Few-shot Fine-tuning is All You Need for Source-free Domain Adaptation

Requirments

python == 3.9.12
torch == 1.12.1
torchvision == 0.13.1
numpy == 1.23.5
scikit-learn == 1.2.0

Prepare Datasets

For vanilla SFDA setting, please prepare datasets in <data_dir> as follows :

|---- data_dir/
|      |---- office_home/
|            |---- Art
|                 |---- class 0 / *.png
|                 |---- ...
|                 |---- class 64 / *.png
|            |---- Clipart
|                 |---- class 0 / *.png
|                 |---- ...
|                 |---- class 64 / *.png

Available datasets : office31, office_home,VISDA-C,VLCS,terra_incognita

Prepare Source Model

We follow the same training procedure from SHOT to generate source pretrained model. Please refer to the official repository of SHOT for source pretraining. All source models we used in the experiments will be released in future.

Few-shot Target Adaptation under Vanilla SFDA setting

3-shot fine-tuning(FT) on Office31 dataset

# Adapt source domain 0 (Amazon) to all possible target domains (DSLR, Webcam)

CUDA_VISIBLE_DEVICES=0 python target_finetune.py --dataset office31 \
--pretrain SHOT --adapt all --few_shot 3 --work_dir <work_dir> --val_frep 10 \
--lr 1e-05 --seed 0 --source 0 --ckpt_dir <Path to source model directory>

3-shot LP-FT on Office31 dataset

First, train linear classifier only(LP) by changing --adapt all to --adapt cls and adding --adapt_step 400.

# Adapt source domain 0 (Amazon) to all possible target domains (DSLR, Webcam)

CUDA_VISIBLE_DEVICES=0 python target_finetune.py --dataset office31 --adapt_step 400 \
--pretrain SHOT --adapt cls --few_shot 3 --work_dir <work_dir> \
--lr 1e-04 --seed 0 --source 0 --ckpt_dir <Path to source model directory>

Next, train whole network(FT) starting from LP checkpoint by changing --pretrain SHOT to --pretrain SHOT_LP and passing the path to LP checkpoint(work_dir in the above command) to ckpt_dir

# Adapt source domain 0 (Amazon) to all possible target domains (DSLR, Webcam)

CUDA_VISIBLE_DEVICES=0 python target_finetune.py --dataset office31 \
--pretrain SHOT_LP --adapt all --few_shot 3 --work_dir <work_dir> --val_frep 10 \
--lr 1e-05 --seed 0 --source 0 --ckpt_dir <Path to LP trained directory>

Results

Log files and checkpoints will be saved as follows :

|---- logs/
|     |---- office31/
|           |---- work_dir/source0
|                |---- target 1
|                     |---- train_log.json
|                     |---- last_ckpt.pth # Last model
|                     |---- test_best_ckpt.pth # Model that gives the highest test accuracy
|                     |---- val_best_ckpt.pth # Model that gives the lowest validation loss
|                |---- target 2
|                |---- config.json
|                |---- best_target.json # Aggregated performance of test best model
|                |---- last_target.json # Aggregated performance of the last model
|                |---- selected_target.json # Aggregated performance of validation best model

Note that all the reported values are from validation best model.

Learning rates used in FT and LP-FT

For both FT and the second stage of LP-FT, we used 1e-05 for all benchmark datasets. For the first stage of LP-FT, we used 1e-04 for all benchmark datsets. All the learning rates are selected through 1-shot validation.

daintlab/fewshot-SFDA