In this repo you can find jupyter notebook with code which I used to train neural network in HackerEarth Deep Learning Challenge 2.
During the challenge I had to identify the class of thorax diseases from the given chest x-ray images. My trained neural network took 2nd place.
- jupyter notebook,
- keras,
- pandas,
- numpy,
- sklearn
In machine learning the most important thing is to search all data which you can use to train your model. So I did my homework and searched on the net all lungs X-ray images. Besides data available in HackerEarth datasets I've used additional ones which I found on the web. I also generated 10k images for classes which have smallest number of images.
I used transfer learning.
Firstly I imported ResNet50 architecture with imagenet weights. I frozen all layers except full connected layer. I ran 20 epochs with optimizer Adam(0.0001, decay=0.00001). Then I unfrozen last 35 layers and ran 20 epochs with optimizer Adam(0.0001, decay=0.00000001). I used AWS p2.xlarge architecture to train my model.