lmartak / distill-nn-tree Goto Github PK

Hi thank you for sharing this code but I had a question :
what do you think about the first log used inside the loss function that you used in line 199 what is the aim of this log

Also why did you move the minus near the second log not the first log as stated in the paper.?

Thank you

Hi , thanks for sharing this useful code, I was wondering what is the goal of self.eps attribute for the Soft Tree ?

Thank you

Hi, I ran your code on my own server, the results are:
[No distill]
10000/10000 [==============================] - 8s 783us/sample - loss: 7.5134 - acc: 0.9085accuracy: 90.85% | loss: 7.513414146804809
10000/10000 [==============================] - 8s 785us/sample - loss: 7.5161 - acc: 0.9032
accuracy: 90.32% | loss: 7.516079863357544

[distill with soft target]
Saving trained model to assets/distilled/tree-model.
10000/10000 [==============================] - 7s 711us/sample - loss: 7.7522 - acc: 0.8254
accuracy: 82.54% | loss: 7.7521795679092405
10000/10000 [==============================] - 8s 758us/sample - loss: 7.7434 - acc: 0.8189
accuracy: 81.89% | loss: 7.7433844789505

thank u very much

Thanks for sharing! How exactly do you label the paths and especially leafs? Are you storing the probabilities to look for likeliest labels at each node and look in which leafs data points ended up from your training set?

Thanks

First of all, thank you for such a good code. I want to ask, when I input a larger size, such as 224 * 224 * 3, I find that the training has no effect, is it necessary to change some parts of the code?