The repo for ZIN: When and How to Learn Invariance Without Environment Partition?

This project is implemented based on IRM, eiil and In-N-Out.

There are two methods to get the environment. You only need to try ONE method that you prefer.

Our code works with the following environment:

• pytorch-transformers==1.2.0
• torch==1.3.1
• torchvision==0.4.2
• Pillow==6.1

To install the necessary packages for the project, please run:

pip install -r requirements.txt

docker pull lutan0914/zin

Run the following command to download the datasets: CelebA, house_price, landcover. They will be put in datasets directory.

bash data_downloader.sh

Dataset Discriptions:

• HousePrice.  This is implemented based on the house price dataset from Kaggle
• CelebA. The dataset is from here. Since IRM suffers from overfitting problem when applied to large models like ResNet-18, we fix the feature extraction backbone (that is, use a pre-trained ResNet-18 with fixed blocks) in this task. One may avoid this limitation by incorperate overfitting robust IRM variants like [1][2][3].
• Landcover. Our implementation on landcover is based on the source code of In-N-Out. Notably, we use exact the same random seed as In-N-Out.

[1] Yong Lin, Hanze Dong, Hao Wang, Tong Zhang, Bayesian Invariant Risk Minimization, CVPR 2022

[2] Xiao Zhou, Yong Lin, Weizhong Zhang, Tong Zhang, Sparse Invariant Risk Minimization, ICML 2022

[3] Xiao Zhou, Yong Lin, Renjie Pi, Weizhong Zhang, Renzhe Xu, Peng Cui, Tong Zhang., Model Agnostic Sample Reweighting for Out-of-Distribution Learning, ICML 2022

• l2_regularizer_weight: L2 regularization weight.
• lr: learning rate
• steps: training steps
• irm_type: which algorithm to use, infer_irmv1 and infer_irmv1_multi_class are for ZIN. If the number of inferred envs are larger than 2, you need to choose infer_irmv1_multi_class. Otherwise, they are the same.
• dataset: which dataset to use
• penalty_anneal_iters: the ERM proceducer befor imposing the IRM penalty on the model, this is also the environmental inference procedure.

The following parameters are only for the synthetic dataset:

• noise_ratio: noise ratio when generation Y from invariant features
• cons_train: the correlation between spurious features and Y in the training domains
• cons_test: the correlation between spurious features and Y in the testing domains
• dim_inv: dimension of invariant features
• dim_sp: dimension of spurious features
• data_num_train: number of training samples
• data_num_test: number of testing samples

The parameter for the Landcover dataset:

• aux_num: the dimension of auxilary information, when Z is longitude and latitude when aux_num=2.

1. To run the ZIN in the temporal dataset with setting p_s=(0.999, 0.9) and p_v=0.9.

   python main.py --l2_regularizer_weight 0.001 --lr 0.005 --noise_ratio 0.1 --cons_train 0.999_0.9 --cons_test 0.999_0.8_0.2_0.001 --penalty_weight 10000 --steps 10000 --dim_inv 5 --dim_sp 5 --data_num_train 5000 --data_num_test 5000 --n_restarts 1 --irm_type infer_irmv1 --dataset logit_z --penalty_anneal_iters 5000
   

   The expected test accuracy is about 82.96.

2. To run ZIN in CelebA dataset with 7 auxiliary information: "Young", "Blond_Hair", "Eyeglasses", "High_Cheekbones", "Big_Nose", "Bags_Under_Eyes", "Chubby"

   python main.py --l2_regularizer_weight 0.001 --lr 0.005 --noise_ratio 0.2 --cons_train 0.999_0.8 --cons_test 0.01_0.2_0.8_0.999 --penalty_weight 10000 --steps 8500 --dim_inv 5 --dim_sp 5 --data_num_train 39996 --data_num_test 20000 --n_restarts 1 --irm_type infer_irmv1 --dataset celebaz_feature --penalty_anneal_iters 8000 --seed 1
   

   The expected test accuracy is about 76.29.

3. To run ZIN in house price prediction task

   python main.py --l2_regularizer_weight 0.001 --lr 0.005 --penalty_weight 1000 --steps 5901   --n_restarts 1 --irm_type infer_irmv1_multi_class --dataset house_price --penalty_anneal_iters 5000 --seed 1 --hidden_dim_infer 64 --hidden_dim 32 
   

   The expected test accuracy is about 0.3013.

4. To run the ZIN with location(latitude and longitude) as auxiliary information in landcover dataset:

   python main.py --aux_num 2 --batch_size 1024 --seed 112 --classes_num 6 --dataset landcover --opt adam --l2_regularizer_weight 0.001 --print_every 1 --lr 0.1 --irm_type infer_irmv1_multi_class --n_restarts 1 --num_classes 6 --z_class_num 2 --penalty_anneal_iters 40 --penalty_weight 10 --steps 400 --scheduler 1
   

   The expected test accuracy is about 66.06.

1. To run MCOLOR with ERM, IRM and EIIL:

   cd eiil
   ./exps/mcolor.sh
   

   Note that the eiil subdirectory is modified from eiil, and has its own license.

2. To run CMNIST with ERM, IRM and EIIL:

   cd eiil
   ./exps/cmnist_with_specified_label_noise.sh 
   

   Note that the eiil subdirectory is modified from eiil, and has its own license.

We provider interface for you to include your own data. You need to inherit the class LYDataProviderMK, and re-implement the function fetch_train and fetch_test. The main function will call fetch_train to get training data for each step. fetch_train should return the following values:

• train_x: the feature tensor;
• train_y: the label tensor;
• train_z: the auxilary information tensor;
• train_g(optional, can set to be None): the tensor contains values indicating which environmnets the data are from;
• train_c(optional, can set to be None): the tensor contains values indicating whether the spurious features align with the labels.
• train_invnoise(optional, can set to be None): the tensor contains values indicating the noisy ratio of the label;

The structure of the return value of fetch_test are similar with fetch_train.

For help or issues using ZIN, please submit a GitHub issue.

For personal communication related to ZIN, please contact Yong Lin ([email protected]).

If you use or extend our work, please cite the following paper:

@inproceedings{lin2022zin,
  title={ZIN: When and How to Learn Invariance Without Environment Partition?},
  author={Lin, Yong and Zhu, Shengyu and Tan, Lu and Cui, Peng},
  booktitle={Advances in neural information processing systems},
  year={2022}
}