2020_ISBI_CNet

A Completion Network for Reconstruction from Compressed Acquisition (ISBI 2020)

We provide the code that produces the results that we report in

Nicolas Ducros, A Lorente Mur, F. Peyrin. A Completion Network for Reconstruction from Compressed Acquisition. 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI), Apr 2020, Iowa City, United States, pp.619-623, ⟨10.1109/ISBI45749.2020.9098390⟩. Download PDF.

Contact: [email protected], CREATIS Laboratory, University of Lyon, France.

Running the main notebook

1. Install SPyRiT and all dependencies. On Windows, you need first to install torch first (see the SPyRiT installation guide).

pip install -e spyrit .

pip install -r extra_requirements.txt

2. (optional) If you already have the STL-10 dataset on your computer, create a symbolic link. Otherwise the STL-10 dataset will be downloaded.

• Linux:

ln -s <stl-10 parent folder> /data/ 

• Windows Powershell:

New-Item -ItemType SymbolicLink -Name \data\ -Target <stl-10 parent folder>

3. Launch JupiterLab from the current folder

jupyter lab

and run main.ipynb.

Re-training the networks

The notebook main.ipynb downloads trained networks from this url. We also provide train.py to train the different variants of the network, from a single command line.

1. Completion network

   python train.py

2. Pseudo inverse network

   python train.py --net_arch 2

3. Free network

   python train.py --net_arch 3

Note that

• the models are saved in the default folder .\models\. To save them at another location consider

python train.py --model_root myfolder

• The defaults training parameters can be changed. For instance, run

python train.py --num_epochs 10 --batch_size 512 --lr 1e-8

to train your network for 10 epochs, with a batch size of 512, and a learning rate of 1e-8.

• you can keep Average_64x64.npy, Cov_64x64.npy and Var_64x64.npy in .\stats\, to avoid re-computing them, which can be time-consuming.