pytorch-cxr

Chest X-ray multi-labels binary classifications

Prerequisites

• Python 3.6+
• PyTorch 1.1.0+

Datasets

• Stanford CheXpert dataset
• MIT MIMIC-CXR dataset

You need to resize and copy from the extracted original data. Modify {src,tar}_path at the end of dataset.py properly for your environment.

$ python dataset.py

Training

If no --cuda <comma-separated gpu-ids> given, cpus will be used for training. If no distributed environment, the first cuda id is used for the training and the others are ignored.

$ python train.py --cuda <comma-separated gpu-ids> --runtime-dir <your-runtime-dir> [--tensorboard] [--slack]

If you want to run in distributed mode, please refer to here.

$ python -m torch.distributed.launch --nproc_per_node=<NUM_GPUS_YOU_HAVE> train.py --cuda <gpu-ids> --runtime-dir <runtime-dir> [any optinal switches]

or using the train.sh with proper modification of node id etc.

if you want to log to slack, add .slack file in yaml format:

token: <your bot-slack-app token>
recipients: <list of log message recipients username>

Prediction

$ python predict.py <study-dir-to-be-predicted>