BLASTNet 2.0 Super-Resolution Benchmark

Code for Super-Resolution (SR) Benchmark Study involving BLASTNet 2.0 Data Written by W.T Chung and B. Akoush with citations for open-source code in individual files.

Requirements

To install requirements:

pip install -r requirements.txt

This code works with Python 3.9. We used conda to manage environments in the work.

Data

Data for this benchmark can be found at Kaggle.

Summary

Brief rundown on code:

1. common contains code for math, data loading, and general utilities.
2. metadata contains csv files with file ids that can be fed into the dataloaders for 5 different splits (train/test/val + 2 OOD sets)
3. create_cubic_files.ipynb precomputes cubic interpolation on files for a baseline comparison.
4. find_addport.py finds master port for multinode training.
5. sample_lsf_{train,test,testcubic}.sh are batch submission scripts for IBM LSF (not Slurm).
6. {train,test,testcubic}.py perform multinode training, and single-node evaluations for ML and cubic interpolation, respectively.
7. requirements.txt provides recommended packages to run this code

Training

To train the models, we provide a sample_lsf_train.sh that provides multi-node training via a batch submission on IBM LSF (not SLURM).

Evaluation

To eval the models, we provide a sample_lsf_test.sh that provides single-gpu evaluation via a batch submission on IBM LSF (not SLURM). Essentially it does:

python test.py \
--data_path=../diverse_2K_with_extrap/ \
--upscale=8 --timeit \
--approx_param=0.5M --case_name=./weights/seed42/rcan_approx0.5M_8xSR.pt \
--precision=32 --num_nodes=1 --model_type=rcan

If you want to compare the ML models with cubic interpolation. You can use create_cubic_files.ipynb to obtain interpolated data. And use sample_lsf_testcubic.sh, which does the same as:

python testcubic.py \
--data_path=../diverse_2K_with_extrap/ \
--cubic_path=../cubic/ \
--upscale=8 --timeit \
--batch_size=2  \
--precision=32 --num_nodes=1 

Pre-trained Models

We also provided pre-trained weights in the same Kaggle repo. A sample is provided in the weights folder.

Inference

You can learn how to plot the features and labels, and perform inference with the pre-trained weights in this Kaggle notebook!