tf.keras.utils.image_dataset_from_directory

Generates a tf.data.Dataset from image files in a directory.

If your directory structure is:

main_directory/
...class_a/
......a_image_1.jpg
......a_image_2.jpg
...class_b/
......b_image_1.jpg
......b_image_2.jpg

Then calling image_dataset_from_directory(main_directory, labels='inferred') will return a tf.data.Dataset that yields batches of images from the subdirectories class_a and class_b, together with labels 0 and 1 (0 corresponding to class_a and 1 corresponding to class_b).

Supported image formats: jpeg, png, bmp, gif. Animated gifs are truncated to the first frame.

directory Directory where the data is located. If labels is "inferred", it should contain subdirectories, each containing images for a class. Otherwise, the directory structure is ignored.
labels Either "inferred" (labels are generated from the directory structure), None (no labels), or a list/tuple of integer labels of the same size as the number of image files found in the directory. Labels should be sorted according to the alphanumeric order of the image file paths (obtained via os.walk(directory) in Python).
label_mode String describing the encoding of labels. Options are:

  • 'int': means that the labels are encoded as integers (e.g. for sparse_categorical_crossentropy loss).
  • 'categorical' means that the labels are encoded as a categorical vector (e.g. for categorical_crossentropy loss).
  • 'binary' means that the labels (there can be only 2) are encoded as float32 scalars with values 0 or 1 (e.g. for binary_crossentropy).
  • None (no labels).
class_names Only valid if "labels" is "inferred". This is the explicit list of class names (must match names of subdirectories). Used to control the order of the classes (otherwise alphanumerical order is used).
color_mode One of "grayscale", "rgb", "rgba". Default: "rgb". Whether the images will be converted to have 1, 3, or 4 channels.
batch_size Size of the batches of data. Default: 32. If None, the data will not be batched (the dataset will yield individual samples).
image_size Size to resize images to after they are read from disk, specified as (height, width). Defaults to (256, 256). Since the pipeline processes batches of images that must all have the same size, this must be provided.
shuffle Whether to shuffle the data. Default: True. If set to False, sorts the data in alphanumeric order.
seed Optional random seed for shuffling and transformations.
validation_split Optional float between 0 and 1, fraction of data to reserve for validation.
subset Subset of the data to return. One of "training", "validation" or "both". Only used if validation_split is set. When subset="both", the utility returns a tuple of two datasets (the training and validation datasets respectively).
interpolation String, the interpolation method used when resizing images. Defaults to bilinear. Supports bilinear, nearest, bicubic, area, lanczos3, lanczos5, gaussian, mitchellcubic.
follow_links Whether to visit subdirectories pointed to by symlinks. Defaults to False.
crop_to_aspect_ratio If True, resize the images without aspect ratio distortion. When the original aspect ratio differs from the target aspect ratio, the output image will be cropped so as to return the largest possible window in the image (of size image_size) that matches the target aspect ratio. By default (crop_to_aspect_ratio=False), aspect ratio may not be preserved.
**kwargs Legacy keyword arguments.

A tf.data.Dataset object.

  • If label_mode is None, it yields float32 tensors of shape (batch_size, image_size[0], image_size[1], num_channels), encoding images (see below for rules regarding num_channels).
  • Otherwise, it yields a tuple (images, labels), where images has shape (batch_size, image_size[0], image_size[1], num_channels), and labels follows the format described below.

Rules regarding labels format:

  • if label_mode is int, the labels are an int32 tensor of shape (batch_size,).
  • if label_mode is binary, the labels are a float32 tensor of 1s and 0s of shape (batch_size, 1).
  • if label_mode is categorical, the labels are a float32 tensor of shape (batch_size, num_classes), representing a one-hot encoding of the class index.

Rules regarding number of channels in the yielded images:

  • if color_mode is grayscale, there's 1 channel in the image tensors.
  • if color_mode is rgb, there are 3 channels in the image tensors.
  • if color_mode is rgba, there are 4 channels in the image tensors.