def __init__(self, image_size, mean_std):
self.transform1 = T.Compose([
T.RandomResizedCrop(image_size,
scale=(0.08, 1.0),
ratio=(3.0 / 4.0, 4.0 / 3.0),
interpolation=Image.BICUBIC),
T.RandomHorizontalFlip(p=0.5),
T.RandomApply([T.ColorJitter(0.4, 0.4, 0.2, 0.1)], p=0.8),
T.RandomGrayscale(p=0.2),
T.GaussianBlur(kernel_size=image_size // 20 * 2 + 1,
sigma=(0.1, 2.0)),
# simclr paper gives the kernel size. Kernel size has to be odd positive number with torchvision
T.ToTensor(),
T.Normalize(*mean_std)
])
self.transform2 = T.Compose([
T.RandomResizedCrop(image_size,
scale=(0.08, 1.0),
ratio=(3.0 / 4.0, 4.0 / 3.0),
interpolation=Image.BICUBIC),
T.RandomHorizontalFlip(p=0.5),
T.RandomApply([T.ColorJitter(0.4, 0.4, 0.2, 0.1)], p=0.8),
T.RandomGrayscale(p=0.2),
# transforms.RandomApply([GaussianBlur(kernel_size=int(0.1 * image_size))], p=0.1),
T.RandomApply([
T.GaussianBlur(kernel_size=image_size // 20 * 2 + 1,
sigma=(0.1, 2.0))
],
p=0.1),
T.RandomApply([Solarization()], p=0.2),
T.ToTensor(),
T.Normalize(*mean_std)
])
def _build_model(self, blur, brightness, need_ray, need_speed):
assert self.obs_shapes[0] == (84, 84, 3)
assert self.obs_shapes[1] == (2,) # ray
assert self.obs_shapes[2] == (8,) # vector
self.need_ray = need_ray
self.need_speed = need_speed
self.conv = m.ConvLayers(84, 84, 3, 'simple',
out_dense_n=64, out_dense_depth=2)
self.dense = m.LinearLayers(self.conv.output_size,
dense_n=64, dense_depth=1)
self.rnn = m.GRU(64 + self.c_action_size, 64, 1)
if blur != 0:
self.blurrer = m.Transform(T.GaussianBlur(blur, sigma=blur))
else:
self.blurrer = None
self.brightness = m.Transform(T.ColorJitter(brightness=(brightness, brightness)))
cropper = torch.nn.Sequential(
T.RandomCrop(size=(50, 50)),
T.Resize(size=(84, 84))
)
self.random_transformers = T.RandomChoice([
m.Transform(SaltAndPepperNoise(0.2, 0.5)),
m.Transform(GaussianNoise()),
m.Transform(T.GaussianBlur(9, sigma=9)),
m.Transform(cropper)
])
def __init__(self, capacity=10000, first_trans=None, second_trans=None):
self.images = []
self.capacity = capacity
self.first_trans = first_trans
self.second_trans = second_trans
if self.first_trans is None:
self.first_trans = transforms.Compose([
transforms.RandomResizedCrop(320),
transforms.RandomApply(
[transforms.ColorJitter(0.8, 0.8, 0.8, 0.2)], p=0.8),
transforms.RandomGrayscale(p=0.2),
transforms.GaussianBlur(33),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
if self.second_trans is None:
self.second_trans = transforms.Compose([
transforms.RandomResizedCrop(320),
transforms.RandomApply(
[transforms.ColorJitter(0.8, 0.8, 0.8, 0.2)], p=0.8),
transforms.RandomGrayscale(p=0.2),
transforms.GaussianBlur(33),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
def _build_model(self, blur, brightness, ray_random, need_speed):
assert self.obs_shapes[0] == (84, 84, 3)
assert self.obs_shapes[1] == (1442, ) # ray (1 + 360 + 360) * 2
assert self.obs_shapes[2] == (6, ) # vector
self.ray_random = ray_random
self.need_speed = need_speed
if blur != 0:
self.blurrer = m.Transform(T.GaussianBlur(blur, sigma=blur))
else:
self.blurrer = None
self.brightness = m.Transform(
T.ColorJitter(brightness=(brightness, brightness)))
self.ray_index = []
for i in reversed(range(RAY_SIZE // 2)):
self.ray_index.append((i * 2 + 1) * 2)
self.ray_index.append((i * 2 + 1) * 2 + 1)
for i in range(RAY_SIZE // 2):
self.ray_index.append((i * 2 + 2) * 2)
self.ray_index.append((i * 2 + 2) * 2 + 1)
self.conv = m.ConvLayers(84,
84,
3,
'simple',
out_dense_n=64,
out_dense_depth=2)
self.ray_conv = m.Conv1dLayers(RAY_SIZE,
2,
'default',
out_dense_n=64,
out_dense_depth=2)
self.vis_ray_dense = m.LinearLayers(self.conv.output_size +
