jingxil / neural-decision-forests Goto Github PK

There were some additional requirements not mentioned in the README (or a requirements file). That might be because of an assumption that the user has Anaconda instead of a vanilla pip environment. In any case, this also requires sklearn and scipy, which you don't get from merely pip installing pytorch.

while updating pi with theta fixed, the result of new_pi in the first step is too large, using softmax to normalize new_pi will get output near to zero. And in next step wo get NAN new_pi

As mentioned in the README file, 'By adding the nonlinearity in the routing function, the accuraries can reach 0.6502 and 0.9753 respectively on the UCI_Yeast and UCI_Letter'.

Actually, I have tried many settings but the accuracies are not that high. So could you tell me how to add non-linearity to get good results (e.g. yeast and letter)? Thank you very much!

Since PyTorch updated, 'volatile' no longer works in the newest version. Running the current code will result in CUDA memory error if 'volatile' was not changed into 'with torch.no_grad():'.

Hi Jingxil, thanks for making this great repository public for everyone! I really appreciate your work here. And I firstly tried MNIST classification using the default setting, everything is perfect. Issues came when I tried UCI_Adult classification, like shown below:
WARNING: RUN WITHOUT GPU Use adult dataset Epoch 1 : Two Stage Learing - Update PI Traceback (most recent call last): File "train.py", line 217, in <module> main() File "train.py", line 213, in main train(model, optim, db, opt) File "train.py", line 124, in train target_batches.append(cls_onehot[target]) IndexError: tensors used as indices must be long, byte or bool tensors

The command I used is:
python train.py -dataset adult -n_class 14 -n_tree 80 -tree_depth 10 -batch_size 1000 -epochs 100
Please note that the n_class parameter is adjusted accordingly already.

Do you have any idea where I did things not correctly? Thanks!

The routing function is a running product of decisions where each decision is a value obtained by employing non-linearity (sigmoid) over f. As you have you mentioned in the readme file, "By adding the nonlinearity in the routing function, the accuraries can reach 0.6502 and 0.9753 respectively on the UCI_Yeast and UCI_Letter." The routing function is already a product of values obtained after non-linear operation. So, how and where to again implement this non-linearity?