fdujay / pytorch-dr Goto Github PK

• Implementation of team o_O solution for the Kaggle Diabetic Retinopathy Detection Challenge in pytorch.
• Solution summary

• Branch "master" tries to apply new techniques to improve team o_O solution.
• Branch "reimplement" is the reimplementation of team o_O solution.
• Branch "two_cates" is used to do binary classification (Normal/DR).

Your should organize your dataset as following:

├── your_data_dir
	├── train
		├── 1
			├── a.jpg
			├── b.jpg
			├── ...
		├── 2
			├── x.jpg
			├── y.jpg
			├── ...
		├── 3
		├── 4
		├── 5
	├── val
	├── test

Here, val and test directory have the same structure of train. The name of directories 1- 5 means the severity of disease.

You can run this code by:

$ python main.py

Most of hyperparameters and configurations are in config.py. You should choose SMALL_NET_CONFIG, MEDIUM_NET_CONFIG or LARGE_NET_CONFIG as STEM_CONFIG in main function of main.py. Function stem will train one inference network and function blend will train a ensemble network which is optional.

def main():
    # network config, you should .
    STEM_CONFIG = SMALL_NET_CONFIG
    stem(STEM_CONFIG)
    
    # blend step config
    # BLEND_CONFIG = BLEND_NET_CONFIG
    # blend(BLEND_CONFIG, STEM_CONFIG)

Moreover, if you want to get the final large network, you should train small and medium network first. More details are in o_O_solution_report.pdf.

This project is still in progress. Now a single large network can achieves 79.84% in EyePACs test set which is close to 80% that the author claims. You can use full ensemble methods that the author designed to get a better result but I haven't implement it yet.