* Remove flag `eval_mode` from Collector.collect
* Replace flag `is_eval` in BasePolicy with `is_within_training_step` (negating usages)
and set it appropriately in BaseTrainer
New method training_step, which
* collects training data (method _collect_training_data)
* performs "test in train" (method _test_in_train)
* performs policy update
The old method named train_step performed only the first two points
and was now split into two separate methods
Closes: #1058
### Api Extensions
- Batch received two new methods: `to_dict` and `to_list_of_dicts`.
#1063
- `Collector`s can now be closed, and their reset is more granular.
#1063
- Trainers can control whether collectors should be reset prior to
training. #1063
- Convenience constructor for `CollectStats` called
`with_autogenerated_stats`. #1063
### Internal Improvements
- `Collector`s rely less on state, the few stateful things are stored
explicitly instead of through a `.data` attribute. #1063
- Introduced a first iteration of a naming convention for vars in
`Collector`s. #1063
- Generally improved readability of Collector code and associated tests
(still quite some way to go). #1063
- Improved typing for `exploration_noise` and within Collector. #1063
### Breaking Changes
- Removed `.data` attribute from `Collector` and its child classes.
#1063
- Collectors no longer reset the environment on initialization. Instead,
the user might have to call `reset`
expicitly or pass `reset_before_collect=True` . #1063
- VectorEnvs now return an array of info-dicts on reset instead of a
list. #1063
- Fixed `iter(Batch(...)` which now behaves the same way as
`Batch(...).__iter__()`. Can be considered a bugfix. #1063
---------
Co-authored-by: Michael Panchenko <m.panchenko@appliedai.de>
This PR adds strict typing to the output of `update` and `learn` in all
policies. This will likely be the last large refactoring PR before the
next release (0.6.0, not 1.0.0), so it requires some attention. Several
difficulties were encountered on the path to that goal:
1. The policy hierarchy is actually "broken" in the sense that the keys
of dicts that were output by `learn` did not follow the same enhancement
(inheritance) pattern as the policies. This is a real problem and should
be addressed in the near future. Generally, several aspects of the
policy design and hierarchy might deserve a dedicated discussion.
2. Each policy needs to be generic in the stats return type, because one
might want to extend it at some point and then also extend the stats.
Even within the source code base this pattern is necessary in many
places.
3. The interaction between learn and update is a bit quirky, we
currently handle it by having update modify special field inside
TrainingStats, whereas all other fields are handled by learn.
4. The IQM module is a policy wrapper and required a
TrainingStatsWrapper. The latter relies on a bunch of black magic.
They were addressed by:
1. Live with the broken hierarchy, which is now made visible by bounds
in generics. We use type: ignore where appropriate.
2. Make all policies generic with bounds following the policy
inheritance hierarchy (which is incorrect, see above). We experimented a
bit with nested TrainingStats classes, but that seemed to add more
complexity and be harder to understand. Unfortunately, mypy thinks that
the code below is wrong, wherefore we have to add `type: ignore` to the
return of each `learn`
```python
T = TypeVar("T", bound=int)
def f() -> T:
return 3
```
3. See above
4. Write representative tests for the `TrainingStatsWrapper`. Still, the
black magic might cause nasty surprises down the line (I am not proud of
it)...
Closes#933
---------
Co-authored-by: Maximilian Huettenrauch <m.huettenrauch@appliedai.de>
Co-authored-by: Michael Panchenko <m.panchenko@appliedai.de>
- [X] I have marked all applicable categories:
+ [ ] exception-raising fix
+ [ ] algorithm implementation fix
+ [ ] documentation modification
+ [X] 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
Closes#914
Additional changes:
- Deprecate python below 11
- Remove 3rd party and throughput tests. This simplifies install and
test pipeline
- Remove gym compatibility and shimmy
- Format with 3.11 conventions. In particular, add `zip(...,
strict=True/False)` where possible
Since the additional tests and gym were complicating the CI pipeline
(flaky and dist-dependent), it didn't make sense to work on fixing the
current tests in this PR to then just delete them in the next one. So
this PR changes the build and removes these tests at the same time.
Preparation for #914 and #920
Changes formatting to ruff and black. Remove python 3.8
## Additional Changes
- Removed flake8 dependencies
- Adjusted pre-commit. Now CI and Make use pre-commit, reducing the
duplication of linting calls
- Removed check-docstyle option (ruff is doing that)
- Merged format and lint. In CI the format-lint step fails if any
changes are done, so it fulfills the lint functionality.