self.ray_conv.output_size,
dense_n=64,
dense_depth=1)
self.rnn = m.GRU(64 + self.c_action_size, 64, 1)
cropper = torch.nn.Sequential(T.RandomCrop(size=(50, 50)),
T.Resize(size=(84, 84)))
self.random_transformers = T.RandomChoice([
m.Transform(SaltAndPepperNoise(0.2, 0.5)),
m.Transform(GaussianNoise()),
m.Transform(T.GaussianBlur(9, sigma=9)),
m.Transform(cropper)
])
def augment(size, dir):
assert dir != None, 'Augment requires directory argument'
labels = [label for label in os.listdir(dir)]
train_augment = transforms.Compose([
transforms.ColorJitter(0.5, 0.5, 0.5),
transforms.RandomAffine(degrees=10),
transforms.GaussianBlur(5),
transforms.RandomPerspective(distortion_scale=0.2)
])
for label in labels:
subdir = os.path.join(dir, label)
output = os.path.join(subdir, 'augmented')
if os.path.isdir(output):
shutil.rmtree(output)
images = glob(os.path.join(subdir, '*'))
num_images = len(images)
if num_images >= size:
print(f'\'{subdir}\' already has enough images ({num_images})')
continue
os.makedirs(output)
for i in tqdm(range(size - num_images)):
img_name = random.choice(images)
image = Image.open(img_name)
transformed = train_augment(image)
transformed = transformed.resize((224, 224))
transformed.save(
os.path.join(output, f'{i}_{img_name.split(os.sep)[-1]}'))
def __init__(self, input_size):
self.training = transforms.Compose([
transforms.Resize([input_size, input_size]),
transforms.ColorJitter(brightness=0.25,
contrast=0.25,
saturation=0.25,
hue=0.1),
transforms.RandomAffine(
degrees=15,
translate=None,
scale=None,
shear=15,
resample=0,
fillcolor=0,
),
transforms.GaussianBlur(kernel_size=3, sigma=(0.001, 2.0)),
transforms.RandomHorizontalFlip(p=0.5),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
self.validation = transforms.Compose([
transforms.Resize([input_size, input_size]),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
def get_transform_1(train):
transforms_list = []
transforms_list.append(transforms.RandomApply(torch.nn.ModuleList([transforms.GaussianBlur(3)]), p=0.5))
transforms_list.append(transforms.RandomApply(torch.nn.ModuleList([transforms.ColorJitter(brightness = 0.1)]), p=0.5))
transforms_list.append(transforms.RandomApply(torch.nn.ModuleList([transforms.ColorJitter(contrast = 0.5)]), p=0.5))
return transforms.Compose(transforms_list)
def __init__(self, train=False, seed=1, resize=510, centerCrop=448):
self.transformations = transforms.Compose([
transforms.Resize(resize),
transforms.CenterCrop(centerCrop),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
self.random_transformations = transforms.Compose([
transforms.RandomRotation(90, expand=False, fill=None),
transforms.RandomHorizontalFlip(p=0.5),
transforms.RandomVerticalFlip(p=0.5),
transforms.GaussianBlur(11, sigma=(0.1, 2.0)),
torch.nn.Dropout2d(0.1),
transforms.RandomErasing(p=0.25)
])
self.train = train
# pandas dataframe that stores image names and labels
self.df = pd.read_csv('./data/train.csv').sample(frac=1,
random_state=seed)
test_section = int(0.8 * len(self.df))
if train:
self.df = self.df.iloc[:test_section]
self.df_transformed = self.df.copy()
self.df = pd.concat([self.df, self.df_transformed])
else:
self.df = self.df.iloc[test_section:]
def __getitem__(self, idx):
(path, label) = self.data[idx]
temp_img = Image.open(path)
picture_h_w = self.image_size
if self.data_augumentation:
result = transforms.Compose([
transforms.CenterCrop((self.crop_size, self.crop_size)),
# transforms.RandomResizedCrop(self.crop_size, scale=(0.1, 1), ratio=(0.5, 2)),
transforms.Resize((picture_h_w, picture_h_w)),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=0.5, hue=0.5, contrast=0.5),
transforms.RandomRotation(20),
