def val(self):
self.model.eval()
tbar = tqdm(self.valid_queue)
with torch.no_grad():
for step, (input, target) in enumerate(tbar):
input = input.cuda(self.device)
target = target.cuda(self.device)
predicts = self.model(input)
val_loss = self.criterion(
predicts if self.aux else predicts[0], target)
self.val_loss_meter.update(val_loss.item())
self.metric_val.update(target, predicts[0])
if step % self.cfg['training']['report_freq'] == 0:
pixAcc, mIoU, dice = self.metric_val.get()
self.logger.info(
'Val loss: {}, pixAcc: {}, mIoU: {}, dice: {}'.format(
self.val_loss_meter.mloss(), pixAcc, mIoU, dice))
tbar.set_description(
'val loss: %.6f, pixAcc: %.3f, mIoU: %.6f, dice: %.6f'
% (self.val_loss_meter.mloss(), pixAcc, mIoU, dice))
# save images
# cause the predicts is a list [pred, aux_pred(may not)]
if len(predicts[0].shape) == 4: #
pred = predicts[0]
else:
pred = predicts
grid_image = store_images(input, pred, target)
self.writer.add_image('Val', grid_image, self.epoch)
# save in tensorboard scalars
pixAcc, mIoU, dice = self.metric_val.get()
cur_loss = self.val_loss_meter.mloss()
self.logger.info(
'Epoch {} Val loss: {}, pixAcc: {}, mIoU: {}, dice: {}'.format(
self.epoch, cur_loss, pixAcc, mIoU, dice))
self.writer.add_scalar('Val/Acc', pixAcc, self.epoch)
self.writer.add_scalar('Val/mIoU', mIoU, self.epoch)
self.writer.add_scalar('Val/dice', dice, self.epoch)
self.writer.add_scalar('Val/loss', cur_loss, self.epoch)
# for early-stopping
if self.best_loss > cur_loss or self.best_mIoU < mIoU:
self.patience = 0
else:
self.patience += 1
# Store best score
self.best_pixAcc = pixAcc if self.best_pixAcc < pixAcc else self.best_pixAcc
self.best_loss = cur_loss if self.best_loss > cur_loss else self.best_loss
if self.best_mIoU < mIoU:
self.best_mIoU = mIoU
self.best_dice_coeff = dice
self.save_best = True
def test(self):
self.model.eval()
predict_list = []
tbar = tqdm(self.test_queue)
with torch.no_grad():
for step, (input, target) in enumerate(tbar):
input = input.cuda(self.device)
if not isinstance(target, list) and not isinstance(
target, str):
target = target.cuda(self.device)
elif isinstance(target, str):
target = target.split('.')[0] + '_mask.tiff'
predicts = self.model(input)
# for cityscapes, voc, camvid
if not isinstance(target, list):
test_loss = self.criterion(
predicts if self.aux else predicts[0], target)
self.test_loss_meter.update(test_loss.item())
self.metric_test.update(target, predicts[0])
else: # for promise12
N = predicts[0].shape[0]
for i in range(N):
predict_list += [
torch.argmax(predicts[0], 1).cpu().numpy()[i]
]
# cause the predicts is a list [pred, aux_pred(may not)]
if len(predicts[0].shape) == 4: #
pred = predicts[0]
else:
pred = predicts
# save images
if not isinstance(target, list):
grid_image = store_images(input, pred, target)
self.writer.add_image('Test', grid_image, self.epoch)
pixAcc, mIoU, dice = self.metric_test.get()
self.logger.info(
'Test/loss: {}, pixAcc: {}, mIoU: {}, dice: {}'.format(
self.test_loss_meter.mloss(), pixAcc, mIoU, dice))
else:
predict_test(predict_list, target, self.save_path + '/test_rst')
def test(self, img_queue, split='val', desc=''):
self.model.eval()
predict_list = []
accuracy = 0
tbar = tqdm(img_queue)
create_exp_dir(desc, desc='=>Save prediction image on')
with torch.no_grad():
for step, (input, target) in enumerate(tbar):
input = input.cuda(self.device)
if not isinstance(target, list):
target = target.cuda(self.device)
predicts = self.model(input)
# for cityscapes, voc, camvid, test have label
if not isinstance(target, list):
test_loss = self.criterion(predicts[0], target)
self.loss_meter.update(test_loss.item())
self.metric.update(target, predicts[0])
if step % self.cfg['training']['report_freq'] == 0:
pixAcc, mIoU = self.metric.get()
self.logger.info(
'{} loss: {}, pixAcc: {}, mIoU: {}'.format(
split, self.loss_meter.mloss, pixAcc, mIoU))
tbar.set_description(
'loss: %.6f, pixAcc: %.3f, mIoU: %.6f' %
(self.loss_meter.mloss, pixAcc, mIoU))
accuracy += dice_coefficient(predicts[0].cpu(),
target.cpu())
else:
N = predicts[0].shape[0]
for i in range(N):
if self.args.crf: # use crf
predict = torch.argmax(predicts[0].cpu(), 1)[i]
predict = dense_crf(
np.array(input[i].cpu()).astype(np.uint8),
predict) > 0.5
img = Image.fromarray(
(predict * 255).astype(np.uint8))
file_name = os.path.split(target[i])[1]
file_name = file_name.split('.')[0] + '_mask.tif'
img.save(os.path.join(desc, file_name))
else:
img = Image.fromarray(
(torch.argmax(predicts[0].cpu(), 1)[i] *
255).numpy().astype(np.uint8))
file_name = os.path.split(target[i])[1]
file_name = file_name.split('.')[0] + '_mask.tif'
img.save(os.path.join(desc, file_name))
if desc == 'promise12': # for promise12, test have not label or have label to calc extra metric
predict_list += [
torch.argmax(predicts[0], dim=1).cpu().numpy()
]
print('==> accuracy: {}'.format(accuracy / len(img_queue)))
# cause the predicts is a list [pred, aux_pred(may not)]
if len(predicts[0].shape) == 4: #
pred = predicts[0]
else:
pred = predicts
# save images
if not isinstance(target, list) and not isinstance(target, str): #
grid_image = store_images(input, pred, target)
pixAcc, mIoU = self.metric.get()
self.logger.info('{}/loss: {}, pixAcc: {}, mIoU: {}'.format(
split, self.loss_meter.mloss, pixAcc, mIoU))
elif desc == 'promise12': # for promise12, test have not label
predict_test(predict_list, target,
self.save_path + '/{}_rst'.format(split))
# for promise12 metirc
if desc == 'promise12' and split == 'val':
val_list = [5, 15, 25, 35, 45]
dir = os.path.join(self.trainset.root_dir, self.trainset.base_dir,
'TrainingData')
biomedical_image_metric(predict_list, val_list, dir + '/')
