Learning through Creation: A Hash-Free Framework for On-the-Fly Category Discovery 🔥

March 13, 2026 · View on GitHub

Accepted to CVPR 2026 Findings Track

Overview

This repository contains the implementation accompanying Learning through Creation (LTC) for on-the-fly category discovery built on CLIP ViT-B/16 visual features. The method combines:

A prototype-based classifier head (cosine similarity)
Contrastive learning with two augmented views
A max-margin objective with an adaptive threshold (initialized by tau_train)
MKEE: entropy–kernel guided pseudo-unknown generation for stronger open-set training

Environment Setup

Install dependencies

Recommended: Python 3.10 (Linux/macOS). Windows is not officially tested.

Install all pinned dependencies (reference environment):

pip install -r requirements.txt

PyTorch installation notes (CUDA)

requirements.txt pins torch/torchvision versions for reproducibility, but GPU builds depend on your CUDA toolchain. If you need CUDA, install PyTorch following the official selector, and then install the rest:

pip install -U pip
# (1) install torch/torchvision with your CUDA build (see official selector)
# (2) install the remaining packages
pip install -r requirements.txt --no-deps
pip install numpy pandas scipy tqdm pillow timm ftfy regex clip-by-openai

Path Configuration

All dataset root paths and the (optional) pretrain checkpoint path are defined in [config.py]

cub_root
car_root (Stanford Cars)
food_101_root
oxford_pet_root
inaturalist_root
cifar_root
imagenet100m_root (ImageNet-100 merged, ImageFolder format)
pretrain_path (currently not required by the CLIP-based default pipeline, but kept for compatibility)

Example

Edit config.py to your local filesystem paths, e.g.:

cub_root = "/data/datasets"
car_root = "/data/datasets/stanford_cars"
food_101_root = "/data/datasets"
oxford_pet_root = "/data/datasets"
inaturalist_root = "/data/datasets"
cifar_root = "/data/datasets"
imagenet100m_root = "/data/datasets/ImageNet-100-merged"
pretrain_path = "/data/pretrain/dino_vitbase16_pretrain.pth"

Reproduction workflow (recommended)

tau_train is the initial value of the adaptive threshold τ used by the max-margin objective and later refined online via a quantile target + EMA. To reproduce, please use the setting in our CVPR paper, then:

Configure dataset roots in config.py
Install the dataset,then run each dataset at a time, e.g. CIFAR-10:

bash scripts/train_ci10.sh

For a full reproduction sweep, run:

bash scripts/train_ci10.sh
bash scripts/train_ci100.sh
bash scripts/train_pets.sh
bash scripts/train_cub.sh
bash scripts/train_food.sh
bash scripts/train_scars.sh
bash scripts/train_im.sh

Expected outputs

Each run creates the following under --output_dir:

train-logs/: training logs (file + console)
checkpoints/last.pth: the latest checkpoint (overwritten each epoch)
tf-logs/: reserved directory (not required for core execution)

-- [arXiv link] --