def _val(self):
self._model.eval()
val_loss = 0.0
val_acc = 0.0
# for plot center lloss
all_features, all_labels = [], []
with torch.no_grad():
for i, (images, targets) in tqdm(enumerate(self._val_loader),
total=len(self._val_loader),
leave=False):
images = images.cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
# compute output, measure accuracy and record loss
outputs, features = self._model(images)
loss = self._criterion(outputs, targets)
acc = accuracy(outputs, targets)[0]
val_loss += loss.item()
val_acc += acc.item()
all_features.append(features.data.cpu().numpy())
all_labels.append(targets.data.cpu().numpy())
i += 1
self._val_loss_list.append(val_loss / i)
self._val_acc_list.append(val_acc / i)
# plot center
all_features = np.concatenate(all_features, 0)
all_labels = np.concatenate(all_labels, 0)
self._plot_features(all_features, all_labels, prefix="test")
def _train(self):
self._model.train()
train_loss = 0.0
train_acc = 0.0
for i, (images, targets) in tqdm(enumerate(self._train_loader),
total=len(self._train_loader),
leave=False):
images = images.cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
# compute output, measure accuracy and record loss
outputs = self._model(images)
loss = self._criterion(outputs, targets)
acc = accuracy(outputs, targets)[0]
# acc = eval_metrics(targets, outputs, 2)[0]
train_loss += loss.item()
train_acc += acc.item()
# compute gradient and do SGD step
self._optimizer.zero_grad()
loss.backward()
self._optimizer.step()
i += 1
self._train_loss_list.append(train_loss / i)
self._train_acc_list.append(train_acc / i)
def _calc_acc_on_private_test_with_tta(self):
self._model.eval()
test_acc = 0.0
print("Calc acc on private test with tta..")
with torch.no_grad():
for idx in tqdm(
range(len(self._test_set)), total=len(self._test_set), leave=False
):
images, targets = self._test_set[idx]
targets = torch.LongTensor([targets])
if not isinstance(images, list):
images = [images]
images = torch.stack(images, 0)
images = images.cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
outputs = self._model(images)
outputs = F.softmax(outputs, 1)
outputs = torch.sum(outputs, 0)
outputs = torch.unsqueeze(outputs, 0)
acc = accuracy(outputs, targets)[0]
test_acc += acc.item()
test_acc = test_acc / (idx + 1)
print("Accuracy on private test with tta: {:.3f}".format(test_acc))
return test_acc
def _val(self):
self._model.eval()
val_loss = 0.0
val_acc = 0.0
os.system("rm -rf debug/*")
for i, (images, targets) in tqdm(enumerate(self._val_loader),
total=len(self._val_loader),
leave=False):
images = images.cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
# compute output, measure accuracy and record loss
outputs = self._model(images)
loss = self._criterion(outputs, targets)
acc = accuracy(outputs, targets)[0]
# acc = eval_metrics(targets, outputs, 2)[0]
val_loss += loss.item()
val_acc += acc.item()
# debug time
outputs = torch.squeeze(outputs, dim=0)
outputs = torch.argmax(outputs, dim=0)
tmp_image = torch.squeeze(images, dim=0)
print(tmp_image.shape)
tmp_image = tmp_image.cpu().numpy()
cv2.imwrite("debug/{}/{}.png".format(outputs, i), tmp_image)
i += 1
self._val_loss.append(val_loss / i)
self._val_acc.append(val_acc / i)
def _calc_acc_on_private_test_with_tta(self):
self._model.eval()
test_acc = 0.0
print("Calc acc on private test..")
transform = transforms.Compose([
transforms.ToPILImage(),
])
for idx in len(self._test_set):
image, label = self._test_set[idx]
with torch.no_grad():
for i, (images, targets) in tqdm(enumerate(self._test_loader),
total=len(self._test_loader),
leave=False):
# TODO: implement augment when predict
images = images.cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
outputs = self._model(images)
acc = accuracy(outputs, targets)[0]
test_acc += acc.item()
test_acc = test_acc / (i + 1)
print("Accuracy on private test: {:.3f}".format(test_acc))
return test_acc
def _train(self):
self._model.train()
train_loss = 0.0
train_acc = 0.0
# for plot center lloss
all_features, all_labels = [], []
for i, (images, targets) in tqdm(enumerate(self._train_loader),
total=len(self._train_loader),
leave=False):
images = images.cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
# compute output, measure accuracy and record loss
outputs, features = self._model(images)
loss = self._criterion(outputs, targets)
if self._current_epoch_num > 20:
loss_cent = self._criterion_cent(features, targets)
loss_cent *= self._configs["cweight"]
loss = loss + loss_cent
acc = accuracy(outputs, targets)[0]
train_loss += loss.item()
train_acc += acc.item()
# compute gradient and do SGD step
self._optimizer.zero_grad()
self._optimizer_cent.zero_grad()
loss.backward()
self._optimizer.step()
if self._current_epoch_num > 20:
# by doing so, weight_cent would not impact on the learning of centers
for param in self._criterion_cent.parameters():
param.grad.data *= 1.0 / self._configs["cweight"]
self._optimizer_cent.step()