transforms.GaussianBlur((5, 5), sigma=(0.1, 2.0)),
transforms.ToTensor(),
transforms.Normalize([0.7906623],
[0.16963087]) #[0.7906623] [0.16963087]
])(temp_img)
else:
result = transforms.Compose([
transforms.CenterCrop((self.crop_size, self.crop_size)),
transforms.Resize((picture_h_w, picture_h_w)),
transforms.ToTensor(),
transforms.Normalize([0.7906623],
[0.16963087]) #[0.7906623] [0.16963087]
])(temp_img)
return {'result': result, 'label': torch.LongTensor([label])}
def __init__(self, df, image_dir, targets=['Type1'], is_train=True):
#for t in targets:
# assert t in ['Type1', 'Type2']
self.df = df
self.image_dir = image_dir
self.targets = targets
self.transform = T.Compose([
T.RandomApply(
[
#transforms.RandomCrop(32, padding=4),
#T.RandomHorizontalFlip(p=0.5)
#T.RandomRotation(degrees, interpolation=<InterpolationMode.NEAREST: 'nearest'>, expand=False, center=None, fill=0, resample=None)
# T.RandomResizedCrop(size, scale=(0.08, 1.0), ratio=(0.75, 1.3333333333333333), interpolation=<InterpolationMode.BILINEAR: 'bilinear'>)
T.ColorJitter(
brightness=0.1, contrast=0.1, saturation=0.1, hue=0),
T.GaussianBlur(11, sigma=(0.1, 2.0))
#T.ToTensor(),
],
p=0.9),
T.RandomHorizontalFlip(),
#T.RandomErasing(p=0.5, scale=(0.02, 0.07), ratio=(0.3, 3.3), value=255, inplace=False),
T.RandomRotation(20, expand=False, center=None, fill=255)
#T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
self.is_train = is_train
def __init__(self, args):
if args.encoder == "clip":
normalize = transforms.Normalize(
mean=[0.48145466, 0.4578275, 0.40821073],
std=[0.26862954, 0.26130258, 0.27577711])
else:
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
randcrop = transforms.RandomResizedCrop(224)
randflip = transforms.RandomHorizontalFlip()
jitter = transforms.ColorJitter(0.4, 0.4, 0.2, 0.1)
randjitter = transforms.RandomApply([jitter], p=0.5)
blur = transforms.GaussianBlur(kernel_size=23, sigma=[1, 5])
randblur = transforms.RandomApply([blur], p=0.4)
noise = GaussianNoise(std=[0.1, 0.5], fixed=False)
randnoise = transforms.RandomApply([noise], p=0.4)
mask = RandomMask(percent_missing=[0.25, 0.50], fixed=False)
randmask = transforms.RandomApply([mask], p=0.1)
self.transform_common = transforms.Compose([randcrop, randflip])
self.transform_clean = transforms.Compose(
[transforms.ToTensor(), normalize])
self.transform_distort = transforms.Compose([
randjitter,
transforms.ToTensor(), randblur, randnoise, randmask, normalize
])
def build_data_aug(size, mode, resnet=False, resizepad=False):
if resnet:
norm_tfm = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
else:
norm_tfm = transforms.Normalize(0.5, 0.5)
if resizepad:
resize_tfm = ResizePad(imgH=size[0], imgW=size[1])
else:
resize_tfm = transforms.Resize(size,
interpolation=InterpolationMode.BICUBIC)
if mode == 'train':
return transforms.Compose([
WeightedRandomChoice([
# transforms.RandomHorizontalFlip(p=1),
transforms.RandomRotation(degrees=(-10, 10),
expand=True,
fill=255),
transforms.GaussianBlur(3),
Dilation(3),
Erosion(3),
transforms.Resize((size[0] // 3, size[1] // 3),
interpolation=InterpolationMode.NEAREST),
Underline(),
KeepOriginal(),
]),
resize_tfm,
transforms.ToTensor(),
norm_tfm
])
else:
return transforms.Compose(
[resize_tfm, transforms.ToTensor(), norm_tfm])
def __init__(
self,
input_height: int = 224,
resize_height: Optional[Union[int, str]] = 'default',
gaussian_blur: bool = True,
jitter_strength: float = 1.,
normalize: Optional[Callable[[Tensor], Tensor]] = None) -> None:
self.jitter_strength = jitter_strength
self.input_height = input_height
self.gaussian_blur = gaussian_blur
self.normalize = normalize
if resize_height is None:
self.resize_height = self.input_height
