Skip to content

Latest commit

 

History

History
32 lines (25 loc) · 1.81 KB

README.md

File metadata and controls

32 lines (25 loc) · 1.81 KB

BDCLearn - Dockerized Keras deep learning tools for BDCat

Jupyter Notebooks

The included Jupyter notebooks are self-contained tools for training each network type. Each notebook contains preprocessing and training with an example dataset, but the intent is that the user can replace to preprocessing section with a custom script and data can be imported.

Pretrained models

Models pretrained with example datasets are available
UNet - trained with lung segmentation masks in the COVID-19 dataset from http://medicalsegmentation.com/covid19/
LeNet - trained with the Fashion MNist set included in Keras

Docker

Networks

Current : UNet
To Do: VGG-16, LeNet, Resnet, SegNet, ...

Commands

/keras/{network name} new {kwargs} : create a new hdf5 model file with the selected network structure
/keras/{network name} train --model_path --image_path --class_path {kwargs} : train a model from image directories (see https://keras.io/api/preprocessing/image/#flowfromdirectory-method)

In progress:
/keras/{network name} test --model_path --image_path --class_path {kwargs} : test a trained model for image directories
/keras/{network name} predict --model_path --image_path : use a model to predict classes from an image directory

Using custom preprocessing scripts in Python

Most use cases will involve customized preprocessing or directory trees not compatible with the Keras flow_from_directory method. By using the new command (see above), the model can be loaded into a custom script without having to instantiate it manually. The scripts in the util directory contain functions that can be called from a custom script to simplify coding for training, testing, and prediction.
This section will be expanded with a detailed description of the available functions