Angorapy

Modeling
- Automized input construction API for model building. Use ´make_input_layers(...)` from `angorapy.utilities.model_utils` to quickly construct input layers for the different modalities of your model. This approach is now also used by the builtin models.

Misc
- Updated requirements for improved simplicity during installation.
- Updated Readme
- Added unit tests for builtin models

Modeling
* __Simple__, __Wider__, and __Deeper__ models now support all input modalities.
* Experimental LiF function.
* Added a validator for model builders that is called when Agents are instantiated.

Environments
* Rolling out the policy now also records auxiliary metrics if the environment supports them in the returned `info` dictionary.

Analysis
* Several new experimental plotting functions.

Environments
* Fixed rendering for reach environments (targets and fingertips are now properly shown)
* Added Hanoi task

Utilities
* Changed plotting utils
* Plotting functions are not longer returning HTML/JS plots but more flexible `bokeh` objects, to be used for both web browser rendering and Jupyter notebooks, too.

Dependencies
* Bumped TensorFlow from 2.9.1 to 2.10.0

Misc
* Better pertaining on object pose

New Features
- Added new policy distribution `ReparameterizedBetaPolicyDistribution`. It's a reparameterized version of the Beta distribution (who could have guessed) that predicts the mean and spread of the PDF instead of alpha and beta. This allows the spread to be controlled without input dependency, leading to more stability.
- Improved customizability of the `Gatherer` class. `Gatherer` now has a `.postprocess()` method called during data collection to postprocess data collected in the data buffer. The default buffer only normalizes advantages via this method, but custom gatherers can apply more postprocessing and even add or filter data to/from the buffer.

Monitoring
- New group view in the web monitor. Experiments now can be assigned an optional group name. The new view can investigate the mean reward progression of grouped experiments. More functionality on this will follow in future updates.
- Better filtering of experiments in the web monitor.
- The hyperparameter view now also shows Gatherer information
- Improved robustness against corrupted JSON files.

Other Changes
- Upgraded from TensorFlow `2.4.2` to `2.9.1`. Should at this point still be backward-compatible though.
- Throughout optimization, several assertions have been added to simplify debugging when facing NaN/Inf values.
- Added new unit tests.

Changelog

New Features/Improvements
- Added Docker support. The Dockerfile can be found at `docker/`. The Dockerfile currently expects you to build with this repository as context.
- Updated dependency on `gym==0.24.0`. This includes a rewrite of the dexterity environments using `mujoco` (deepmind's own bindings). These environments now do not anymore depend on the gym implementation but are native in `angorapy`.
- With the migration to `mujoco`, MuJoCo itself needs not to be installed manually anymore. Checks for installed MuJoCo have thus been changed and now depend on whether `mujoco` is installed.
- Added new variants of Manipulate (just registrations)
- Added new examples
- Added more information to the README
- Improved pull script
- The monitor now also stores hyperparameters set in the drill call (number of epochs and batch size)

Bugfixes
- Several bug fixes w.r.t. MuJoCo environments.
- Fixed automatic MuJoCo detection. AngoraPy now checks for either an existing `.mujoco` directory in your home directory or a set environment variable MUJOCO_PY_MUJOCO_PATH. If either exist, MuJoCo will be attempted to load.

Breaking changes
- Moved package data to `angorapy`, so all imports now are `import angorapy` instead of `import dexterity`. We recommend doing `import angorapy as ang`. Additional shortcuts to major function (e.g. `make_env`) have been added and will be in future releases.