Load ImageNet with TensorFlow-Datasets tfds from a PyTorch Folder Structure

Dear ImageNet and TensorFlow specialists

Originally, I have downloaded the ImageNet dataset tar files ILSVRC2012_img_train.tar, ILSVRC2012_img_val.tar and used this PyTorch script to extract the files. This results in the folder structure

#  imagenet/train/
#  ├── n01440764
#  │   ├── n01440764_10026.JPEG
#  │   ├── n01440764_10027.JPEG
#  │   ├── ......
#  ├── ......
#  imagenet/val/
#  ├── n01440764
#  │   ├── ILSVRC2012_val_00000293.JPEG
#  │   ├── ILSVRC2012_val_00002138.JPEG
#  │   ├── ......
#  ├── ......

How can I load this resulting folder structure into a TensorFlow Dataset (tfds)? An example is shown here, but it assumes extraction from the tar files (which I no longer have available). Please help : )

You don't need to use TensorFlow Datasets (tfds) to load the ImageNet dataset from the existing folder structure, but you can utilize tf.data.Dataset from the TensorFlow library.

Using the Keras API:

import tensorflow as tf
import os

# Define paths to the train and validation datasets
train_dir = 'path_to/imagenet/train'
val_dir = 'path_to/imagenet/val'

# Load training dataset
train_dataset = tf.keras.preprocessing.image_dataset_from_directory(
    train_dir,
    image_size=(224, 224),  # Define the image size you want
    batch_size=32,  # Define the batch size you want
    label_mode='categorical',  # For multi-class classification problem
    shuffle=True,
    seed=42,
)

# Load validation dataset
val_dataset = tf.keras.preprocessing.image_dataset_from_directory(
    val_dir,
    image_size=(224, 224),  # Define the image size you want
    batch_size=32,  # Define the batch size you want
    label_mode='categorical',  # For multi-class classification problem
    shuffle=False,
    seed=42,
)

That will create TensorFlow datasets from your folder structure.
Do replace 'path_to/imagenet/train' and 'path_to/imagenet/val' with the actual paths to the training and validation datasets.

The function tf.keras.utils.image_dataset_from_directory automatically infers the labels from the subdirectory names in the provided directory path. The labels are one-hot encoded which suits multi-class classification tasks.
See "Building a One Hot Encoding Layer with TensorFlow" from George Novack for illustrating that concept.
The label_mode='categorical' argument in the tf.keras.preprocessing.image_dataset_from_directory function instructs TensorFlow to one-hot encode the labels. That means that the labels are converted into a binary vector where 1 stands for the correct class, and 0 stands for the rest.
That one-hot encoding format suits multi-class classification tasks because in these tasks, an instance can belong to one class out of multiple possible classes. (as opposed to "binary", when there are only two possible classes).
In the context of the ImageNet dataset, there are 1,000 possible classes, so one-hot encoding is a suitable choice for representing the class labels.

The function returns a tf.data.Dataset object that yields batches of images and labels. Images are in the format of float32 with values between 0 and 1, and the labels are also in float32 format.

You can then use these datasets to train your model.