elif resize_height == 'default':
self.resize_height = int(self.input_height +
0.1 * self.input_height)
else:
assert isinstance(resize_height, int)
self.resize_height = resize_height
self.color_jitter = transforms.ColorJitter(0.8 * self.jitter_strength,
0.8 * self.jitter_strength,
0.8 * self.jitter_strength,
0.2 * self.jitter_strength)
data_transforms = [
transforms.RandomResizedCrop(size=self.input_height),
transforms.RandomHorizontalFlip(p=0.5),
transforms.RandomApply([self.color_jitter], p=0.8),
transforms.RandomGrayscale(p=0.2)
]
if self.gaussian_blur:
kernel_size = int(0.1 * self.input_height)
if kernel_size % 2 == 0:
kernel_size += 1
data_transforms.append(
transforms.RandomApply(
[transforms.GaussianBlur(kernel_size=kernel_size)], p=0.5))
data_transforms = transforms.Compose(data_transforms)
self.online_eval_transform = transforms.Compose([
transforms.Resize(self.resize_height),
transforms.CenterCrop(self.input_height),
])
if normalize is None:
self.final_transform = transforms.ToTensor()
else:
self.final_transform = transforms.Compose(
[transforms.ToTensor(), normalize])
self.simclr_transform = transforms.Compose(
[data_transforms, self.final_transform])
self.online_eval_transform = transforms.Compose(
[self.online_eval_transform, self.final_transform])
def __init__(self, args, fixed_distortion=None, epoch=None):
if args.encoder == "clip":
normalize = transforms.Normalize(
mean=[0.48145466, 0.4578275, 0.40821073],
std=[0.26862954, 0.26130258, 0.27577711])
else:
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
if fixed_distortion is not None:
distortion = fixed_distortion
elif args.distortion == "squaremask":
distortion = SquareMask(length=args.length,
offset=args.offset,
fixed=args.fixed_mask)
elif args.distortion == "randommask":
distortion = RandomMask(percent_missing=convnoise(
args.percent_missing, epoch),
fixed=args.fixed_mask)
elif args.distortion == "gaussiannoise":
distortion = GaussianNoise(std=convnoise(args.std, epoch),
fixed=args.fixed_mask)
elif args.distortion == "gaussianblur":
distortion = transforms.GaussianBlur(kernel_size=args.kernel_size,
sigma=args.sigma)
self.transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(), distortion, normalize
])
self.distortion = distortion
def create_cell_dataset(
folders,
scale=0.125,
crop=(128, 128),
n_streams=12,
batch_size=16,
cell_prominence_min=0.4,
cell_prominence_max=float('inf'),
):
def validate_crop(tensor):
mean = tensor[1:].sum(dim=0).mean()
if cell_prominence_min < mean < cell_prominence_max:
return tensor[:1]
return None
images = Dataset.ImageIterableDataset(
folders,
transforms.Compose((
transforms.GaussianBlur(3, (.01, 1.)),
data_aug.flip_scale_pipeline(scale, ),
#data_aug.pipeline(scale, degrees=0, noise_p=0.01),
)),
Dataset.CropGenerator(crop, validate_crop=validate_crop),
n_streams=n_streams, # large memory impact
indices=[*range(0, 204), *range(306, 2526)],
)
return DataLoader(images, batch_size=batch_size, drop_last=True)
def load_dataset(args, mode):
if mode == 'val':
data_transforms = transforms.Compose([
transforms.Resize((512, 512)),
transforms.ToTensor(),
# transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
else:
data_transforms = transforms.Compose([
transforms.ColorJitter(0.5, 0.5, 0.5, 0.5),
transforms.RandomGrayscale(p=0.05),
transforms.RandomApply([transforms.GaussianBlur(kernel_size=3)],
p=0.5),
transforms.RandomRotation((-90, 90)),
transforms.RandomHorizontalFlip(),
transforms.Resize((512, 512)),
transforms.ToTensor(),
# transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
dataset = datasets.ImageFolder(os.path.join(args.data_dir, mode),
data_transforms)
dataloader = DataLoader(dataset,
