soumith / kaggle_retinopathy_starter.torch

A starter kit in Torch for Kaggle Diabetic Retinopathy Detection

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

• Maxout Networks: http://arxiv.org/abs/1302.4389
• All Convolution Net: http://arxiv.org/abs/1412.6806
• Separable Convolution filters
• Dropout/DropConnect
• PReLU: https://github.com/torch/nn/blob/master/doc/transfer.md#prelu
• Batch Normalization: https://github.com/torch/nn/blob/master/doc/simple.md#batchnormalization
• L2 Pooling: https://github.com/torch/nn/blob/master/SpatialLPPooling.lua
• Different data augmentation
• Polar/Log-polar images: https://github.com/torch/image#res-imagepolardst-src-interpolation-mode
• Different color spaces: https://github.com/torch/image#color-space-conversions
• All of Sander Deileman's tricks:
  • http://benanne.github.io/2014/04/05/galaxy-zoo.html
  • http://benanne.github.io/2015/03/17/plankton.html

##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