Mark Nishimura
EE 367 Final Project, Winter 2018
PyTorch implementation of Choi et. al. 2017 "High-Quality Hyperspectral Reconstruction using a Spectral Prior" at https://github.com/KAIST-VCLAB/deepcassi
All analysis was conducted in the notebooks.
prepare_data.ipynb- Constructs .csv files that allow PyTorch data manipulation objects to access patches from the data in an efficient manner.autoencoder.ipynb- Constructs and trains the autoencoder. Saves checkpoints so the model weights can be loaded later.ADMM.ipynb- The ADMM reconstruction. Loads the validation set and provides several functions for displaying the data.naive_cs.ipynb- An implementation of our naive DCT-based smooth spectum ADMM algorithm.
To train the model, it is recommended that you have a GPU with CUDA and CUDNN installed. For help setting that up, see http://docs.nvidia.com/cuda/index.html and https://developer.nvidia.com/cudnn
Most of the code is available in the jupyter notebooks. To make sure they run correctly, I recommend setting up the conda environment using:
$ conda env create -f environment.yml
This command creates an environment called pytorch91. Once activated, all the relevant packages should be available.
The hyperspectral-cs/data folder contains a script for downloading the data for the CAVE and KAIST datasets. To use it, run
$ source download_data.sh
from the data folder.