def infer(self, dev):
hparams = self.hparams
sess = self.sess
preds = []
total_loss = []
a = hparams.batch_size
hparams.batch_size = hparams.infer_batch_size
dev_single_features = dev[hparams.single_features].values
if hparams.multi_features is not None:
dev_multi_features = dev[hparams.multi_features].values
if hparams.dense_features is not None:
dev_dense_features = dev[hparams.dense_features].values
if hparams.kv_features is not None:
dev_kv_features = dev[hparams.kv_features].values
if hparams.cross_features is not None:
dev_cross_features = dev[hparams.cross_features].values
for idx in tqdm(range(len(dev) // hparams.batch_size + 1),
total=len(dev) // hparams.batch_size + 1):
single_batch = dev_single_features[idx * hparams.batch_size:min(
(idx + 1) * hparams.batch_size, len(dev))]
if len(single_batch) == 0:
break
feed_dic = {}
feed_dic[self.use_norm] = False
if hparams.single_features is not None:
single_batch = utils.hash_single_batch(single_batch, hparams)
feed_dic[self.single_features] = single_batch
if hparams.multi_features is not None:
multi_batch = dev_multi_features[idx * hparams.batch_size:min(
(idx + 1) * hparams.batch_size, len(dev))]
multi_batch, multi_weights = utils.hash_multi_batch(
multi_batch, hparams)
feed_dic[self.multi_features] = multi_batch
feed_dic[self.multi_weights] = multi_weights
if hparams.dense_features is not None:
feed_dic[self.dense_features]=dev_dense_features[idx*hparams.batch_size:\
min((idx+1)*hparams.batch_size,len(dev))]
if hparams.kv_features is not None:
feed_dic[self.kv_features]=dev_kv_features[idx*hparams.batch_size:\
min((idx+1)*hparams.batch_size,len(dev))]
if hparams.cross_features is not None:
cross_batch = dev_cross_features[idx * hparams.batch_size:min(
(idx + 1) * hparams.batch_size, len(dev))]
cross_batch = utils.hash_single_batch(cross_batch, hparams)
feed_dic[self.cross_features] = cross_batch
feed_dic[self.use_norm] = False
pred = sess.run(self.val, feed_dict=feed_dic)
preds.append(pred)
preds = np.concatenate(preds)
dev['temp'] = preds
preds = hparams.test_scaler.inverse_transform(dev[['temp']])[:, 0]
del dev['temp']
hparams.batch_size = a
return preds
def train(self,train,dev):
hparams=self.hparams
sess=self.sess
train_single_features=train[hparams.single_features].values
train_label=train[hparams.label].values
if hparams.multi_features is not None:
train_multi_features=train[hparams.multi_features].values
train_multi_weights=train[hparams.multi_weights].values
if hparams.dense_features is not None:
train_dense_features=train[hparams.dense_features].values
if hparams.kv_features is not None:
train_kv_features=train[hparams.kv_features].values
if hparams.cross_features is not None:
train_cross_features=train[hparams.cross_features].values
for epoch in range(hparams.epoch):
info={}
info['loss']=[]
info['norm']=[]
start_time = time.time()
for idx in range(len(train)//hparams.batch_size+3):
if idx*hparams.batch_size>=len(train):
T=(time.time()-start_time)
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)
if dev is not None:
self.eval(T,dev,hparams,sess)
break
feed_dic={}
single_batch=train_single_features[idx*hparams.batch_size:min((idx+1)*hparams.batch_size,len(train))]
single_batch=utils.hash_single_batch(single_batch,hparams)
feed_dic[self.single_features]=single_batch
if hparams.multi_features is not None:
multi_batch=train_multi_features[idx*hparams.batch_size:min((idx+1)*hparams.batch_size,len(train))]
multi_weight=train_multi_weights[idx*hparams.batch_size:min((idx+1)*hparams.batch_size,len(train))]
multi_batch,multi_weights=utils.hash_multi_batch(multi_batch,multi_weight,hparams)
feed_dic[self.multi_features]=multi_batch
feed_dic[self.multi_weights]=multi_weights
if hparams.dense_features is not None:
feed_dic[self.dense_features]=train_dense_features[idx*hparams.batch_size:\
min((idx+1)*hparams.batch_size,len(train))]
if hparams.kv_features is not None:
feed_dic[self.kv_features]=train_kv_features[idx*hparams.batch_size:\
min((idx+1)*hparams.batch_size,len(train))]
if hparams.cross_features is not None:
cross_batch=train_cross_features[idx*hparams.batch_size:min((idx+1)*hparams.batch_size,len(train))]
cross_batch=utils.hash_single_batch(cross_batch,hparams)
feed_dic[self.cross_features]=cross_batch
label=train_label[idx*hparams.batch_size: min((idx+1)*hparams.batch_size,len(train))]
feed_dic[self.label]=label
feed_dic[self.use_norm]=True
loss,_,norm=sess.run([self.loss,self.update,self.grad_norm],feed_dict=feed_dic)
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']=[]
T=(time.time()-start_time)
return self.best_score