Train a classifier for CheXpert dataset and visually explain the model.
- PyTorch
- Torchvision
- TensorboardX
- NumPy
- Matplotlib
- sklearn
- Pandas
- PIL
- RandAugment (from Ildoonet 's GitHub)
- Captum
CheXpert is one of the largest chest x-ray image dataset to date, which contains 224,316 chest radiographs of 65,240 patients. For gaining access to the CheXpert dataset, please send registration to StanfordML 's website. The dataset we used here is the smaller dataset but might be improved if you use the larger version. After downloading the dataset, you can change the directory description in config/ and validation.py to easily fit your own environment.
The file structure should be:
<Main Directory>
└─── <Our Repo>
└─── CheXpert-v1.0/
└─── (opional) ChestX-ray14/
The working directory should be at our repository.
cd BME_Final_Project_ChestXRay/
For training a new model, add a new config file in the config/. You can customize your training settings like the template file in config/base_config.yaml. PyTorch pretrained models are available except inception-v3. EfficientNet pretrained models from EfficientNet PyTorch are also available.
After adding a new config file, you can start training by
python3 main.py -c config/<your_config_file>
For monitoring training process, TensorboardX is available.
tensorboard --logdir log/<your_config_name>
To generate predictions for CheXpert Challenge
python3 predict.py <input-data-csv> <output_prediction-csv>
Our work supports the visualization from pytorch/captum.To visualize your model of a validation image,
python3 visualization.py -i <which-validation-image> -m <model#> -u <if-using-cached-threshold> -s <search-space-for-new-threshold>
Our work supports writing pseudo-label from NIH Chest X-ray dataset. Note that in current work, this method doesn't help.
python3 pseudo_labeller.py -o <output-pred-csv> -m <model#> -u <if-using-cached-threshold> -s <search-space-for-new-threshold> -r <label-resolution>