Skip to content


Repository files navigation

Kaggle Diabetic Retinopathy Solution.

Fifth place solution for the Kaggle Diabetic Retinopathy competition including some trained network. For more information see my blog post.

The code is quite badly written (lots of ideas, not so much time) but I figured it was more important to release it quickly than to rewrite it all and only publish it weeks later. The most interesting file will probably be the notebook at notebooks/Sample_prediction.ipynb which shows how to load the model dump, do some predictions with it and see the activations for the different layers.

Important to note: the code is built on top of Lasagne at commit cf1a23c21666fc0225a05d284134b255e3613335. For Theano I have been using the latest master and have no problems. Some specific versions:

  • Theano: 9a653e3e91c0e38b6643e4452199931e792a24a2
  • Lasagne: cf1a23c21666fc0225a05d284134b255e3613335
  • Numpy: 1.9.2
  • Pandas: 0.15.2
  • Scikit-learn: 0.16.0
  • Scipy: 0.15.1
  • IPython: 3.0.0
  • Matplotlib: 1.4.2

The basic model included is dependent on cuDNN (but not tested with the latest cuDNN 3 RC). I have, however, also made an export of the raw parameter values for the network using You could then replace the cuDNN layers by the layers of your choice or use these parameters to initialise layers in other frameworks.


Fifth place solution of the Kaggle Diabetic Retinopathy competition.







No releases published