Observations.txt

1. Simple model + batchnorm = 0.8 f1 score, complex vgg net like architecture is not giving good results (with and without dropout). 0.8 was achieved after upsampling to make number of positive frames = numer of negative. 
2. Using relu after batchnorm is worse (0.8 before versus 0.7 after)
3. Not using batchnorm destroys f1 score to 0. Seems like network is not able to learn anything without it.

Things to try:
1. Rename two frames before transition also to 1s.
2. Use ResNet pertained features as inputs