This work was based on the following implementation: https://github.com/MarkPrecursor/SRCNN-keras.
This is also an attempt of implementation of the following work: Image Super-Resolution Using Deep Convolutional Networks, also available in Arxiv.
This work was mostly developed and executed within the following environment settings:
Python 3.6.5
pip 9.0.1 from /usr/lib/python3/dist-packages (python 3.6)
Distributor ID: Ubuntu
Description: Ubuntu 18.04 LTS
Release: 18.04
Codename: bionic
Docker version 17.12.1-ce, build 7390fc6
This repository was also tested in mybinder environment, but did not run, due to its heavy processing. You could give it a try! (Note that the environment takes very long to build, so please be patient)
The notebook within deliver/Final_Version.ipynb downloads the data in a tarball and extracts it within the working directory. If that does not work, one can try and download them manually from figshare, using one of the following sources:
-
The data.tar.gz that contains plotting info and 2 samples (basically this repository's data directory): https://ndownloader.figshare.com/files/15483431
-
If one wishes to download the h5 separately, the data are all available within this project in figshare: https://figshare.com/account/home#/projects/64340
The files named ground* must be downloaded within data/ground_data/ and the ones named sub*, within data/sub_data/
If you are using the docker environment, it contains the data.tar.gz, so there is no need to download it.
Please set your environment to run on Python3, since this project was written and tested to run under Python3.
You might need to install the following dependencies in order to run this project:
- scikit-image
- numpy
- h5py
- matplotlib
- scipy==1.1.0
- tensorflow
- keras
- opencv-python
- requests
In order to install them, you can simply run:
pip3 install -v requests numpy h5py matplotlib tensorflow keras opencv-python scikit-image scipy==1.1.0
One must start an instance of jupyter notebook on the root directory of their local copy of this repository. Then, change directory to 'deliver' and open the file Final_Version.ipynb. From then on, the notebook should run.
The notebook named deliver/Final_Version.ipynb runs the report of the experiments on the data and displays some of the predictions, as well as the ground truth of the test data.
The slice from where the prediction will run can be chosen. As default, an additional of 20 slices will then be also included in the prediction and visualization of the result. This can also be changed, but be aware that it might be demanding on your hardware.
For the following configuration:
Intel(R) Core(TM) i7-3630QM CPU @ 2.40GHz
total used free shared buff/cache available
Mem: 7,7G 2,4G 3,4G 744M 1,9G 4,3G
Swap: 14G 2,0G 12G
The program was not able to run for more than 25 additional slices.
DISCLAIMER: It is not recommended that one executes the training by this time. That is due to the heavy processing it does. For instance, with the setup presented below, the complete training took over 15h to execute. An execution with CPUs was not tested.
It is possible to run the training. One should take the following steps to do so:
- Run the notebook dev/prepare_data.ipynb
- Verify that the file crop_data.h5 was generated
- Run deliver/train-9-3-5-64.ipynb
The training was executed on a server with 8 NVIDIA V100-32GB and the following setup:
Intel(R) Xeon(R) Gold 6134 CPU @ 3.20GHz
total used free shared buff/cache available
Mem: 1.0T 2.8G 96G 35M 908G 1.0T
Swap: 975M 3.5M 972M
This work can be executed in a docker environment, with no previous dependencies, other than docker-ce (or docker.io) on your machine.
In order to run in a docker, with docker installed on your machine:
- Clone this repository, with:
git clone https://github.com/ferseiti/reproducibility
- Navigate to the cloned directory like so:
cd <path_to>/reproducibility
- Then run the docker:
docker run -it --rm -p 12345:8888 -v $PWD:/home ferseiti/jupyter:reproducible
Finally, just access, through your browser, the address http://localhost:12345.
Note that the port 12345 can be changed if 12345 is already in use by your OS.
Once in the jupyter, open deliver/Final_Version.ipynb and run the paper.
