def test(self, data, label):
'''
Test the model with unnormalized ndarray test dataset.
@param data [list or ndarray, (batch, input_dim)]
@param label [list or ndarray, (batch, output_dim)]
'''
pred = self.predict(data)
pred = CUDA(torch.tensor(pred).long())
labels = CUDA(torch.tensor(label).long())
acc = pred.eq(labels.view_as(pred)).mean().item()
return acc
def fit(self, x=None, y=None, use_data_buf=True, normalize=True):
'''
Train the model either from external data or internal data buf.
@param x [list or ndarray, (batch, input_dim)]
@param y [list or ndarray, (batch, output_dim)]
'''
if use_data_buf:
x, y = self.data_buf.get_all()
train_loader, test_loader = self.make_dataloader(
x, y, normalize=normalize)
else: # use external data loader
train_loader, test_loader = self.make_dataloader(
x, y, normalize=normalize)
for epoch in range(self.n_epochs):
loss_train, acc_train = 0, 0
self.model.train()
for datas, labels in train_loader:
datas = CUDA(datas)
labels = CUDA(labels)
self.optimizer.zero_grad()
outputs = self.model(datas)
labels = torch.squeeze(labels)
loss = self.criterion(outputs, labels)
loss.backward()
self.optimizer.step()
loss_train += loss.item() * datas.shape[0] # sum of the loss
pred = outputs.argmax(dim=1, keepdim=True) # get the index
acc_train += pred.eq(labels.view_as(pred)).sum().item()
if self.save and (epoch + 1) % self.save_freq == 0:
self.save_model(self.save_path)
if (epoch + 1) % self.test_freq == 0:
loss_train /= len(train_loader.dataset)
acc_train /= len(train_loader.dataset)
loss_test, acc_test = -0.1234, -0.12
if len(test_loader) > 0:
loss_test, acc_test = self.test_model(test_loader)
print(
f"epoch[{epoch}/{self.n_epochs}],train l|acc: {loss_train:.4f}|{100.*acc_train:.2f}%, test l|acc {loss_test:.4f}|{100.*acc_test:.2f}%"
)
else:
print(
f"epoch[{epoch}/{self.n_epochs}],train l|acc: {loss_train:.4f}|{100.*acc_train:.2f}%"
)
if self.save:
self.save_model(self.save_path)
def test_model(self, testloader):
'''
Test the model with normalized test dataset.
@param test_loader [torch.utils.data.DataLoader]
'''
self.model.eval()
loss_test, acc_test = 0, 0
for datas, labels in testloader:
datas = CUDA(datas)
labels = CUDA(labels)
outputs = self.model(datas)
labels = torch.squeeze(labels)
loss = self.criterion(outputs, labels)
loss_test += loss.item() * datas.shape[0]
pred = outputs.argmax(
dim=1,
keepdim=True) # get the index of the max log-probability
acc_test += pred.eq(labels.view_as(pred)).sum().item()
loss_test /= len(testloader.dataset)
acc_test /= len(testloader.dataset)
self.model.train()
return loss_test, acc_test