NOTE: This code was originally written by Yuan Xue and Sharon Huang at Lehigh University. I have put this into a GitHub repository so I can maintain configuration control of the code while I make changes to get it to work in the NLM/Lister Hill Center environment - Mike Bopf
1. Pytorch available on http://pytorch.org/
TensorFlow available on Github: https://github.com/tensorflow/tensorflow. (We use tensorboard in TF to monitor the training process)
2. Python 2.7 or Python 3.5/3.6
3. In addition, please pip install the following packages:
numpy
scipy
pillow
If I miss anything here, you may encounter some errors and please let me know those errors.
4. Using Virtual Environment or pip to install simpleITK
1. download dataset from https://sites.google.com/site/braintumorsegmentation/home/brats2015 (We trained our model on Brats 2015 training data)
2. Preprocess the dataset using prepare_data.py
$ python prepare_data.py
Note that you may need to change path in __init__ method in class Preprocess (line 47 in prepare_data.py). The path should contain HGG/ folder and LGG/ folder.
Also, after run prepare_data.py, you may need to manually create a val/ folder under root folder and randomly move some training samples from train/ folder to val/ for validation.
The number of samples in val/ is better be multiple of 5 so that you can always see 5 validation images in tensorboard.
3. To train a model with the preprocessed dataset:
$ CUDA_VISIBLE_DEVICES=X python train_adversarial.py --cuda --batchsize 15
Note that X means the id of your GPU, for now we only support training with ONE GPU. If you have only one GPU, then X is 0.
The number after --batchsize is the training batch size. I trained our model on a Titan X Pascal GPU with 12G memory. If you have access to a GPU with large memory such as 12G, you can keep this number. Otherwise you may need to use a smaller batchsize. I suggest you to use GPU with at least 6G memory, then you can set the batch size to be 6 or 7. If you encounter some errors such as "...out of memory...", then probably you need to use a GPU with larger memory or smaller batchsize.
The default output folder is SegAN/, you can find tensorboard event file and trained models in that folder.
4. The code is just for training for now, and I haven't cleaned the code, sorry for inconvenience.
To monitor the training process, please run the command below under your Brats/ folder (the root folder for your codes and data)
$ tensorboard --logdir='./' --port=6006 --reload_interval=5
Then, open your browser and go to: http://0.0.0.0:6006, then you can monitor the training process via tensorboard.
HOWEVER, I'm having some trouble using pytorch and tensorflow together, it worked pretty well but after I upgraded cuda, pytorch, tensorflow, etc. I always have "dlopen: cannot load any more object with static TLS" when I import torch and tensorflow together. It may work for you though.
If you also encouter any problems with tensorboard, you can delete the code for tensorboard which is seperated by #============ TensorBoard logging ============#
If you have any questions, please feel free to contact me, my email is yux715@lehigh.edu.