Skip to content

Latest commit

 

History

History
37 lines (29 loc) · 1.97 KB

README.md

File metadata and controls

37 lines (29 loc) · 1.97 KB

About

This repository contains the codes and experiments for the paper, Solving Inverse Problems in Snapshot Compressive Imaging with Score Generative Models. Implementation of the score model was forked from Yang Song's repository and modified accordingly for our experiments, which was carried out in the environment defined in score_sci/environment.yml.

Note that, you may encounter troubles setting up the exact environment and dependencies specified in the configuration file depending on your GPU model and CUDA version. If so, please refer to the original installation guides for relevant packages.

If you have any doubts or queries, please feel free to reach out at my email. Thank you :)

Getting started

Downloading model checkpoints

  1. Download model checkpoints provided by Yang Song et al.
  2. Store checkpoints under score_sci/checkpoints/. E.g., 
    | score_sci/checkpoints/
|---- ve/
|-------- cifar10_ncsnpp_continuous/
|------------ checkpoint_24.pth
|-------- ffhq_256_ncsnpp_continuous/
|------------ checkpoint_48.pth
  3. Setup Conda environment. 
    conda env create --name envname --file=score_sci/environment.yml

Viewing the SCI dataset

  1. From the dataset is available at Google Drive, download the matlab.zip and test_gray.zip under the SCI folder.
  2. Load the dataset by running the load_dataset_mat_example.py.

Main experiments

  1. Navigate to the main folder, cd ./score_sci
  2. Run demo either through main.py or main.ipynb
    • Note that, if main.ipynb is used, you may have to restart your kernel to clear the GPU memory if CUDA memory limit errors are encountered.
  3. Generated samples will be saved under assets/{scene}_{sampler} by default