kaggle_retinopathy_starter.torch
A starter kit in Torch for Kaggle Diabetic Retinopathy Detection.
It showcases:
- Classification
- Regression
DID NOT FINISH IMPLEMENTING:
- Metric Learning (Siamese and triplet networks)
- Averaging model ensembles
What else?
- 1-GPU or multi-GPU convolution neural networks
- multi-threaded data loading (data is loaded in compressed-form into memory and decompressed + jittered on the fly)
- test script to take your trained model (or models) and produce Kaggle-compatible CSV file ready for upload
Getting started
-
Install torch + dependencies, instructions are here: INSTALL.md
-
If you already have torch installed, update it and it's packages. These scripts use some recently added features.
-
Run script with:
th main.lua
##Ideas to try
##Broad Strategy
- Train several models
- Average the network predictions over many of these models
- For each image, average the prediction over several crops and orientations of the image (for example at test time, every image's prediction is the average of 4 rotations of the image)
- Win