mxnet starter for https://www.kaggle.com/c/statoil-iceberg-classifier-challenge modified from https://github.com/zackchase/mxnet-the-straight-dope
download train.json and test.json to the folder input, run ''' ./creat_files.sh ''' to create idx files, a 80% train split is used, need to be improved.
I tried resnet18 as an example, in script folder
''' python gluon.py '''
Predict script need to be modified to have the stored params from traning '''
'''
1, It seem that some prediction is 0 or 1 which make leadbboard score bad, need to find a better way to cap the predicted values. Using this script as is can only get 0.61 in public lb. A larger resolution is used which may be a bad idea. Data generation is navie, may affect result, at least should not use imagenet mean and var