def data_augmentations(size=640):
train_augmentations = Compose([
Scale(size),
RandomHorizontallyFlip(0.5),
RandomVerticallyFlip(0.5),
AdjustContrast(0.25),
AdjustBrightness(0.25),
AdjustSaturation(0.25)
])
val_augmentations = Compose([Scale(size)])
return train_augmentations, val_augmentations
def __getitem__(self, item):
img = cv2.imread(self.image_dir[item])
img = np.array(img, dtype=np.uint8)
lbl = cv2.imread(self.label_dir[item])
lbl = np.array(lbl, dtype=np.uint8)
# Step 1
# Rotation -10~10
img, lbl = Rotation(img, lbl, 5)
# Scale 0.9~1.3
img, lbl = Scale(img, lbl)
# Translate -50~50
img, lbl = Translate(img, lbl, 25, 25)
# Flip
img, lbl = Flip(img, lbl, 0.5)
# Step2
selection_num = random.randint(0, 7)
if selection_num == 0:
img = White_Noise(img)
elif selection_num == 1:
img = Gray(img)
elif selection_num == 2:
img = Brightness(img)
elif selection_num == 3:
img = Contrast(img)
elif selection_num == 4:
img = Color(img)
elif selection_num == 5:
img = Equalization(img)
elif selection_num == 6:
img = Shapness(img)
elif selection_num == 7:
img = Power_Law(img)
img = cv2.resize(img, self.img_size)
lbl = cv2.resize(lbl, self.img_size)
lbl = process_gt_image(lbl)
# NHWC -> NCHW
img = img.transpose(2, 0, 1)
lbl = lbl.transpose(2, 0, 1)
img = torch.tensor(img).float()
lbl = torch.tensor(lbl).float()
return img, lbl
test_imgs_dir = os.path.join(data_dir, 'test')
train_df, test_df = get_train_test_df(data_dir)
immean = [0.315, 0.317, 0.214] # mean for wheat train images
imstd = [0.207, 0.209, 0.176] # std
from augmentations import Normalizer, Flip, Resizer, GaussBlur, AdjustBrightness, AdjustContrast, AdjustGamma, RandomRotate, Scale
from torchvision import transforms
train_transform = transforms.Compose([
Flip(),
GaussBlur(p=0.5),
AdjustContrast(p=0.3),
AdjustBrightness(p=0.3),
AdjustGamma(p=0.3),
Scale(),
Normalizer(mean=immean, std=imstd),
Resizer(1280)
])
test_transform = transforms.Compose(
[Scale(), Normalizer(mean=immean, std=imstd),
Resizer(1280)])
train_dataset = WheatDataset(train_df,
train_imgs_dir,
train_transform,
mixup=True)
test_dataset = WheatDataset(test_df,
test_imgs_dir,
test_transform,
rgb[:, :, 2] = b / 255.0
return rgb
def encode_segmap(self, mask):
# Put all void classes to zero
for _voidc in self.void_classes:
mask[mask == _voidc] = self.ignore_index
for _validc in self.valid_classes:
mask[mask == _validc] = self.class_map[_validc]
return mask
if __name__ == "__main__":
import matplotlib.pyplot as plt
augmentations = Compose([Scale(2048), RandomRotate(10), RandomHorizontallyFlip(0.5)])
local_path = "/datasets01/cityscapes/112817/"
dst = CityscapesLoader(local_path, is_transform=True, augmentations=augmentations)
bs = 4
trainloader = data.DataLoader(dst, batch_size=bs, num_workers=0)
for i, data_samples in enumerate(trainloader):
imgs, labels = data_samples
import pdb
pdb.set_trace()
imgs = imgs.numpy()[:, ::-1, :, :]
imgs = np.transpose(imgs, [0, 2, 3, 1])
f, axarr = plt.subplots(bs, 2)
for j in range(bs):
axarr[j][0].imshow(imgs[j])
data = [item[0] for item in batch]
target = [item[1] for item in batch]
return [data, target]
# Leave code for debugging purposes
# import ptsemseg.augmentations as aug
if __name__ == '__main__':
from augmentations import Compose, RandomHorizontallyFlip, RandomVerticallyFlip, Scale
from augmentations import AdjustContrast, AdjustBrightness, AdjustSaturation
import matplotlib.pyplot as plt
bs = 4
augmentations = Compose([
Scale(512),
RandomHorizontallyFlip(0.5),
RandomVerticallyFlip(0.5),
AdjustContrast(0.25),
AdjustBrightness(0.25),
AdjustSaturation(0.25)
])
dst = CustomDataset(root_dir='../dataset/train',
augmentations=augmentations)
trainloader = DataLoader(dst,
batch_size=bs,
collate_fn=custom_collate,
pin_memory=True)
criterion = nn.MSELoss()
parser.add_argument('--img_cols', type=int, default=64, help='resized image width')
parser.add_argument('--img_rows', type=int, default=64, help='resized image height')
parser.add_argument('--workers', type=int, default=4, help='Data loader workers')
args = parser.parse_args()
random.seed(args.seed)
torch.manual_seed(args.seed)
plt.switch_backend('agg') # Allow plotting when running remotely
save_epoch = 100 # save log images per save_epoch
# 02 rotation + flip augmentation option
# Setup Augmentations
data_aug_tr = Compose([Scale(args.img_cols), # resize longer side of an image to the defined size
CenterPadding([args.img_rows, args.img_cols]), # zero pad remaining regions
RandomHorizontallyFlip(), # random horizontal flip
RandomRotate(180)]) # ramdom rotation
data_aug_te = Compose([Scale(args.img_cols),
CenterPadding([args.img_rows, args.img_cols])])
result_path = 'results_' + args.dataset
if not os.path.exists(result_path):
os.makedirs(result_path)
outimg_path = "./img_log_" + args.dataset
if not os.path.exists(outimg_path):
os.makedirs(outimg_path)
f_loss = open(os.path.join(result_path, "log_loss.txt"), 'w')
mask = np.ones(lbl.shape[:2], dtype=np.uint8) * self.ignore_index
for _validc in self.valid_classes:
clr = self.class_map[_validc]
mask_bool = np.array(lbl == clr).all(axis=2)
mask[mask_bool] = _validc
return mask
if __name__ == "__main__":
img = imread(
'/home/robotics/rssrai2019/data_preprocessed1/val/images/0/20160421_L1A0001537716_55.png'
)
import matplotlib.pyplot as plt
augmentations = Compose(
[Scale(2048),
RandomRotate(10),
RandomHorizontallyFlip(0.5)])
local_path = "/home/robotics/ma_thesis_data/lgln_3city/dataset/C1_20cm"
dst = threeCityLoader(local_path,
is_transform=True,
augmentations=augmentations)
bs = 4
trainloader = data.DataLoader(dst, batch_size=bs, num_workers=0)
for i, data_samples in enumerate(trainloader):
imgs, labels = data_samples
import pdb
pdb.set_trace()
imgs = imgs.numpy()[:, ::-1, :, :]