def __getitem__(self, idx):
# print(os.getcwd())
# print(path_test)
path_test = os.path.join(os.getcwd(),'Chexpert/data/nas',self._image_paths[idx])
# print(path_test)
# print(self._image_paths[idx])
# testing = cv2.imread()
# image = cv2.imread(self._image_paths[idx], 0)
image = cv2.imread(path_test, 0)
# print(image)
image = Image.fromarray(image)
if self._mode == 'train':
image = GetTransforms(image, type=self.cfg.use_transforms_type)
image = np.array(image)
if self.cfg.use_equalizeHist:
image = cv2.equalizeHist(image)
image = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB).astype(np.float32)
# changed the resolution of the img from 512 to 640 all around to see if produces better results
# chaged the interpolation method from INTER_LINEAR to INTER_AREA which resamples using pixel area relation.
# possibly better results with inter_area if no zooming is applied
if self.cfg.fix_ratio:
image = self._fix_ratio(image)
else:
image = cv2.resize(image, dsize=(self.cfg.width, self.cfg.height),
interpolation=cv2.INTER_AREA)
# chaged the smoothing of the image:
# changed from gaussian_blur to bilateralFiltering, which is a better version, which simplifies the img but preserve edges
# it can run a little slower than gaussian_blur since it a more complex process, but might improve results
# left the if statement unchanged, just to filter the same imgs than before.
if self.cfg.gaussian_blur > 0:
image = cv2.bilateralFilter(image, 9, 75, 75)
# normalization
image -= self.cfg.pixel_mean
# vgg and resnet do not use pixel_std, densenet and inception use.
if self.cfg.use_pixel_std:
image /= self.cfg.pixel_std
# normal image tensor : H x W x C
# torch image tensor : C X H X W
image = image.transpose((2, 0, 1))
labels = np.array(self._labels[idx]).astype(np.float32)
path = self._image_paths[idx]
if self._mode == 'train' or self._mode == 'dev':
return (image, labels)
elif self._mode == 'test':
return (image, path)
elif self._mode == 'heatmap':
return (image, path, labels)
else:
raise Exception('Unknown mode : {}'.format(self._mode))
def __getitem__(self, idx):
# print(os.getcwd())
# print(path_test)
path_test = os.path.join(os.getcwd(), 'Chexpert/data/nas',
self._image_paths[idx])
# print(path_test)
# print(self._image_paths[idx])
# testing = cv2.imread()
# image = cv2.imread(self._image_paths[idx], 0)
image = cv2.imread(path_test, 0)
# print(image)
image = Image.fromarray(image)
# print('working,----------------------------------')
if self._mode == 'train':
image = GetTransforms(image, type=self.cfg.use_transforms_type)
image = np.array(image)
if self.cfg.use_equalizeHist:
image = cv2.equalizeHist(image)
image = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB).astype(np.float32)
if self.cfg.fix_ratio:
image = self._fix_ratio(image)
else:
image = cv2.resize(image,
dsize=(self.cfg.width, self.cfg.height),
interpolation=cv2.INTER_LINEAR)
if self.cfg.gaussian_blur > 0:
image = cv2.GaussianBlur(
image, (self.cfg.gaussian_blur, self.cfg.gaussian_blur), 0)
# normalization
image -= self.cfg.pixel_mean
# vgg and resnet do not use pixel_std, densenet and inception use.
if self.cfg.use_pixel_std:
image /= self.cfg.pixel_std
# normal image tensor : H x W x C
# torch image tensor : C X H X W
image = image.transpose((2, 0, 1))
labels = np.array(self._labels[idx]).astype(np.float32)
path = self._image_paths[idx]
if self._mode == 'train' or self._mode == 'dev':
return (image, labels)
elif self._mode == 'test':
return (image, path)
elif self._mode == 'heatmap':
return (image, path, labels)
else:
raise Exception('Unknown mode : {}'.format(self._mode))