Awesome
Memory-Efficient Network for Large-scale Video Compressive Sensing
This repository contains the code for the paper Memory-Efficient Network for Large-scale Video Compressive Sensing (CVPR 2021) by Ziheng Cheng, Bo Chen, Guanliang Liu, Hao Zhang, Ruiying Lu, Zhengjue Wang and Xin Yuan.
The large-scale color video reconstruction version are added in the "color model" floder.
Requirements
PyTorch > 1.3.0
numpy
scipy
skimage
Data
The training data for RevSCI-net is the same as the previous work BIRNAT. Please see the above link to generate the training set. To train the RevSCI-net, should generate the data in train/
.
Train
Reversible training:
python train.py --mode reverse --num_block 18 --num_group 2
Normal training (automatic differentiation routine):
python train.py --mode normal --num_block 18 --num_group 2
If the GPU memory is enough, recommend using normal training (about 1.5x faster than reversible training). Change the number of blocks and groups to different numbers to support different models.
Test
Run
python test.py
where will evaluate the preformance on simulation data using the pre-trained model in model/
.