batch_size=args.train_batch_size,
shuffle=True,
num_workers=16)
data_set_size = len(dataset)
return dataloader, data_set_size
def __init__(self, batch_size):
super().__init__()
self.batch_size = batch_size
self.test_transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
# random crop, color jitter etc
self.train_transform = transforms.Compose([
transforms.Resize(256),
transforms.RandomResizedCrop(224, scale=(0.2, 1.)),
transforms.RandomApply(
[
transforms.ColorJitter(0.4, 0.4, 0.4,
0.1) # not strengthened
],
p=0.8),
transforms.RandomGrayscale(p=0.2),
transforms.RandomApply([transforms.GaussianBlur([1, 1])],
p=0.5), # perhaps this blur is too much
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.491, 0.482, 0.446],
std=[0.247, 0.243, 0.261]),
])
self.train_dataset = MSCOCO(train=True,
image_transforms=self.train_transform)
self.test_dataset = MSCOCO(train=False,
image_transforms=self.test_transform)
def get_reduced_transform(resize, size, blur, means, stds,
adaptive_thresholding):
"""Reduced transforms applied to original inputs
Arguments:
resize {int} -- resize before cropping
size {int} -- expected size
blur {float} -- sigma of the gaussian blur
means {list} -- pixel-wise means
stds {list} -- pixel-wise stds
"""
tfs = []
tfs.append(transforms.Resize(size=resize))
tfs.append(transforms.RandomCrop(size))
# gaussian blur
if blur is not None:
tfs.append(transforms.GaussianBlur(5, blur))
if adaptive_thresholding:
tfs.append(AdaptiveGaussianThreshold())
tfs.append(transforms.ToTensor())
if means is not None and stds is not None:
tfs.append(transforms.Normalize(means, stds))
return transforms.Compose(tfs)
def create_dataloaders(self):
self.data_val = self.dataset_val
examples_to_discard = len(self.data_val) - self.size_experiment
self.data_val_without_blur,subdataset_original_val_to_discard=\
random_split( self.data_val,
( self.size_experiment,examples_to_discard)
)
self.data_val_with_blur = deepcopy(self.data_val_without_blur)
new_transform = transforms.Compose([
transforms.Resize((32, 32), Image.BILINEAR),
transforms.GaussianBlur((13), sigma=(10, 20)),
transforms.ToTensor(),
transforms.Normalize(0.5, 0.5)
])
self.data_val_with_blur.dataset.dataset.transform = new_transform
self.data_val_without_blur = DataLoader(
dataset=self.data_val_without_blur,
batch_size=self.config.batch_size,
num_workers=self.config.NUM_WORKERS,
pin_memory=True,
shuffle=False,
)
self.data_val_with_blur = DataLoader(
dataset=self.data_val_with_blur,
batch_size=self.config.batch_size,
num_workers=self.config.NUM_WORKERS,
pin_memory=True,
shuffle=False,
)
def noise_pipeline(p, sigma):
return random_apply(
[
transforms.GaussianBlur(3, sigma=sigma),
GaussianNoise(0.0, (0.02, 0.08))
],
p=p,
)
def all_transforms():
return transforms.RandomApply(
[
# transforms.RandomHorizontalFlip(p=1.0),
transforms.GaussianBlur(kernel_size=5, sigma=10.0),
# CustomRescalingTransform([-5, 5], [0.85, 0.95], [0.85, 0.95])
],
p=1.0)
def get_transforms(baseSize=(256, 256), targetSize=(224, 224), advance=False):
if not advance:
train_transformations = transforms.Compose([
transforms.Resize(baseSize),
# transforms.RandomResizedCrop(targetSize),
transforms.RandomCrop(targetSize),
# transforms.RandomChoice([
# transforms.Resize(targetSize),
# transforms.RandomResizedCrop(targetSize),
# transforms.RandomCrop(targetSize),
# transforms.CenterCrop(targetSize)
# ]),
transforms.RandomApply([
transforms.ColorJitter(0.05, 0.05, 0.05, 0.05),
transforms.GaussianBlur(5),
]),
transforms.RandomHorizontalFlip(),
# transforms.RandomRotation(10),
transforms.RandomAffine(10),
transforms.RandomGrayscale(),
# Imgaug(),
transforms.ToTensor(), # 转成0.～1.
# transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)), # -1.~1.
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
transforms.RandomErasing()
])
else:
train_transformations = transforms.Compose([
transforms.Resize(baseSize),
# transforms.RandomResizedCrop(targetSize),
transforms.RandomCrop(targetSize),
# transforms.RandomChoice([
# transforms.Resize(targetSize),
# transforms.RandomResizedCrop(targetSize),
# transforms.RandomCrop(targetSize),
# transforms.CenterCrop(targetSize)
# ]),
Imgaug(),
transforms.ToTensor(), # 转成0.～1.
# transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)), # -1.~1.
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
test_transformations = transforms.Compose([
transforms.Resize(baseSize),
transforms.CenterCrop(targetSize), # 0.936
# transforms.Resize(targetSize), # 0.941
transforms.ToTensor(), # 转成0.～1.
# transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)), # -1.~1.
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
return train_transformations, test_transformations
def __call__(self, data):
label, input = data['label'], data['input']
t = transforms.GaussianBlur(5, sigma=(0.1, 2.0))
if random.random() < 0.5:
input = t(input)
data = {'label': label, 'input': input}
return data
def __call__(self, image, bboxes):
if type(image) == torch.Tensor:
image = T.ToPILImage()(image)
assert type(image) in (
JpegImageFile,
Image), 'R.H.FLIP: Problems with type of the image: {}'.format(
type(image))
image = T.GaussianBlur(kernel_size=3, sigma=(0.1, 2.0))(
image) # TODO: what about contrast and hue?
return image, bboxes
def __init__(self, batch_size, val_split=0.2):
super().__init__()
self.batch_size = batch_size
self.val_split = val_split
self.test_transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
# random crop, color jitter etc
self.train_transform = transforms.Compose([
transforms.Resize(256),
transforms.RandomResizedCrop(224, scale=(0.2, 1.)),
transforms.RandomApply(
[
transforms.ColorJitter(0.4, 0.4, 0.4,
0.1) # not strengthened
],
p=0.8),
transforms.RandomGrayscale(p=0.2),
transforms.RandomApply([transforms.GaussianBlur([1, 1])], p=0.5),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.491, 0.482, 0.446],
std=[0.247, 0.243, 0.261]),
])
# train and test transforms needed
self.questions_file = "/data5/shashank2000/final_json/OpenEnded_mscoco_train2014_questions.json"
self.answers_file = "/data5/shashank2000/final_json/mscoco_train2014_annotations.json"
self.coco_loc = "/mnt/fs0/datasets/mscoco/train2014"
# Enter path for questions_file, answers_file and directory for COCO images that correspond to these
#self.questions_file = "../../datasets/questions.json"
#self.answers_file = "../../datasets/answers.json"
# copy and then unarchive instead?
#self.coco_loc = "../../datasets/train2014"
