Rl4co

Major release: `v0.4.0` is here! 🚀

This release adds several new features and major refactorings in both modeling and environment sides!


Changelog

✨ Features
- DeepACO + ACO Furffico henry-yeh
- Non-autoregressive (NAR) models and NARGNN Furffico henry-yeh
- Add modular environment data generator with support to new distributions cbhua
- New decoding techniques based on the decoding strategy class LTluttmann
- Top-p (nucleus sampling)
- Top-k
- Select start nodes functions LTluttmann


⚙️ Refactoring
- **Major modeling refactoring** (summarized [here](https://github.com/ai4co/rl4co/pull/165)). Now we categorize NCO approaches (which are not necessarily trained with RL!) into the following: 1) constructive (AR and NAR), 2) improvement, 3) transductive. This translates into code, which is now fully customizable. For instance, in constructive methods, now encoders / decoders can be fully replaced or removed in an abstract way!
- **Major environment refactoring** (summarized [here](https://github.com/ai4co/rl4co/pull/169)): we further modularize the environments into components (logic under `env`, data generation under `generator`, and so on), with several components moved inside the `RL4COEnvBase`. Importantly, we introduce data generators that can be customized!
- Use Abstract classes if class should not be ngastzepeda


📝 Documentation
- Hydra documentation and tutorial LTluttmann
- New modularized examples under `examples/`
- Updated RL4CO structure in ReadTheDocs
- Move to MIT license with AI4CO for inclusiveness
- New RL4CO / AI4CO swag. You may also find them [here](https://github.com/ai4co/assets)!

🐛 Bug Fixes
- MatNet and FFSP bugfix LTluttmann
- Best solution gathering from POMO ahottung
- Tests now passing on MPS; compatibility with TorchRL https://github.com/pytorch/rl/pull/2125
- Miscellaneuous LTluttmann , bokveizen , tycbony

New Routing Envs and more :rocket:

Changelog

:sparkles: Features

* Add CVRPTW Environment ngastzepeda
* Add Solomon instance / solution loader via `vrplib`
* Add basic Skill-VRP (SVRP) ngastzepeda

:page_with_curl: Documentation

* [Minor] improve decoding strategies documentation

:bug: Bug Fixes

* Avoid `deepcopy` bug by not saving intermediate steps of decoding strategy 123
* Allow passing `select_start_nodes_fn` and other kwargs in decoding strategies

New Decoding Types and more :rocket:

Changelog

Features
- Beam Search 109 110 LTluttmann
- Decoding type class 109 110 LTluttmann

Documentation

- Add (simple , API work in progress!) tutorial notebooks for TSPLib and CVRPLib 84
- Add decoding strategies notebook LTluttmann + small fix Haimrich

Optimization

* `torch.no_grad` to `torch.inference_mode`
* Faster testing

Bug Fixes

* Batch size initialization ngastzepeda
* Bump up naming to align with `0.4.0` release of TorchRL
* MatNet bug fix 108

QoL and BugFixes 🚀

Changelog

- Better multi start decoding 102
- Add modular `select_start_nodes` function for POMO
- Improve efficiency of multistart function
- Add testing and selection function for more envs
- Fix OP selecting too far away nodes in POMO
- Automatic multistart, no need to manually choose beforehand when running POMO
- Fix CVRP capacity bug ngastzepeda 105
- Add critic init embedding support
- Fix data generation and add better docs 106
- Better dataset handling: add dataset choice; use low CPU usage dataset by default
- Better solution plotting and better quickstart notebook 103
- Library winter cleanup
- Miscellaneous minor fixes here and there

Faster Library, Python 3.11 and new TorchRL support, Envs, Models, Multiple Dataloaders, and more 🚀

Faster Library, new Python 3.11 and TorchRL
- Update to latest TorchRL 72, solving several issues as 95 97 (also see [this](https://github.com/wouterkool/attention-learn-to-route/issues/58))
- Benchmarking:
- Up to 20% speedup in training epochs thanks to faster TensorDict and new env updates
- Almost instant data generation (avoid list comprehension, e.g. from ~20 seconds to <1 second per epoch!)
- Python 3.11 now available 97

New SMTWTP environment
- Add new scheduling problem: Single Machine Total Weighted Tardiness Problem environment as in [DeepACO](https://arxiv.org/pdf/2309.14032.pdf) henry-yeh

New MatNet model
- Add MatNet version for square matrices (faster implementation ideal for routing problems)
- Should be easy to implement scheduling from here

Multiple Dataloaders
- Now it is possible to have multiple dataloaders, with naming as well!
- For example, to track generalization during training

Miscellaneous
- Fix POMO shapes hyeok9855 , modularizing PPO etc
- Fix precion bug for PPO
- New AI4CO transfer!

Add FlashAttention2 support ⚡

- Add FlashAttention2 support as mentioned [here](https://github.com/kaist-silab/rl4co/issues/85)
- Remove old wrapper for `half()` precision since Lightning already deals with this
- Fix `scaled_dot_product_attention` implementation in PyTorch < 2.0
- Minor fixes