Directional Label Diffusion Model for Learning from Noisy Labels

1. Preparing python environment

Install requirements.

conda env create -f environment.yml
conda activate DLD-env
conda env update --file environment.yml --prune

The name of the environment is set to DLD-env by default. You can modify the first line of the environment.yml file to set the new environment's name.

2. Pre-trained model & Checkpoints

• ViT pre-trained models are available in the python package at here. Install without dependency:
  

pip install ftfy regex tqdm
pip install git+https://github.com/openai/CLIP.git  --no-dependencies

Trained checkpoints for the directional diffusion models will be available soon.

3. Generate the Instance-Dependent Noisy (IDN) Labels

The IDN used in our experiments are provided in folder noise_label_IDN. The noisy labels are generated following the original paper.

4. Run demo scripts to train the DLD models

4.1 CIFAR-10 and CIFAR-100

Default values for input arguments are given in the code. An example command is given:

python train_on_CIFAR.py --device cuda:0 --noise_type cifar10-idn-0.1 --nepoch 200 --warmup_epochs 5 --log_name cifar10-idn-0.1.log

4.2 Animal10N

The dataset should be downloaded according to the instruction here: Aniaml10N
Default values for input arguments are given in the code. An example command is given:

python train_on_Animal10N.py --device cuda:0 --nepoch 200 --warmup_epochs 5 --log_name Animal10N.log

4.3 WebVision and ILSVRC2012

Download WebVision 1.0 and the validation set of ILSVRC2012 datasets.

python train_on_WebVision.py --gpu_devices 0 1 2 3 4 5 6 7 --nepoch 200 --warmup_epochs 5  --log_name Webvision.log

python test_on_ILSVRC2012.py --gpu_devices 0 1 2 3 4 5 6 7 --log_name ILSVRC2012.log

4.4 Clothing1M

The dataset should be downloaded according to the instruction here: Clothing1M. Default values for input arguments are given in the code.

python train_on_Clothing1M.py --gpu_devices 0 1 2 3 4 5 6 7 --nepoch 200 --warmup_epochs 5  --log_name Clothing1M.log