tkchris93 / forwardthinking Goto Github PK

I see that in table Paragraph 5, page 7, FTDRF with 500 trees achieves almost the same accuracy as FTDRF with 2000 trees and gcForest with 4000 trees. Do you have any thoughts on this apparent overcapacity and how one can come up with the most economical model? Thanks

Btw, thank you and @kingfengji for publishing such interesting alternatives to DNNs.

... can you publish results for CIFAR-10? Zhou and Ji Feng published a supplement to their v2 article where their gcForest showed significantly worse accuracy than ResNet and AlexNet, which was quite disappointing given the claims they made. Also, I don't think MNIST is regarded as challenging these days.

hi~thanks a lot for sharing your codes!
I forked this repo and read the codes, it seems that the FTDRF.py file contains only the codes that should exist in FTDRF.sh, and the FTDRF.sh is actually empty. Is this a typo in the file name?

Hey there,

I tried running the Deep RF and DNN code. I got the DNN code working no problem, but the Deep RF code apparently has a couple issues. In FTDRF_test.py ("y_sp = y_sp.astype('uint8')") seems to fail as y_sp was not yet initialized. After commenting that out (it doesn't appear to be used at all), it then fails in structure.py at the line "for tree in models[count].estimators_: # make half of the trees completely random Decision Trees" - similar issue (I think) - to me it appears as though the trees have not been generated yet and thus the forest has no attribute "estimators_" (the error generated is an AttributeError when trying to iterator over the "esimators_" attribute in the RandomForestClassifier).

Is there something I'm missing here?

Thanks,

Ryan