Daniel Lawson
Daniel Lawson
Hi Danijar, Great research and implementation. Do you have future plans to release pre-trained model weights for some environments? This could aid in research that aims to study the transfer...
### Parallel episode sampling I have a use case where we have a dataset consisting of image-based observations, and I notice that sampling speed seems to be slower than with...
## Background [BabyAI](https://arxiv.org/abs/1810.08272) is a "gridworld environment whose levels consist of instruction-following tasks that are described by a synthetic language". Gato generates their dataset using the built-in BabyAI bot, with...
We have a guide on doing distributed training w/ Vast here: https://docs.google.com/document/d/1W_dN3qarCOcLRDdEZ75LBtkLGiwUziWWDtVTjd43Ad4/edit?usp=sharing . However, we have not performed full distributed training runs. This issue does not specify specific issues, but...
We want fast model inference for quickly evaluating the performance of our checkpoints during training or after training. Evaluation is called in trainer.py, looping over each task: https://github.com/ManifoldRG/gato-control/blob/93009abfaa1e0a9efcfcb8eba1435352dfdbcd4b/gato/training/trainer.py#L77-L84 task.evaluate() a...
This issue relates to how we embed each (16x16) patch. Additionally, we discuss the positional encodings we add to each patch's embedding. # Patch Embedding Let's review, we split the...
This issue does not go into all the detail regarding our dataset considerations, but I am currently converting datasets for 45 Atari games to Minari. I utilize dqn-replay, which in...
Make this in parallel: https://github.com/ManifoldRG/gato-control/blob/0f4f7d2c8a31ca1e6a6e85a6d3f0d763910bfb39/gato/training/trainer.py#L148-L165 and potentially Minari's own function: https://github.com/Farama-Foundation/Minari/blob/c0669fc3a8829dec4a7a1fbee198a6be4f668ea1/minari/dataset/minari_storage.py#L153 When training on several games w/ image observations (Atari) the training loop is too slowed and is being dominated...
Performed: - 3 MuJoCo tasks - Breakout (on a small, relatively arbitrary subset of steps converted to Minari) What's next __imediately__: - Convert more Atari data to Minari (The details...
Gato's prompting for control tasks:  We follow this general strategy, but fill the gaps as some details are not specified. Let's look at some arguments related to prompting in...