DGSML is the first method of semi-supervised meta learning that employs two novel loss functions to learn a predictive domain-invariant representation by taking both labeled and unlabeled samples from different input domains. Please read the draft here:
https://arxiv.org/pdf/2009.12658.pdf
- Python 3
- Conda
To get the source files of DGSML you need to clone into its repository:
git clone https://github.com/hosseinshn/DGSML.git
All the required packages to run DGSML are specified in environment subdirectory of dgsml-gpu and dgsml-resnet-cpu separately.
To install these packages run the following command:
conda env create -f ENVIRONMENT.yml
Where ENVIRONMENT.yml is the name of the requirement file in the environment directory.
So to create the environment required for DGSML-GPU run:
conda env create -f dgsml-gpu/environment/dgsml_gpu.yml
This will create a conda environment called dgsml_gpu.
And to create the environment required for DGSML-CPU (resnet only) run:
conda env create -f dgsml-resnet-cpu/environment/dgsml_cpu.yml
This will create a conda environment called dgsml_cpu.
After the successful installation of the packages into environmet you would be able to load them using conda activate.
We used PACS and VLCS benchmarks to evaluate the performance of DGSML and the studied baselines. PACS was obtained from here and VLCS from here. Before running DGSML, unzip the data using following commands:
cd data/DATASET_NAME/
for F in *.tar.gz; do tar -xzvf ${F}; done
Where DATASET_NAME is either PACS or VLCS.
To run a small test:
cd dgsml-gpu/scripts/
bash test_run.bs
It can also be submitted to a cluster.
This section provides instructions to reproduce the reported experimnets using AlexNet and ResNet-18. It is important to note that AlexNet experiments are designed to be performed on GPU only and ResNet-18 experiments on CPU only.
The scripts for submitting VLCS and PACS jobs using AlexNet to a cluster are provided in dgsml-gpu/scripts/ directory.
GPU resources should be added to the header section of the batch scripts based on the GPU settings on your system before submitting them to the cluster.
On a slurm job scheduling system they can be submitted to a cluster using:
cd dgsml-gpu/scripts/
mkdir -p output_vlcs
mkdir -p log_vlcs
sbatch test_VLCS.bs
mkdir -p output_pacs
mkdir -p log_pacs
sbatch test_PACS.bs
The scripts for submitting PACS jobs using ResNet-18 to a cluster are provided in dgsml-gpu/scripts/ directory.
On a slurm job scheduling system they can be submitted to a cluster using:
cd dgsml-resnet-cpu/scripts/
mkdir -p output_pacs
mkdir -p log_pacs
sbatch test_PACS.bs
The structure of an output directory (e.g. save0.95_1_photo) after a successful run will be as follows:
In this example 0.95 is the rate of unlabeled samples, 1 is the run id (a number between 1 to 5 corresponding to an independent run with a specific seed value), and photo is the target domain.
save0.95_1_photo
├── args.txt # Input arguments passed to DGSML and training losses
├── Best_CLS.pt # Best saved classifier
├── Best_FT.pt # Best saved feature extractor
├── Best_val.txt # Saved validation accuracies
└── Target.txt # Target accuracy
To obtain a table of the accuracy results you can run show_vlcs_result.sh for AlexNet and show_pacs_result.sh for both AlexNet and ResNet-18.
These scripts are provided in dgsml-gpu/scripts/ and dgsml-resnet-cpu/scripts/.
For AlexNet results run:
cd dgsml-gpu/scripts/
bash show_vlcs_result.sh
bash show_pacs_result.sh
For ResNet-18 results run:
cd dgsml-resnet-cpu/scripts/
bash show_pacs_result.sh
Note that for each rate of unlabeled samples and each target domain, 5 independent runs with different seed values are added to the job list which their accuracies appear in the tables generated by the bash scripts mentioned above.
@article{sharifi2020domain,
title={Domain Generalization via Semi-supervised Meta Learning},
author={Sharifi-Noghabi, Hossein and Asghari, Hossein and Mehrasa, Nazanin and Ester, Martin},
journal={arXiv preprint arXiv:2009.12658},
year={2020}
}
In the implementation of DGSML, we reused some codes from the following repositories.