Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
VN-Transformer: Rotation-Equivariant Attention for Vector Neurons
Authors: Serge Assaad, Carlton Downey, Rami Al-Rfou', Nigamaa Nayakanti, Benjamin Sapp
TMLR 2023 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical analysis. Finally, we evaluate our VN-Transformer on (i) the Model Net40 shape classification task, (ii) a modified Model Net40 which includes per-point non-spatial attributes, and (iii) a modified version of the Waymo Open Motion Dataset trajectory forecasting task (see Section 8). |
| Researcher Affiliation | Collaboration | Serge Assaad EMAIL Duke University Carlton Downey EMAIL Waymo LLC Rami Al-Rfou EMAIL Waymo LLC Nigamaa Nayakanti EMAIL Waymo LLC Ben Sapp EMAIL Waymo LLC |
| Pseudocode | No | The paper describes methods and architectures through text and diagrams (e.g., Figure 1, 2, 3, 4, 5) and mathematical definitions, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present procedures in a code-like structured format. |
| Open Source Code | No | The paper does not contain an explicit statement or link indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | We evaluate our VN-Transformer classifier on the commonly used Model Net40 dataset (Wu et al., 2015), a 40-class point-cloud classification problem. The Model Net40 dataset (Wu et al., 2015) is publicly available at https://modelnet.cs.princeton.edu We evaluate the model on a simplified version of the Waymo Open Motion Dataset (WOMD; Ettinger et al., 2021): The Waymo Open Motion Dataset (Ettinger et al., 2021) is publicly available at https://waymo.com/open/data/motion/ under a non-commercial use license agreement. Full license details can be found here: https://waymo.com/open/terms/. |
| Dataset Splits | Yes | We select 4904 trajectories (3915 for training, 979 for testing). Each trajectory consists of 91 [x, y, z] points for a single vehicle sampled at 5 Hz. We use the first 11 points (the past) as input and we predict the remaining 80 points (the future). |
| Hardware Specification | Yes | We trained our models on TPU-v3 devices. which are accessible through Google Cloud. Our longest training jobs ran for less than 3 hours on 32 TPU cores. |
| Software Dependencies | No | Optimizer Adam W (Loshchilov & Hutter, 2019). While the optimizer is specified, no specific software library or its version number (e.g., TensorFlow, PyTorch, JAX, etc.) is provided to reproduce the exact software environment. |
| Experiment Setup | Yes | Table 5: Model hyperparameter ranges for Model Net40, Model Net40 Polka-dot, and Waymo Open Motion Dataset. Hyperparameter Value/Range Feature dimension of VN-Transformer {32, 64, 128, 256, 512, 1024} Number of attention heads {4, 8, 16, 32, 64, 128} Hidden layer dimension in encoder s VN-MLP {32, 64, 128, 256, 512} Learning rate 10 3 Learning rate schedule Linear decay Optimizer Adam W (Loshchilov & Hutter, 2019) Epochs 4000 ϵ of VN-Linear With Bias {0, 10 6} |