- Fix the current bug discussed in #844 in `test_ppo.py`.
- Add warning for `ActorProb ` if both `max_action ` and
`unbounded=True` are used for model initializations.
- Add warning for PGpolicy and DDPGpolicy if they find duplicate usage
of action-bounded actor and action scaling method.
- [x] I have marked all applicable categories:
+ [ ] exception-raising fix
+ [x] algorithm implementation fix
+ [ ] documentation modification
+ [ ] new feature
- [x] I have reformatted the code using `make format` (**required**)
- [x] I have checked the code using `make commit-checks` (**required**)
- [x] If applicable, I have mentioned the relevant/related issue(s)
- [x] If applicable, I have listed every items in this Pull Request
below
While trying to debug Atari PPO+LSTM, I found significant gap between
our Atari PPO example vs [CleanRL's Atari PPO w/
EnvPool](https://docs.cleanrl.dev/rl-algorithms/ppo/#ppo_atari_envpoolpy).
I tried to align our implementation with CleaRL's version, mostly in
hyper parameter choices, and got significant gain in Breakout, Qbert,
SpaceInvaders while on par in other games. After this fix, I would
suggest updating our [Atari
Benchmark](https://tianshou.readthedocs.io/en/master/tutorials/benchmark.html)
PPO experiments.
A few interesting findings:
- Layer initialization helps stabilize the training and enable the use
of larger learning rates; without it, larger learning rates will trigger
NaN gradient very quickly;
- ppo.py#L97-L101: this change helps training stability for reasons I do
not understand; also it makes the GPU usage higher.
Shoutout to [CleanRL](https://github.com/vwxyzjn/cleanrl) for a
well-tuned Atari PPO reference implementation!
* When clip_loss_grad=True is passed, Huber loss is used instead of the MSE loss.
* Made the argument's name more descriptive;
* Replaced the smooth L1 loss with the Huber loss, which has an identical form to the default parametrization, but seems to be better known in this context;
* Added a fuller description to the docstring;
- collector.collect() now returns 4 extra keys: rew/rew_std/len/len_std (previously this work is done in logger)
- save_fn() will be called at the beginning of trainer
- Batch: do not raise error when it finds list of np.array with different shape[0].
- Venv's obs: add try...except block for np.stack(obs_list)
- remove venv.__del__ since it is buggy
Change the behavior of to_numpy and to_torch: from now on, dict is automatically converted to Batch and list is automatically converted to np.ndarray (if an error occurs, raise the exception instead of converting each element in the list).
This PR focus on some definition change of trainer to make it more friendly to use and be consistent with typical usage in research papers, typically change `collect-per-step` to `step-per-collect`, add `update-per-step` / `episode-per-collect` accordingly, and modify the documentation.
This is the third PR of 6 commits mentioned in #274, which features refactor of Collector to fix#245. You can check #274 for more detail.
Things changed in this PR:
1. refactor collector to be more cleaner, split AsyncCollector to support asyncvenv;
2. change buffer.add api to add(batch, bffer_ids); add several types of buffer (VectorReplayBuffer, PrioritizedVectorReplayBuffer, etc.)
3. add policy.exploration_noise(act, batch) -> act
4. small change in BasePolicy.compute_*_returns
5. move reward_metric from collector to trainer
6. fix np.asanyarray issue (different version's numpy will result in different output)
7. flake8 maxlength=88
8. polish docs and fix test
Co-authored-by: n+e <trinkle23897@gmail.com>
This is the PR for QR-DQN algorithm: https://arxiv.org/abs/1710.10044
1. add QR-DQN policy in tianshou/policy/modelfree/qrdqn.py.
2. add QR-DQN net in examples/atari/atari_network.py.
3. add QR-DQN atari example in examples/atari/atari_qrdqn.py.
4. add QR-DQN statement in tianshou/policy/init.py.
5. add QR-DQN unit test in test/discrete/test_qrdqn.py.
6. add QR-DQN atari results in examples/atari/results/qrdqn/.
7. add compute_q_value in DQNPolicy and C51Policy for simplify forward function.
8. move `with torch.no_grad():` from `_target_q` to BasePolicy
By running "python3 atari_qrdqn.py --task "PongNoFrameskip-v4" --batch-size 64", get best_result': '19.8 ± 0.40', in epoch 8.
This is the PR for C51algorithm: https://arxiv.org/abs/1707.06887
1. add C51 policy in tianshou/policy/modelfree/c51.py.
2. add C51 net in tianshou/utils/net/discrete.py.
3. add C51 atari example in examples/atari/atari_c51.py.
4. add C51 statement in tianshou/policy/__init__.py.
5. add C51 test in test/discrete/test_c51.py.
6. add C51 atari results in examples/atari/results/c51/.
By running "python3 atari_c51.py --task "PongNoFrameskip-v4" --batch-size 64", get best_result': '20.50 ± 0.50', in epoch 9.
By running "python3 atari_c51.py --task "BreakoutNoFrameskip-v4" --n-step 1 --epoch 40", get best_reward: 407.400000 ± 31.155096 in epoch 39.
