An implementation of the Deep Neural Decision Forests(dNDF) in PyTorch.

• Two stage optimization as in the original paper Deep Neural Decision Forests (fix the neural network and optimize $\pi$ and then optimize $\Theta$ with the class probability distribution in each leaf node fixed )
• Jointly training $\pi$ and $\Theta$ proposed by chrischoy in his work Fully Differentiable Deep Neural Decision Forest
• Shallow Neural Decision Forest (sNDF)
• Deep Neural Decision Forest (dNDF)

MNIST, UCI_Adult, UCI_Letter and UCI_Yeast datasets are available. For datasets other than MNIST, you need to go to corresponding directory and run the get_data.sh script.

• Python 3.x
• PyTorch >= 1.0.0
• numpy
• sklearn

python train.py --ARG=VALUE

in the case of training the sNDF on MNIST with alternating optimization, the command is like

python train.py -dataset mnist -n_class 10 -gpuid 0 -n_tree 80 -tree_depth 10 -batch_size 1000 -epochs 100

Not spending much time on picking hyperparameters and without bells and whistles, I got the accuracy results(obtained by training $\pi$ and $\Theta$ seperately) as follows:

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.

Some people may experience the 'loss is NaN' situation which could be caused by the output probability being zero. Please make sure you have normalized your data and used a large enough tree size and depth. In the case that you want to stick with your tree setting, a workaround could be to clamp the output value.