PIL library loads an Image from a path and returns the Image in PIL.Image.Image format ...
im = pil_loader(_IMAGE)
show_image(im)
assert isinstance(im, Image.Image)
Contrary to PIL, cv2 loads and returns the Image in np.ndarray ...
im = cv2_loader(_IMAGE)
show_image(im)
assert isinstance(im, np.ndarray)
augs = [T.RandomRotation(90), T.RandomHorizontalFlip(1.0), T.ColorJitter(0.4, 0.4, 0.4)]
augs = T.Compose(augs)
mapper = ClassificationMapper(augmentations=augs)
mapper
For training datasets ClassificationMapper takes in a DatasetDict which contains the keys file_name, targets; these corresponds to the file path of the Image and the Interger target label for the Image repectively ...
datas = DatasetDict(file_name=_IMAGE, target=0)
im, targ = mapper.encodes(datas)
# de-normalize image
im = im.permute(1, 2, 0) * torch.tensor(mapper.std) + torch.tensor(mapper.mean)
show_images([im], titles=[targ], imsize=5)
ClassificationMapper with torchvision default datasets ...
import torchvision
from gale.collections.download import download_and_extract_archive
from nbdev.export import Config
from fastcore.all import Path
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)
# create torchvision dataset instance
path = data_path / "hymenoptera_data"
ds = torchvision.datasets.ImageFolder(root=path / "train")
im, targ = mapper.encodes(ds[0])
# de-normalize image
im = im.permute(1, 2, 0) * torch.tensor(mapper.std) + torch.tensor(mapper.mean)
show_images([im], titles=[targ], imsize=5)
uint8 or PIL images. Normalization and conversion to tensors are handled independently by the library. ClassificationMapper is compatible both with albumentation augmentations and torchvision augmentations.Arguments to FolderParser:
root: Root directory path.class_map: Path to a.txtfile which contains the class mapping
parser = FolderParser(root=path / "train")
img, targ = mapper.encodes(parser[0])
img = img.permute(1, 2, 0) * torch.tensor(mapper.std) + torch.tensor(mapper.mean)
show_images([img], titles=[targ], imsize=5)
from gale.collections.pandas import folder2df, dataframe_labels_2_int
df = folder2df(path / "train")
df, class_map = dataframe_labels_2_int(df, label_column="target", return_labelling=True)
df.head()
parser = PandasParser(df, path_column="image_id", label_column="target")
img, targ = mapper.encodes(parser[0])
img = img.permute(1, 2, 0) * torch.tensor(mapper.std) + torch.tensor(mapper.mean)
show_images([img], titles=[targ], imsize=5)
To create a Image classification dataset for Gale we need to do the following -
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"
path.ls(), (path / "train").ls()
Data is stored in folders named after the classes. Also the data is divided into 2 subsets, we will only work with the train subset of the current dataset. We first initialize our parser :
parser = FolderParser(root=path / "train")
ims = []
targs = []
# Let's check some samples from our parser
for i in range(5):
img = parser[i].file_name
targ = parser[i].target
ims.append(Image.open(img))
targs.append(targ)
show_images(ims)
Now that the Parser is corectly loading the Images, to get our data in Dataset's and DataLoader's we need to first create a maper for ClassificationDataset.
The mapper will contain all the image augmentations and preprocessing necessary to prepare our data. While applying augmentations we do not need to normalize or convert our data to Tensors, this is automatically done by the Mapper :
# you can add even more facncy augmentations here
augs = A.Compose([A.Resize(128,128,p=1.0),A.HueSaturationValue(p=1.0)])
# also supports torchvision augmentations
# augs = T.Compose([T.Resize((128,128)), T.ColorJitter(0.3, 0.3, 0.3)])
mapper = ClassificationMapper(augmentations=augs)
dset = ClassificationDataset(mapper=mapper, parser=parser)
# Let's creat the dataloader
loader = torch.utils.data.DataLoader(dset, batch_size=8, shuffle=True)
View images from the DataLoader for sanity check -
samples = next(iter(loader))
show_image_batch(samples)