---------
Co-authored-by: Jiayi Weng <jiayi@openai.com>
# Goals of the PR
The PR introduces **no changes to functionality**, apart from improved
input validation here and there. The main goals are to reduce some
complexity of the code, to improve types and IDE completions, and to
extend documentation and block comments where appropriate. Because of
the change to the trainer interfaces, many files are affected (more
details below), but still the overall changes are "small" in a certain
sense.
## Major Change 1 - BatchProtocol
**TL;DR:** One can now annotate which fields the batch is expected to
have on input params and which fields a returned batch has. Should be
useful for reading the code. getting meaningful IDE support, and
catching bugs with mypy. This annotation strategy will continue to work
if Batch is replaced by TensorDict or by something else.
**In more detail:** Batch itself has no fields and using it for
annotations is of limited informational power. Batches with fields are
not separate classes but instead instances of Batch directly, so there
is no type that could be used for annotation. Fortunately, python
`Protocol` is here for the rescue. With these changes we can now do
things like
```python
class ActionBatchProtocol(BatchProtocol):
logits: Sequence[Union[tuple, torch.Tensor]]
dist: torch.distributions.Distribution
act: torch.Tensor
state: Optional[torch.Tensor]
class RolloutBatchProtocol(BatchProtocol):
obs: torch.Tensor
obs_next: torch.Tensor
info: Dict[str, Any]
rew: torch.Tensor
terminated: torch.Tensor
truncated: torch.Tensor
class PGPolicy(BasePolicy):
...
def forward(
self,
batch: RolloutBatchProtocol,
state: Optional[Union[dict, Batch, np.ndarray]] = None,
**kwargs: Any,
) -> ActionBatchProtocol:
```
The IDE and mypy are now very helpful in finding errors and in
auto-completion, whereas before the tools couldn't assist in that at
all.
## Major Change 2 - remove duplication in trainer package
**TL;DR:** There was a lot of duplication between `BaseTrainer` and its
subclasses. Even worse, it was almost-duplication. There was also
interface fragmentation through things like `onpolicy_trainer`. Now this
duplication is gone and all downstream code was adjusted.
**In more detail:** Since this change affects a lot of code, I would
like to explain why I thought it to be necessary.
1. The subclasses of `BaseTrainer` just duplicated docstrings and
constructors. What's worse, they changed the order of args there, even
turning some kwargs of BaseTrainer into args. They also had the arg
`learning_type` which was passed as kwarg to the base class and was
unused there. This made things difficult to maintain, and in fact some
errors were already present in the duplicated docstrings.
2. The "functions" a la `onpolicy_trainer`, which just called the
`OnpolicyTrainer.run`, not only introduced interface fragmentation but
also completely obfuscated the docstring and interfaces. They themselves
had no dosctring and the interface was just `*args, **kwargs`, which
makes it impossible to understand what they do and which things can be
passed without reading their implementation, then reading the docstring
of the associated class, etc. Needless to say, mypy and IDEs provide no
support with such functions. Nevertheless, they were used everywhere in
the code-base. I didn't find the sacrifices in clarity and complexity
justified just for the sake of not having to write `.run()` after
instantiating a trainer.
3. The trainers are all very similar to each other. As for my
application I needed a new trainer, I wanted to understand their
structure. The similarity, however, was hard to discover since they were
all in separate modules and there was so much duplication. I kept
staring at the constructors for a while until I figured out that
essentially no changes to the superclass were introduced. Now they are
all in the same module and the similarities/differences between them are
much easier to grasp (in my opinion)
4. Because of (1), I had to manually change and check a lot of code,
which was very tedious and boring. This kind of work won't be necessary
in the future, since now IDEs can be used for changing signatures,
renaming args and kwargs, changing class names and so on.
I have some more reasons, but maybe the above ones are convincing
enough.
## Minor changes: improved input validation and types
I added input validation for things like `state` and `action_scaling`
(which only makes sense for continuous envs). After adding this, some
tests failed to pass this validation. There I added
`action_scaling=isinstance(env.action_space, Box)`, after which tests
were green. I don't know why the tests were green before, since action
scaling doesn't make sense for discrete actions. I guess some aspect was
not tested and didn't crash.
