yuzhi535 / ITKD

code for the paper "instance temporary knowledge distillation"

ITKD

code for the paper "instance temporary knowledge distillation"

Project page | arXiv

installation

# create env
conda create -n ITKD python=3.9
conda activate ITKD
# pytorch version
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
# other dependencies
pip install -r requirements.txt

Weights

weights are putted into pretrained folder.

Training on cifar100

All training scripts reside within the file train.sh. Adapt this file to align with your specific environment, then proceed to select a configuration from the available options. Note that the dataset will be automatically downloaded during the execution phase.

# modify the content inside and just run it
bash train.sh

Inference on cifar100

The testing scripts mirror the training scripts in their structure and usage. Simply modify the content within the scripts to harmonize with your environment.

# modify the content inside and just run it
bash test.sh