def predict(self, model: nn.Module, data: BaseModel.Dataset) -> np.ndarray:
"""
The returned prediction is a 2D-array, each row corresponds to all the candidates,
and the ground-truth item poses the first.
Example: ground-truth items: [1, 2], 2 negative items for each instance: [[3,4], [5,6]]
predictions order: [[1,3,4], [2,5,6]]
"""
model.eval() #同样的样本，预测结果不一样
predictions = list()
dl = DataLoader(data,
batch_size=self.eval_batch_size,
shuffle=False,
num_workers=self.num_workers,
collate_fn=data.collate_batch,
pin_memory=self.pin_memory)
for batch in tqdm(dl,
leave=False,
ncols=100,
mininterval=1,
desc='Predict'):
prediction = model(utils.batch_to_gpu(
batch, model.device))['prediction'] #将batch中的数据转换为tensor
#对象初始化调用init方法，对象传参调用__call__方法。给model中传参会调用forword方法，因为在model中继承nn.Module,
#nn.Module中定义的__call__方法，其中调用了forword方法
predictions.extend(prediction.cpu().data.numpy())
return np.array(predictions)
def fit(self,
model: nn.Module,
data: BaseModel.Dataset,
epoch=-1) -> float:
gc.collect() #垃圾回收
torch.cuda.empty_cache() #显存才会在Nvidia-smi中释放
if model.optimizer is None:
model.optimizer = self._build_optimizer(model)
data.actions_before_epoch(
) # must sample before multi thread start 多线程启动前必须采样（负样本）
model.train()
loss_lst = list()
dl = DataLoader(data,
batch_size=self.batch_size,
shuffle=False,
num_workers=self.num_workers,
collate_fn=data.collate_batch,
pin_memory=self.pin_memory)
for batch in tqdm(dl,
leave=False,
desc='Epoch {:<3}'.format(epoch),
ncols=100,
mininterval=1): #batch_size中商品的另一个标签怎么确定？
batch = utils.batch_to_gpu(batch, model.device)
model.optimizer.zero_grad()
out_dict = model(batch)
loss = model.loss(out_dict)
loss.backward()
model.optimizer.step()
loss_lst.append(loss.detach().cpu().data.numpy())
return np.mean(loss_lst).item()
def fit(self, data: BaseModel.Dataset, epoch=-1) -> float:
model = data.model
if model.optimizer is None:
model.optimizer = self._build_optimizer(model)
data.actions_before_epoch() # must sample before multi thread start
model.train()
loss_lst = list()
dl = DataLoader(data,
batch_size=self.batch_size,
shuffle=True,
num_workers=self.num_workers,
collate_fn=data.collate_batch,
pin_memory=self.pin_memory)
for batch in tqdm(dl,
leave=False,
desc='Epoch {:<3}'.format(epoch),
ncols=100,
mininterval=1):
batch['epoch'] = epoch
batch = utils.batch_to_gpu(batch, model.device)
model.optimizer.zero_grad()
out_dict = model(batch)
loss = model.loss(out_dict)
loss.backward()
model.optimizer.step()
loss_lst.append(loss.detach().cpu().data.numpy())
return np.mean(loss_lst).item()
def predict(self, data: BaseModel.Dataset):
"""
The returned prediction is a 2D-array, each row corresponds to all the candidates,
and the ground-truth item poses the first.
Example: ground-truth items: [1, 2], 2 negative items for each instance: [[3,4], [5,6]]
predictions like: [[1,3,4], [2,5,6]]
"""
data.model.eval()
predictions = list()
js_div = list()
dis = list()
dl = DataLoader(data,
batch_size=self.eval_batch_size,
shuffle=False,
num_workers=self.num_workers,
collate_fn=data.collate_batch,
pin_memory=self.pin_memory)
for batch in tqdm(dl,
leave=False,
ncols=100,
mininterval=1,
desc='Predict'):
out_dict = data.model(utils.batch_to_gpu(batch, data.model.device))
predictions.extend(out_dict['prediction'].cpu().data.numpy())
if 'js' in out_dict:
js_div.extend(out_dict['js'].cpu().data.numpy())
if 'dis' in out_dict:
dis.extend(out_dict['dis'].cpu().data.numpy())
if len(js_div) > 0:
print('JS DIV:', np.mean(js_div))
np.save('../log/js', np.array(js_div))
if len(dis) > 0:
print('Interest Dis:', np.mean(dis))
return np.array(predictions)
def predict(self, model, data):
"""
The returned prediction is a 2D-array, each row corresponds to all the candidates,
and the ground-truth item poses the first.
Example: ground-truth items: [1, 2], 2 negative items for each instance: [[3,4], [5,6]]
predictions order: [[1,3,4], [2,5,6]]
"""
model.eval()
predictions = list()
dl = DataLoader(data, batch_size=self.eval_batch_size, shuffle=False, num_workers=self.num_workers,
collate_fn=data.collate_batch, pin_memory=self.pin_memory)
for batch in tqdm(dl, leave=False, ncols=100, mininterval=1, desc='Predict'):
prediction = model(utils.batch_to_gpu(batch))
predictions.extend(prediction.cpu().data.numpy())
return np.array(predictions)
def fit(self, model, data, epoch=-1):
gc.collect()
torch.cuda.empty_cache()
if model.optimizer is None:
model.optimizer = self._build_optimizer(model)
data.negative_sampling() # must sample before multi thread start
model.train()
loss_lst = list()
dl = DataLoader(data, batch_size=self.batch_size, shuffle=True, num_workers=self.num_workers,
collate_fn=data.collate_batch, pin_memory=self.pin_memory)
for batch in tqdm(dl, leave=False, desc='Epoch {:<3}'.format(epoch), ncols=100, mininterval=1):
batch = utils.batch_to_gpu(batch)
model.optimizer.zero_grad()
prediction = model(batch)
loss = model.loss(prediction)
loss.backward()
model.optimizer.step()
loss_lst.append(loss.detach().cpu().data.numpy())
return np.mean(loss_lst)