image_datasets

The Python library provides easy access to images in a dataset. It is an alternative when you don't work with PyTorch or Keras (or whatever) but you still want to load a dataset as it provides basic functionnalities with minimal requirements and dependencies.

For now support classification and segmentation datasets. The lib assumes a certain structure for the dataset files (explained below).

Directory structures

• For segmentation:
• IMAGE_FOLDER/[SET_FOLDER/]/INPUT_DIR/FILENAME

Classification datasets

Files should be organized as follows: IMAGE_FOLDER/[SET_FOLDER/]CLASS_FOLDER/IMAGE_FILENAME

Constructor

Parameters:

• image_folder: path to the dataset root folder
• dirs: list of of fold directories (None if the dataset is not splitted)
• classes: list of classes (if None, auto-detected from the folder structure)
• encode_classes: True for encoding classes as numbers in {0, ..., n_class}, False to keep folder names

dataset = ImageClassificationDataset("~/my_dataset", dirs=["train", "test"])
dataset = ImageClassificationDataset("~/my_dataset", dirs=["train", "test"], classes=[5562, 23178], encode_classes=True)

Method all

Returns the list of image filenames and classes (optionally for a given fold).

Parameters:

• dirs: names of the folds of which the images information must be extracted (None, for taking images from all folds)
• classes: a subset of classes to select (None for all classes)
• positive: a subset of classes to consider as positive (None for keeping it multiclass)
• negative: a subset of classes to consider as negative (None for keeping it multiclass)
• one_hot: true for a one-hot encoding of the classes

dataset = ImageClassificationDataset("~/my_dataset", dirs=["train", "test"])
x_train, y_train = dataset.all(dirs=["train"], one_hot=False)

# x_train is a list of strings containing the filenames
# y_train is the list of classes as integers 

Method batch

Return a random batch of images and their classes.

Parameters:

• size: size of the batch
• dirs: see method all
• replace: True for sampling with replacement
• random_state: a random state
• classes: see method all
• positive: see method all
• negative: see method all
• one_hot: see method all

dataset = ImageClassificationDataset("~/my_dataset", dirs=["train", "test"])
x_batch, y_batch = dataset.batch(12, dirs=["train"], one_hot=False)

# x_batch is a list of 12 strings containing the filenames of the batch
# y_batch is the list of classes as integers 

Segmentation dataset

TODO

Other

Other utilities functions provided by the library.

Function ImageDataset.load_images

Load all images from a list of filepaths into memory. Return them as a list.

Parameters

• files: list of filepaths of the file to open
• open_fn: the function to use to open the images. By default, try to import OpenCV's imread.

dataset = ImageClassificationDataset("~/my_dataset", dirs=["train", "test"])
x_train, y_train = dataset.all(dirs=["train"], one_hot=False)
images_train = ImageDataset.load_images(x_train)

Function ImageClassificationDataset.load_crops

Load one or several crops from all given images.

• files: The path of the files
• crop_size: The size of the square crop
• y: Output vector, for it to be updated if several are taken per image
• load_min_size: Minimum loading size for smallest side of the image (>= crop_size)
• load_max_size: Maximum loading size for smallest side of the image (>= load_min_size)
• grayscale: True for loading as a grayscale
• crops_per_image: Number of crops to extract per image
• random: True for extracting crop at a random location in the image
• random_state: Random state

dataset = ImageClassificationDataset("~/my_dataset", dirs=["train", "test"])
x_train, y_train = dataset.all(dirs=["train"], one_hot=False)

# for taking 10 random crops per image, each crops taken as between 50% and 100% of the image
crops_train, crops_y = ImageClassificationDataset.load_images(
    x_train, crop_size=224,
    load_size_min=224,
    load_size_max=448
    y=y_train,
    crops_per_image=10,
    random=True,
    random_state=42
)

# for taking one maximum centered random crop per image
crops_train_2, crops_y_2 = ImageClassificationDataset.load_images(
    x_train, crop_size=224,
    y=y_train, random=False
)