Conditional Permutation Invariant Flows
Authors: Berend Zwartsenberg, Adam Scibior, Matthew Niedoba, Vasileios Lioutas, Justice Sefas, Yunpeng Liu, Setareh Dabiri, Jonathan Wilder Lavington, Trevor Campbell, Frank Wood
TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We illustrate the utility of this model via applications including (1) complex traffic scene generation conditioned on visually specified map information, and (2) object bounding box generation conditioned directly on images. We train our model by maximizing the expected likelihood of labeled conditional data under our flow, with the aid of a penalty that ensures the dynamics are smooth and hence efficiently solvable. Our method significantly outperforms non-permutation invariant baselines in terms of log likelihood and domain-specific metrics (offroad, collision, and combined infractions), yielding realistic samples that are difficult to distinguish from data. |
| Researcher Affiliation | Collaboration | 1Inverted AI 2Department of Computer Science, University of British Columbia 3Department of Statistics, University of British Columbia 4Mila |
| Pseudocode | No | The paper describes the dynamics functions and mathematical derivations, but does not include any clearly labeled pseudocode or algorithm blocks. The methods are explained in prose and mathematical equations (e.g., Eq. 9, Eq. 12). |
| Open Source Code | Yes | Since there are to the best of our knowledge no existing benchmark tasks for conditional set generation, as a part of our contributions we have released code to recreate the datasets and metrics in our work.1 1https://github.com/inverted-ai/conditional-permutation-invariant-flows-datasets |
| Open Datasets | Yes | The datasets used in this study are the INTERACTION datset (Zhan et al., 2019) (available for research purposes), and the CLEVR dataset (Johnson et al., 2017) (available under the Creative Commons CC BY 4.0 license). |
| Dataset Splits | No | The paper mentions training on varying amounts of agents and comparing on a 'held-out test dataset' and 'held out validation sets'. It also describes a filtering process for the traffic scene data: 'We exclude all data with less than seven agents, and cases with more than seven agents are pruned to retain only the agents closest to the center.' However, it does not specify explicit split percentages (e.g., 80/10/10), absolute sample counts for each split, or refer to standard predefined splits for the train/validation/test sets. |
| Hardware Specification | No | All our experiments were performed on a single GPU, all permutation invariant models were trained between 2 and 7 days of wall-clock time. The vanilla continuous normalizing flow, real NVP, and autoregressive model were trained over 14 days of wall-clock time. |
| Software Dependencies | No | We use the adaptive solver of Dormand and Prince of order 4 to solve the ODE (Dormand & Prince, 1980). To calculate the gradients of the ODE with respect to its parameters we use the adjoint method (Coddington & Levinson, 1955). This functionality is all available in the torchdiffeq package (Chen et al., 2018), which is the implementation we use in our experiments. |
| Experiment Setup | Yes | In all experiments gθ, fθ are implemented as neural networks of n layers and h neurons per layer. The convolutional embedding network has n layers of c channels, followed by a single feed forward layer of h neurons. We use sigmoid-linear units in all our dynamics functions... For each experiment, these parameters are presented in Table 3. |