TorchRL: A data-driven decision-making library for PyTorch

Authors: Albert Bou, Matteo Bettini, Sebastian Dittert, Vikash Kumar, Shagun Sodhani, Xiaomeng Yang, Gianni De Fabritiis, Vincent Moens

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we experimentally demonstrate its reliability and flexibility and show comparative benchmarks to demonstrate its computational efficiency.
Researcher Affiliation Collaboration Albert Bou UPF, Acellera Matteo Bettini University of Cambridge Sebastian Dittert UPF Vikash Kumar Meta AI Shagun Sodhani Meta AI Xiaomeng Yang Meta AI Gianni De Fabritiis ICREA, UPF, Acellera Vincent Moens Py Torch, Meta vmoens[at]meta.com
Pseudocode No Figures 5 and 6 contain code examples but are not presented as pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes Torch RL fosters long-term support and is publicly available on Git Hub for greater reproducibility and collaboration within the research community. The code is open-sourced on Git Hub.
Open Datasets Yes Torch RL offers a range of downloadable datasets for this purpose (such as D4RL (Fu et al., 2020) or Open ML (Vanschoren et al., 2013)).
Dataset Splits No The paper mentions using D4RL datasets and MuJoCo/Atari environments but does not specify explicit train/validation/test splits (e.g., percentages or counts) within the paper itself.
Hardware Specification Yes These experiments were run on an AWS cluster with a single node (96 CPU cores and 1 A100 GPUs per run).
Software Dependencies Yes Compatibility with all versions of Gym is guaranteed starting from v0.13, including the latest transition to Gymnasium. We measured Tensor Dict overhead for the lastest release at the time of writing the manuscript (v0.2.1) against other existing solutions.
Experiment Setup Yes Tables 6, 7 and 8 display all hyperparameters values and network architecture details required to reproduce our online RL results. Table 9 displays all hyperparameter values and network architecture details required to reproduce our offline RL results.