README.md

Datasets

This file provides explanations for understanding the structure of the datasets directory.

Adding Datasets

Adding a new dataset requires three (optionally four) parts.

Dataset Definition

First, please define your new dataset following the same pattern that exists in each currently existing dataset file. This basically entails writing two functions "prepare_train_loaders" and "prepare_test_loaders". Each of these functions prepares (1) a dataloader for the full dataset, and (2) optionally a number dataloaders corresponding to subdatasets of the full dataset. These correspond to dataloaders that only use a subset of the full data images and labels (e.g., any dataset that does not have a validation set by default).

Add Dataset Config

Second, you will need to add a config to describe the dataset you've created. This can be done by adding a succinct dictionary to the "configs.py" file in this directory. Please see the file for more information. Below is an example config:

stanford_cars = {
    'wrapper': Cars,
    'batch_size': 128,
    'res': 224,
    'type': 'stanford_cars',
    'num_workers': 8,
    'shuffle_train': True,
    'shuffle_test': False,
    'dir': './data/stanford_cars'
}

Here the variable name corresponds to the name of the dataset provided in the experiment config

(Optional) Extract CLIP Encodings

If your experiment involves a CLIP-style model, you need to extract the CLIP label encodings for your task. You can do this by running "parsing/generate_clip_heads.py" and following the instructions there. Depending on the dataset you wish to train, you may need to also add a set of label templates pertaining to the dataset in the "templates.py" file.

Add Dataset Parser

The last step entails adding a couple lines to parse your newly added dataset, in the utils.py file. Line 368 (prepare_data) contains the information you want. All that needs to be done is adding an elif that checks for the name of your dataset, and if it is a match, loads in the train and test loaders you've just created, and passes them through the pipeline.

That's it! You now have added a new dataset.