss)
info['norm'].append(norm)
if (idx + 1) % hparams.num_display_steps == 0:
info['learning_rate'] = hparams.learning_rate
info["train_ppl"] = np.mean(info['loss'])
info["avg_grad_norm"] = np.mean(info['norm'])
utils.print_step_info(" ", epoch, idx + 1, info)
del info
info = {}
info['loss'] = []
info['norm'] = []
if (idx + 1) % hparams.num_eval_steps == 0 and dev_data:
T = (time.time() - start_time)
self.eval(T, dev_data, hparams, sess)
"""保存模型"""
self.saver.restore(sess, 'model_tmp/model')
T = (time.time() - start_time)
self.eval(T, dev_data, hparams, sess)
os.system("rm -r model_tmp")
def get_embedding(self, dev_data):
hparams = self.hparams
sess = self.sess
assert len(dev_data[0]) == len(
dev_data[1]), "Size of features data must be equal to label"
embedding = []
total_loss = []
for idx in range(len(dev_data[0]) // hparams.batch_size + 1):
batch=dev_data[0][idx*hparams.batch_size:\
min((idx+1)*hparams.batch_size,len(dev_data[0]))]
if len(batch) == 0:
break
batch = utils.hash_batch(batch, hparams)
label=dev_data[1][idx*hparams.batch_size:\
min((idx+1)*hparams.batch_size,len(dev_data[1]))]
temp=sess.run(self.emb_inp_v2,\
feed_dict={self.features:batch,self.label:label})
embedding.append(temp)
embedding = np.concatenate(embedding, 0)
return embedding