Code (pytorch) for 'G2KD: Gradual Geometry-guided Knowledge Distillation for Source Data-free Domain Adaptation' on Digit, Office-31, Office-Home, VisDA-C.
You need to download the Office-31, Office-Home, VisDA-C dataset, modify the path of images in each '.txt' under the folder './data/'.
Concerning the Digits dsatasets, the code will automatically download three digit datasets (i.e., MNIST, USPS, and SVHN) in './digit/data/'.
The experiments are conducted on one GPU (NVIDIA RTX TITAN).
- python == 3.7.3
- pytorch ==1.6.0
- torchvision == 0.7.0
We choose R50-ViT-B_16 as our encoder.
wget https://storage.googleapis.com/vit_models/imagenet21k/R50+ViT-B_16.npz
mkdir ./model/vit_checkpoint/imagenet21k
mv R50+ViT-B_16.npz ./model/vit_checkpoint/imagenet21k/R50+ViT-B_16.npz- First training model on the source data, Office-31 dataset on Resnet is shown here.
~/anaconda3/bin/python g2kd_source.py --trte val --output g2kd2020r0/source/ --da uda --gpu_id 0 --dset office --max_epoch 100 --s 0 --seed 2020
- Then adapting source model to target domain, with only the unlabeled target data.
~/anaconda3/bin/python g2kd_target.py --cls_par 0.05 --da uda --dset office --gpu_id 0 --s 0 --t 1 --output_src g2kd2020r0/source/ --output g2kd2020r0/target_mix/ --seed 2020
The results of g2kd is display under the folder './G2KD_Res/object/results/accuracy/'.
The codes are based on SHOT (ICML 2020, also source-free).