Welcome to the official implementation of the DTSemNet architecture, as proposed in the paper:
- “Vanilla Gradient Descent for Oblique Decision Trees,” ECAI-2024.
- "Approximation-Free Differentiable Oblique Decision Trees," JMLR-2026.
🆕 NEW: the
v2branch has a new, more tree-centric implementation of DTSemNet (seedtsemnet.py). Instead of only training a fixed architecture, you build and hold onto the underlyingObliqueDTtree directly — inspect it, edit it, graft/replace subtrees, freeze/unfreeze nodes, and update the tree structure at any time: then convert it to a trainable neural network (DTSemNet) and back. It's currently only wired up for MNIST, viatrain.py, as a usage example (hyperparameters there are not tuned).
DTSemNet is a novel invertible encoding of Oblique Decision Trees (ODT) as a Neural Network (NN) facilitating the training of ODT using vanilla gradient descent. This repository contains all the necessary files and scripts to replicate the experiments and results presented in the paper. We have reused some parts of code from CRO-DT, DGT and ICCT.
- src/dtsemnet.py: Core implementation of the DTSemNet model.
- src/net_train.py: Script for training on small classification datasets.
- src/net_train2.py: Script for training on large classification datasets, with GPU support.
- src/reg_train_linear.py: Script for training DTSemNet on regression tasks.
- /results: Directory where training logs are saved. Previous logs used in our paper are available in /results/ecai-reported.
- /results/combined_mean.py: In the case of multiple splits, the reported mean ± standard deviation are calculated by combining the results from each split.
Included in the `datasets' directory: [breast_cancer,car,banknote,balance,acute−1,acute−2,transfusion,climate,sonar,optical,drybean,avila,wine−red,wine−white]
Due to size constraints, the following datasets are not included. It needs to be downloaded to `datasets' directory.
| Dataset Name | Download Link |
|---|---|
| mnist | PyTorch Auto-Download |
| letter | Download |
| connect | Download |
| segment | Download |
| satimages | Download |
| pendigits | Download |
| protein | Download |
| sensit | Download |
| Dataset Name | Download Link |
|---|---|
| abalone | Download |
| ailerons | Download |
| cpu_active | Download |
| pdb_bind | Download |
| year | Download |
| ctslice | Download |
| ms | Download |
Please install conda environment using environment.yml. In case, there is an error in PyTorch installation, please install the same version manually.
conda env create -f environment.yml
conda activate dtsemnetAlso, install the packages in `src' directory:
python -m pip install -e .python -m src.net_train --model dtsemnet --dataset all --depth 4 -s 1 --output_prefix dtsement --verbose True- replace `dtsemnet` with `dgt` for DGT evaluation
- For time computation restrict number of cores to 8, use `taskset -cpa 0-7`
- s: Number of simulations for averaging (100 for small DTs and 10 for large DTs)
- depth: Height of the DT
- dataset: Use appropriate dataset name. `all' to evaluate all datasets
- output_prefix: helps in managing the name of the log file
python -m src.net_train2 --model dtsemnet --dataset mnist -s 1 --output_prefix dtsement --verbose True -g- Height is included in the configuration file, there is no need to specify
- model: `dtsemnet' or `dgt'
- dataset: use appropriate name of the dataset
- add `-g` for GPU operation
Use the following terminal command for regression datasets ["abalone", "ailerons", "cpu_active", "pdb_bind", "year", "ctslice", "ms"].
python -m src.reg_train_linear --model dtregnet --dataset ailerons -s 1 --output_prefix ailerons --verbose True -g- Use `--model dtregnet` for for DTSemNet-regression
- Please switch to branch `topk' for the proposed topk method which avoids STE in regression setting
- The branch is standalone, so you have to follow the README file in that branch for installation and evaluation.
- Please switch to branch rl_d (discrete action: PPO) or rl_c (continuous action: SAC) for the RL experiments.
- The branches are standalone, so you have to follow respective README files and do the environment installation for RL experiment.
If you find DTSemNet useful in your research, please cite our work:
Subrat Panda, Blaise Genest, Arvind Easwaran, and Ponnuthurai Suganthan, "Vanilla Gradient Descent for Oblique Decision Trees," European Conference on Artificial Intelligence (ECAI), 2024. Link to paper
Subrat Panda, Blaise Genest, and Arvind Easwaran, "Approximation-Free Differentiable Oblique Decision Trees," Journal of Machine Learning Research (JMLR), 2026. Link to paper