Learning_to_diversify

This is the official code repository for ICCV2021 'Learning to Diversify for Single Domain Generalization'.

Paper Link: http://arxiv.org/abs/2108.11726

Update: Single DG with Resnet-18

Recently, we receive increasing enquiry about single DG on PACS with Resnet-18 Backbone. (In the paper, we reported Alexnet result) Please try hyperparameters lr=0.002 and e=50, to start your experiment.

We report the following single DG result on PACS, with resnet-18 as the backbone network:

Src. domain	P	A	C	S	avg.
Avg. Tar. Acc.	52.29	76.91	77.88	53.66	65.18

Quick start: (Generalizing from art, cartoon, sketch to photo domain with ResNet-18)

1. Install the required packages.
2. Download PACS dataset.
3. Execute the following code.

bash run_main_PACS.sh

Change dataset

In line 266-300 of train.py, we provide 3 different datasets settings (PACS, VLCS, OFFICE-HOME). You can simply uncomment it to start your own experiment. It may require hyper-parameter fine tuning for some of the tasks.