# do something about self.dims?
self.train_dataset = JeopardyDataset(self.questions_file,
self.answers_file,
self.coco_loc,
self.train_transform,
train=True)
self.test_dataset = JeopardyDataset(
self.questions_file,
self.answers_file,
self.coco_loc,
self.test_transform,
word2idx=self.train_dataset.word2idx,
train=False)
self.vl = self.get_vocab_length()
def __init__(self, args, mode='train', downstream=False):
if mode == 'train':
data_list = self.train_list
else:
data_list = self.test_list
self.targets = []
self.data = []
self.args = args
self.downstream = downstream
for file_name, checksum in data_list:
file_path = os.path.join(args.base_dir, file_name)
with open(file_path, 'rb') as f:
entry = pickle.load(f, encoding='latin1')
self.data.append(entry['data'])
if 'labels' in entry:
self.targets.extend(entry['labels'])
else:
self.targets.extend(entry['fine_labels'])
self.data = np.vstack(self.data).reshape(-1, 3, 32, 32)
self.data = self.data.transpose((0, 2, 3, 1)) # convert to HWC
self.transform1 = transforms.Compose([
transforms.RandomResizedCrop(self.args.img_size, scale=(0.2, 1.0)),
transforms.RandomHorizontalFlip(),
transforms.RandomApply([transforms.ColorJitter(0.4, 0.4, 0.4, 0.1)], p=0.8),
transforms.RandomGrayscale(0.2),
transforms.GaussianBlur(kernel_size=int(self.args.img_size * 0.1), sigma=(0.1, 2.0)),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
self.transform2 = transforms.Compose([
transforms.RandomResizedCrop(self.args.img_size, scale=(0.2, 1.0)),
transforms.RandomHorizontalFlip(),
transforms.RandomApply([transforms.ColorJitter(0.4, 0.4, 0.4, 0.1)], p=0.8),
transforms.RandomGrayscale(0.2),
transforms.GaussianBlur(kernel_size=int(self.args.img_size * 0.1), sigma=(0.1, 2.0)),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
def _make_default_aug(image_size):
# default SimCLR augmentation
return torch.nn.Sequential(
RandomApply(T.ColorJitter(0.8, 0.8, 0.8, 0.2), p=0.3),
T.RandomGrayscale(p=0.2),
T.RandomHorizontalFlip(),
RandomApply(T.GaussianBlur((3, 3), (1.0, 2.0)), p=0.2),
T.RandomResizedCrop((image_size, image_size)),
T.Normalize(mean=torch.tensor([0.485, 0.456, 0.406]),
std=torch.tensor([0.229, 0.224, 0.225])),
)
def train(img_size=224, target_frame=56):
return transforms.Compose([
ReduceFrame(target_frame=target_frame),
transforms.RandomChoice([
transforms.GaussianBlur(kernel_size=3),
transforms.ColorJitter(brightness=0.5, contrast=0.5, saturation=0.5, hue=0.5)
]),
transforms.RandomHorizontalFlip(p=0.5),
transforms.RandomResizedCrop(img_size, scale=(0.66, 1.0), ratio=(3. / 4., 4. / 3.)),
transforms.Normalize(mean=[0.471, 0.448, 0.408], std=[0.234, 0.239, 0.242]),
Expandframe(target_frame=target_frame),
])
def __init__(self, config: dict, from_timm=True, from_th_vision=False):
super(Model, self).__init__()
try:
self.save_hyperparameters(config)
except:
pass
self.train_transforms = nn.Sequential(
transforms.Resize(size=(Config.resize, Config.resize)),
transforms.RandomHorizontalFlip(p=.7),
transforms.RandomVerticalFlip(p=.3),
transforms.RandomRotation(degrees=25),
transforms.CenterCrop(size=(Config.img_h, Config.img_w)),
transforms.ColorJitter(brightness=(0.4, 1),
contrast=.2,
saturation=0,
hue=0),
transforms.GaussianBlur(kernel_size=3))
self.validation_transforms = nn.Sequential(
transforms.Resize(size=(Config.resize, Config.resize)),
transforms.RandomRotation(degrees=25),
transforms.CenterCrop(size=(Config.img_h, Config.img_w)),
transforms.ColorJitter(brightness=(0.45, 1),
contrast=.1,
saturation=.1,
hue=0.1),
transforms.GaussianBlur(kernel_size=3))
# get backbone
if from_timm:
self.encoder = create_model(model_name=self.hparams.base_model,
pretrained=True)
else:
self.encoder = getattr(torchvision.models,
self.hparams.base_model)(pretrained=True)
# create classification layer
self.classifier = th.nn.Linear(in_features=1000, out_features=5)
self.dropout = th.nn.Dropout(p=.35)
def __init__(self, n_detectors, embed_size, pretrained_alex, net, div_transform):
super(CNN_layers, self).__init__()
self.n_detectors = n_detectors
# conv nets
self.conv = nn.ModuleList(retrieve_convnets(self.n_detectors, embed_size, pretrained_alex, net=net))
self.transforms = [to_hsv(),
kornia.filters.Sobel(),
remove_channel(0),
transforms.Grayscale(num_output_channels=3),
transforms.GaussianBlur(5, sigma=(0.1, 2.0)),
transforms.ColorJitter(brightness=0.25, contrast=0.25, saturation=0.25, hue=0.4),
identity()]
self.div_transform = div_transform