Sanghyeok Choi

안녕하세요. Thumbnail이 어떤 경우엔 post에 있는 사진이 자동으로 등록이 되고, 어떤 경우엔 그렇지 않더라구요. 자동 등록이 되지 않은 경우에 Thumbnail을 따로 업로드 하면 본문 상단에 자동 추가되는 게 마음에 안...

[Feat] Add Local Search for Solution Improvement

2

# **Open for Review; Do Not Merge for Now!** ## Description Add local search operators as a post-processing to improve a given solution. Here we use the [LocalSearch](https://github.com/PyVRP/PyVRP/tree/main/pyvrp/search) operator provided...

Add DAG-GFlowNet (Bayesian Structure learning, Deleu et al., 2022)

1

### PR Summary This PR includes the `BayesianStructure` env to reproduce [DAG-GFlowNet](https://arxiv.org/abs/2202.13903). You can find a script for training with an MLP in `tutorials/examples/train_bayesian_structure.py`. ### Issues - Training is very...

Currently, `Trajectories` with `GraphStates` do not support `Trajectories.reverse_backward_trajectories`. The `LocalSearchSampler`, which depends on `Trajectories.reverse_backward_trajectories`, also cannot be supported. The tricky part is `states.tensor.transpose(0, 1)`, meaning that `states.tensor` is assumed to...

Issues in `ReplayBuffer`

12

Here I note several issues in the current `ReplayBuffer`, including the points raised by @younik in #193. 1. **Types of objects**, (#193): Currently, `ReplayBuffer` can take `Trajectories`, `Transitions` (which are...

Refactoring ideas regarding `Sampler`

9

There are two key components, `GFlowNet` (in `src/gfn/gflownet`) and `Sampler` (in `src/gfn/samplers.py`). I think `Sampler` is a redundant class, and we could get rid of it by incorporating the sampling...

I'm not sure, but I believe the goal of this project is to support 1. Various GFN algorithms, including advanced techniques, e.g., [Forward-looking](https://arxiv.org/abs/2302.01687), [Local search](https://arxiv.org/abs/2310.02710), or [Reward augmentation](https://arxiv.org/abs/2210.03308) 2. Various...

Suggestion to improve the Tutorials

1

I believe the tutorial for torchgfn could be more friendly with the following ideas: - Using Jupyter Notebook with Google Colab support and step-by-step descriptions. - Visualization of learning progresses...

`val_data_size` and `test_data_size` have no effect.

3

The configs `val_data_size` and `test_data_size` (see [here](https://github.com/ai4co/rl4co/blob/643ef99d118ce535e615a9441838782a2decf412/configs/experiment/routing/deepaco.yaml#L26)) have no effect. It seems that the default `env.val_file` and `env.test_file` are used. You can simply check this by 1) change those configs...

- [x] I've read the [.github/CONTRIBUTING.md](.github/CONTRIBUTING.md) file - [x] My code follows the typing guidelines - [x] I've added appropriate tests - [x] I've run pre-commit hooks locally ## Description...