Diary

2018-10-09T215410EST
7 layer CNN with 128 pixels input. Too large, not learning fast enough. stuck at loss 5.15
Increase epoch. Reduce each epoch.
reduced batch size.
Going to reduce each epoch time to a few minutes.

2018-10-09T230814EST
Vastly underfitting the model.
Not going to solve the issue in a reasonable amount of time. Parameters need to be rechosen.
Reduced batch size further with reduce steps per epoch.
Intergrated tensorboard to see.
Trying to achieve over fit first before reducing it.

;time
Even first epoch, loss is lower... validation is... about the same.
crappy learning. No improvement.
Trying batch normalization and reduce dropout and randomness to introduce overfit.

1:53:
increase stride definitely helped the loss to come down MUCH faster. 10 to 3! val loss around 1.
Loss is now at 0.6! and won

10:23
without dropout, the val loss continues to rise while loss is low.
added keras save to the callback list.
Reduced dropout to 0.1, 0.2 might be too high in my case with already augmented data.
FIRST epoch you should see drastic loss decrease or else the model is not working.
adadelta might be a great idea to try see examples: http://ruder.io/optimizing-gradient-descent/index.html#challenges

14:40
0.1 dropout does not seems to be enough. But I also used AdaDelta.
Going to increase dropout in the final stage too.
So far, most network it seems within 30 minutes, I should be able to tell how well it is performing. NO point training longer than that.
Also increased batch size to see.

15:35
After increasing dropout, definitely seeing better trend where both loss and val-loss are decreasing.


15:50
Remove RGB to gray scale for more data throughput.

8:45
After increasing the view area size, the val accuracy increase from 55% to 75% .
added leaky relu and further view size to see if it helps.

todo
remove softmax to provide continuous label for w p r.

2018-10-26 20:34
Constructed the multile regression for WPR but it is probably overlearning as the validation accuracy hover around 40% at best.

Added batch normalization to all process.