Deep Learning on Biomedial Data

This Repository aims to provide boiler plate code to quickly train, test, and evaluate deep learning architectures in Pytorch.

Install Guide

There are two ways to install the package:

# 1.) First clone the repo and then install it where it is located
git clone https://github.com/pgruening/dlbio.git
cd dlbio
pip install -e .

# 2.) Pip install directly from Github
pip install git+https://github.com/pgruening/dlbio

# Note, if you use docker, you may also need to run this:
RUN pip install git+https://github.com/pgruening/dlbio
RUN apt update
RUN apt install libgl1-mesa-glx -y

Interesting modules:

• helpers.py: utility functions for path operations, number conversions, detection rectangles, and so on. With MyDataFrame, a simple data structure based on dictionaries is provided to quickly setup Pandas DataFrames.
• pt_training.py: contains a class ‘training’ that comprises all necessary objects (optimizer, loss-functions, data-loader, …) and, upon calling, runs a typical pytorch training session. Furthermore, a default ArgumentParser is given that functions as a standard interface to select hyperparameters. Training results are saved to a log-file that is managed by the Printer.
• pt_train_printer.py: an object that prints intermediate training results in the terminal window and additionally, writes them to a .json file.
• exe_log_gui.py: opens a TKinter GUI in the current directory. Searches it and all subfolders for json files (search can be narrowed by with regular expressions). The found files are presented in a list. One can select a file that is assumed to be a dictionary with strings as keys and lists as values. The keys can be selected and plotted in a window.
• pytorch_helpers.py: contains some convenience functions specifically for the use of pytorch. For example, getting the current device, count the number of parameters in a model, or transform a torch image tensor to a numpy image array.
• embedding_gui.py: provides classes to visualize embeddings. It is possible to click on one of the datapoints in the embedding and display additional information about it in a new window.
• process_image_patchwise.py: meant for segmenting large images with a network. The whole image is processed in patches, inspired by the ‘seamless tiling’ strategy in the u-net paper.
• ds_pt_dataset.py: boiler plate code to implement custom datasets. For example, there is a SegmentationDataset class, that can be used on small sets that fit into RAM.
• pt_run_parallel.py: useful for running several training processes. ITrainingProcess is a good interface to quickly setup and run a training process.

Naming Conventions:

Module prefixes

• ‘pt’: pytorch, any module that imports, or at least heavily uses, pytorch packages.
• ‘ds’: dataset, should contain a class that inherits a pytorch Dataset and has a getter function for an appropriate pytorch DataLoader.
• ‘exe’: execute, modules that are meant to be called directly in the terminal. In general, exe modules should not be imported by other modules.
• ‘run’: a module that can be executed, but heavily relies on input parameters. Hence, it should be called by an exe module (e.g. with a class that implements ITrainingProcess).

Class Prefixes

• ‘ISomeClass’: interface. Similar to interfaces in other languages, this class should not be instantiated, but rather, other classes should inherit it and implement the necessary functions.