Plot with evolution of MSB terms during training

mtewes / tenbilac

Neural network for inverse regression problems

1 stars 2 forks source link

Closed mtewes closed 7 years ago

mtewes commented 7 years ago

Status report: figure_1

kuntzer commented 7 years ago

Nice. The fact that the colours become lighter with iteration means that the differences between the samples increase?

mtewes commented 7 years ago

No, alpha and lightness do not change here. It's just that subtle differences between cases become visible only after a bit of training.

kuntzer commented 7 years ago

okay!

mtewes commented 7 years ago

In a more realistic case: screen shot 2016-12-07 at 17 07 58

Issues with this

many iterations -> takes long time to plot. Don't save every iteration ?
after a restart of the training, training data is shuffled, and colors (features) that the plot uses don't match to the biases that where saved before. So only the last block of colors is good.

It's interesting to see and understand this plot, but maybe this is not a top priority.

mtewes commented 7 years ago

Done with the plot -- it is activated by a kwarg to the training constructor. Will now find a solution to pass it as option from MegaLUT.

mtewes commented 7 years ago

Note that the problem (above) with the many iterations is now fixed. We only save some snapshots.

kuntzer commented 7 years ago

:+1:

mtewes commented 7 years ago

Done and merged.