all_features.append(features.data.cpu().numpy())
all_labels.append(targets.data.cpu().numpy())
i += 1
self._train_loss_list.append(train_loss / i)
self._train_acc_list.append(train_acc / i)
# plot center
all_features = np.concatenate(all_features, 0)
all_labels = np.concatenate(all_labels, 0)
self._plot_features(all_features, all_labels, prefix="train")
def _train(self):
self._model.train()
train_loss = 0.0
train_acc = 0.0
for i, (images, targets) in tqdm(enumerate(self._train_loader),
total=len(self._train_loader),
leave=False):
images = images.cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
# compute output, measure accuracy and record loss
outputs = self._model(images)
loss = self._criterion(outputs, targets)
acc = accuracy(outputs, targets)[0]
# acc = eval_metrics(targets, outputs, 2)[0]
train_loss += loss.item()
train_acc += acc.item()
# compute gradient and do SGD step
self._optimizer.zero_grad()
loss.backward()
self._optimizer.step()
# log
if i == 0:
grid = torchvision.utils.make_grid(images)
self._writer.add_image("images", grid, 0)
# self._writer.add_graph(self._model, images)
# self._writer.close()
if self._configs["little"] == 1:
mask = torch.squeeze(outputs, 0)
mask = mask.detach().cpu().numpy() * 255
mask = np.transpose(mask, (1, 2, 0)).astype(np.uint8)
cv2.imwrite(
os.path.join("debug",
"e{}.png".format(self._current_epoch_num)),
mask[..., 1],
)
i += 1
self._train_loss.append(train_loss / i)
self._train_acc.append(train_acc / i)
def _calc_acc_on_private_test(self):
self._model.eval()
test_acc = 0.0
print("Calc acc on private test..")
with torch.no_grad():
for i, (images, targets) in tqdm(
enumerate(self._test_loader), total=len(self._test_loader), leave=False
):
images = images.cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
outputs = self._model(images)
print(outputs.shape, outputs)
acc = accuracy(outputs, targets)[0]
test_acc += acc.item()
test_acc = test_acc / (i + 1)
print("Accuracy on private test: {:.3f}".format(test_acc))
return test_acc
def _calc_acc_on_private_test_with_tta(self):
self._model.eval()
test_acc = 0.0
print("Calc acc on private test with tta..")
f = open(
"private_test_log_{}_{}.txt".format(
self._configs["arch"], self._configs["model_name"]
),
"w",
)
with torch.no_grad():
for idx in tqdm(
range(len(self._test_set)), total=len(self._test_set), leave=False
):
images, targets = self._test_set[idx]
targets = torch.LongTensor([targets])
images = make_batch(images)
images = images.cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
outputs = self._model(images)
outputs = F.softmax(outputs, 1)
# outputs.shape [tta_size, 7]
outputs = torch.sum(outputs, 0)
outputs = torch.unsqueeze(outputs, 0)
# print(outputs.shape)
# TODO: try with softmax first and see the change
acc = accuracy(outputs, targets)[0]
test_acc += acc.item()
f.writelines("{}_{}\n".format(idx, acc.item()))
test_acc = test_acc / (idx + 1)
print("Accuracy on private test with tta: {:.3f}".format(test_acc))
f.close()
return test_acc
def _calc_acc_on_private_test(self):
self._model.eval()
test_acc = 0.0
print("Calc acc on private test..")
f = open("private_test_log.txt", "w")
with torch.no_grad():
for i, (images, targets) in tqdm(enumerate(self._test_loader),
total=len(self._test_loader),
leave=False):
images = images.cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
outputs = self._model(images)
#print(outputs.shape, outputs)
acc = accuracy(outputs, targets)[0]
print(f"test image {i+1} acc: {acc.item()}")
test_acc += acc.item()
f.writelines("{}_{}\n".format(i, acc.item()))
test_acc = test_acc / (i + 1)
print("Accuracy on private test: {:.3f}".format(test_acc))
f.close()
return test_acc
def _val(self):
self._model.eval()
val_loss = 0.0
val_acc = 0.0
with torch.no_grad():
for i, (images, targets) in tqdm(enumerate(self._val_loader),
total=len(self._val_loader),
leave=False):
images = images.cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
# compute output, measure accuracy and record loss
outputs = self._model(images)
loss = self._criterion(outputs, targets)
acc = accuracy(outputs, targets)[0]
val_loss += loss.item()
val_acc += acc.item()
i += 1
self._val_loss_list.append(val_loss / i)
self._val_acc_list.append(val_acc / i)
def _calc_acc_on_private_test(self):
self._model.eval()
test_acc = 0.
print('Calc acc on private test..')
f = open('private_test_log.txt', 'w')
with torch.no_grad():
for i, (images, targets) in tqdm(enumerate(self._test_loader),
total=len(self._test_loader),
leave=False):
images = images.cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
outputs = self._model(images)
print(outputs.shape, outputs)
acc = accuracy(outputs, targets)[0]
test_acc += acc.item()
f.writelines("{}_{}\n".format(i, acc.item()))
test_acc = test_acc / (i + 1)
print("Accuracy on private test: {:.3f}".format(test_acc))
print("Checkpoint saved at", self._checkpoint_path)
f.close()
return test_acc
