from fastcore.all import Path
from nbdev.export import Config
from torch.utils.data import DataLoader
from gale.classification.core import *
from gale.collections.download import download_and_extract_archive
URL = "https://download.pytorch.org/tutorial/hymenoptera_data.zip"
data_path = Path(Config().path("nbs_path")) / "data"
# download a toy dataset
download_and_extract_archive(url=URL, download_root=data_path, extract_root=data_path)
# take a peek at the structure of the dataset
path = data_path / "hymenoptera_data"
parser = FolderParser(path / "train")
Args:
size: Desired output size. If size is a sequence like (h, w), output size will be matched to this. If size is an int, smaller edge of the image will be matched to this number. i.e, if height > width, then image will be rescaled to (size * height / width, size).interpolation: Desired interpolation
transforms = imagenet_no_augment_transform(size=(224, 224), interpolation="bicubic")
mapper = ClassificationMapper(transforms)
dset = ClassificationDataset(mapper, parser)
dls = DataLoader(dset, batch_size=8, shuffle=False)
samples = next(iter(dls))
show_image_batch(samples)
Args:
size: Desired output size. If size is a sequence like (h, w), output size will be matched to this. If size is an int, smaller edge of the image will be matched to this number. i.e, if height > width, then image will be rescaled to (size * height / width, size).scale: Range of size of the origin size cropped.ratio: Range of aspect ratio of the origin aspect ratio cropped.interpolation: Desired interpolation.hflip: Probability for random horizontal flip. Active ifnot False and > 0.vflip: Probability for random vertical flip. Active ifnot False and > 0.color_jitter: Values of brightness, contrast, saturation and hue fortorchvision.transforms.ColorJitter.auto_augment: String defining configuration of auto augmentation policies. Currently supports only AutoAugment and RandAugment.mean: Image mean, required only ifauto_augmentis not None.
transforms = imagenet_augment_transform(
size=(224, 224),
hflip=0.5,
vflip=0.5,
color_jitter=0.4,
)
mapper = ClassificationMapper(transforms)
dset = ClassificationDataset(mapper, parser)
dls = DataLoader(dset, batch_size=8, shuffle=False)
samples = next(iter(dls))
show_image_batch(samples)
Applying AutoAugment:
To use AutoAugment we need to create a config str and specify the parameters. Under the hood the above functions uses auto_augment_transform from timm. Config str must be like should consists of multiple sections separated by dashes ('-'). The first section defines the AutoAugment policy (one of 'v0', 'v0r', 'original', 'originalr'). The remaining sections, not order sepecific determine
mstd(float std deviation of magnitude noise applied)
Ex - original-mstd0.5 results in AutoAugment with original policy, magnitude_std 0.5
transforms = imagenet_augment_transform(
size=(224, 224), hflip=0.5, vflip=0.5, auto_augment="original-mstd0.5"
)
mapper = ClassificationMapper(transforms)
dset = ClassificationDataset(mapper, parser)
dls = DataLoader(dset, batch_size=8, shuffle=False)
samples = next(iter(dls))
show_image_batch(samples)
Applying RandAugment:
To use RandAugment we need to create a config str and specify the parameters. Under the hood the above functions uses rand_augment_transform from timm. Config str must be like should consists of multiple sections separated by dashes ('-'). The first section defines the specific variant of rand augment (currently only 'rand').
The remaining sections, not order sepecific determine:
m- integer magnitude of rand augmentn- integer num layers (number of transform ops selected per image)w- integer probabiliy weight index (index of a set of weights to influence choice of op)mstd- float std deviation of magnitude noise appliedinc- integer (bool), use augmentations that increase in severity with magnitude (default: 0)
Ex - rand-m9-n3-mstd0.5 results in RandAugment with magnitude 9, num_layers 3, magnitude_std 0.5 & rand-mstd1-w0 results in magnitude_std 1.0, weights 0, default magnitude of 10 and num_layers 2
transforms = imagenet_augment_transform(
size=(224, 224), hflip=0.5, vflip=0.5, auto_augment="rand-m9-n3-mstd0.5"
)
mapper = ClassificationMapper(transforms)
dset = ClassificationDataset(mapper, parser)
dls = DataLoader(dset, batch_size=8, shuffle=False)
samples = next(iter(dls))
show_image_batch(samples)
This function utilizes transformations from albumentations library. First the Image is resized to presize , transformations are applied after which we RandomCrop to size.
HorizontalFlip (or VerticalFlip if vflip=True) with p=hlip (or vflip) is added when hflip=True. With p_rotate we apply a albumentations.Rotate of max_rotate degrees and if max_warp a albumentations.IAAAffine with scale=max_warp and model=pad_mode. With p_lighting we apply a change in brightness and contrast of max_lighting. Custonm xtra_tfms can be added, these must be a List contraining transformations from albumentations. max_rotate,max_lighting,max_warp are multiplied by mult so you can more easily increase or decrease augmentation with a single parameter.
transforms = aug_transforms(260, 224, mult=1.0)
mapper = ClassificationMapper(transforms)
dset = ClassificationDataset(mapper, parser)
dls = DataLoader(dset, batch_size=8, shuffle=False)
samples = next(iter(dls))
show_image_batch(samples)
RandomErasing with these transformations. But random erasing must be applied after normalization of the batch so to apply random erasing we need to pass this transform directly into our mapper.from timm.data.random_erasing import RandomErasing
re = RandomErasing(1.0, mode="pixel", device="cpu")
# # or alternatively:
# from torchvision.transforms import RandomErasing
# re = RandomErasing(1.0)
transforms = aug_transforms(260, 224, mult=2.0)
mapper = ClassificationMapper(transforms, xtras=re)
dset = ClassificationDataset(mapper, parser)
dls = DataLoader(dset, batch_size=8, shuffle=False)
samples = next(iter(dls))
show_image_batch(samples)