I also added Literal in some places, in particular for
`action_bound_method`. Now it is no longer allowed to pass an empty
string, instead one should pass `None`. Also here there is input
validation with clear error messages.
@Trinkle23897 The functional tests are green. I didn't want to fix the
formatting, since it will change in the next PR that will solve #914
anyway. I also found a whole bunch of code in `docs/_static`, which I
just deleted (shouldn't it be copied from the sources during docs build
instead of committed?). I also haven't adjusted the documentation yet,
which atm still mentions the trainers of the type
`onpolicy_trainer(...)` instead of `OnpolicyTrainer(...).run()`
## Breaking Changes
The adjustments to the trainer package introduce breaking changes as
duplicated interfaces are deleted. However, it should be very easy for
users to adjust to them
---------
Co-authored-by: Michael Panchenko <m.panchenko@appliedai.de>
- [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!
- A DummyTqdm class added to utils: it replicates the interface used by trainers, but does not show the progress bar;
- Added a show_progress argument to the base trainer: when show_progress == True, dummy_tqdm is used in place of tqdm.
The new proposed feature is to have trainers as generators.
The usage pattern is:
```python
trainer = OnPolicyTrainer(...)
for epoch, epoch_stat, info in trainer:
print(f"Epoch: {epoch}")
print(epoch_stat)
print(info)
do_something_with_policy()
query_something_about_policy()
make_a_plot_with(epoch_stat)
display(info)
```
- epoch int: the epoch number
- epoch_stat dict: a large collection of metrics of the current epoch, including stat
- info dict: the usual dict out of the non-generator version of the trainer
You can even iterate on several different trainers at the same time:
```python
trainer1 = OnPolicyTrainer(...)
trainer2 = OnPolicyTrainer(...)
for result1, result2, ... in zip(trainer1, trainer2, ...):
compare_results(result1, result2, ...)
```
Co-authored-by: Jiayi Weng <trinkle23897@gmail.com>
- Fixes an inconsistency in the implementation of Discrete CRR. Now it uses `Critic` class for its critic, following conventions in other actor-critic policies;
- Updates several offline policies to use `ActorCritic` class for its optimizer to eliminate randomness caused by parameter sharing between actor and critic;
- Add `writer.flush()` in TensorboardLogger to ensure real-time result;
- Enable `test_collector=None` in 3 trainers to turn off testing during training;
- Updates the Atari offline results in README.md;
- Moves Atari offline RL examples to `examples/offline`; tests to `test/offline` per review comments.
- 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
This PR focus on refactor of logging method to solve bug of nan reward and log interval. After these two pr, hopefully fundamental change of tianshou/data is finished. We then can concentrate on building benchmarks of tianshou finally.
Things changed:
1. trainer now accepts logger (BasicLogger or LazyLogger) instead of writer;
2. remove utils.SummaryWriter;
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 PR separates the `global_step` into `env_step` and `gradient_step`. In the future, the data from the collecting state will be stored under `env_step`, and the data from the updating state will be stored under `gradient_step`.
Others:
- add `rew_std` and `best_result` into the monitor
- fix network unbounded in `test/continuous/test_sac_with_il.py` and `examples/box2d/bipedal_hardcore_sac.py`
- change the dependency of ray to 1.0.0 since ray-project/ray#10134 has been resolved
Add an indicator(i.e. `self.learning`) of learning will be convenient for distinguishing state of policy.
Meanwhile, the state of `self.training` will be undisputed in the training stage.
Related issue: #211
Others:
- fix a bug in DDQN: target_q could not be sampled from np.random.rand
- fix a bug in DQN atari net: it should add a ReLU before the last layer
- fix a bug in collector timing
Co-authored-by: n+e <463003665@qq.com>
Cherry-pick from #200
- update the function signature
- format code-style
- move _compile into separate functions
- fix a bug in to_torch and to_numpy (Batch)
- remove None in action_range
In short, the code-format only contains function-signature style and `'` -> `"`. (pick up from [black](https://github.com/psf/black))
- replace DiagGuassian with Independent(Normal) (pytorch has already supported this)
- detach alpha from autograd
- add value/alpha to result (more informational)
- revert #204 to fix#211
Co-authored-by: Trinkle23897 <463003665@qq.com>
