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
| Dataset | sNDF | dNDF |
|---|---|---|
| MNIST | 0.9794 | 0.9963 |
| UCI_Adult | 0.8558 | NA |
| UCI_Letter | 0.9507 | NA |
| UCI_Yeast | 0.6031 | NA |
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.