For Atari observation spaces, RLlib defaults to using the `DeepMind preprocessors <https://github.com/ray-project/ray/blob/master/rllib/env/atari_wrappers.py>`__
(``preprocessor_pref=deepmind``). However, if the Trainer's config key ``preprocessor_pref`` is set to "rllib",
the following mappings apply for Atari-type observation spaces:
- Images of shape ``(210, 160, 3)`` are downscaled to ``dim x dim``, where
``dim`` is a model config key (see default Model config below). Also, you can set
``grayscale=True`` for reducing the color channel to 1, or ``zero_mean=True`` for
producing -1.0 to 1.0 values (instead of 0.0 to 1.0 values by default).
The dict above (or an overriding sub-set) is handed to the Trainer via the ``model`` key within
the main config dict like so:
..code-block:: python
algo_config = {
# All model-related settings go into this sub-dict.
"model": {
# By default, the MODEL_DEFAULTS dict above will be used.
# Change individual keys in that dict by overriding them, e.g.
"fcnet_hiddens": [512, 512, 512],
"fcnet_activation": "relu",
},
# ... other Trainer config keys, e.g. "lr" ...
"lr": 0.00001,
}
Built-in Models
~~~~~~~~~~~~~~~
After preprocessing (if applicable) the raw environment outputs, the processed observations are fed through the policy's model.
In case, no custom model is specified (see further below on how to customize models), RLlib will pick a default model
based on simple heuristics:
- A vision network (`TF <https://github.com/ray-project/ray/blob/master/rllib/models/tf/visionnet.py>`__ or `Torch <https://github.com/ray-project/ray/blob/master/rllib/models/torch/visionnet.py>`__)
for observations that have a shape of length larger than 2, for example, ``(84 x 84 x 3)``.
- A fully connected network (`TF <https://github.com/ray-project/ray/blob/master/rllib/models/tf/fcnet.py>`__ or `Torch <https://github.com/ray-project/ray/blob/master/rllib/models/torch/fcnet.py>`__)
for everything else.
These default model types can further be configured via the ``model`` config key inside your Trainer config (as discussed above).
Available settings are `listed above <#default-model-config-settings>`__ and also documented in the `model catalog file <https://github.com/ray-project/ray/blob/master/rllib/models/catalog.py>`__.
Note that for the vision network case, you'll probably have to configure ``conv_filters``, if your environment observations
have custom sizes. For example, ``"model": {"dim": 42, "conv_filters": [[16, [4, 4], 2], [32, [4, 4], 2], [512, [11, 11], 1]]}`` for 42x42 observations.
Thereby, always make sure that the last Conv2D output has an output shape of ``[B, 1, 1, X]`` (``[B, X, 1, 1]`` for PyTorch), where B=batch and
X=last Conv2D layer's number of filters, so that RLlib can flatten it. An informative error will be thrown if this is not the case.
.._auto_lstm_and_attention:
Built-in auto-LSTM, and auto-Attention Wrappers
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In addition, if you set ``"use_lstm": True`` or ``"use_attention": True`` in your model config,
your model's output will be further processed by an LSTM cell
(`TF <https://github.com/ray-project/ray/blob/master/rllib/models/tf/recurrent_net.py>`__ or `Torch <https://github.com/ray-project/ray/blob/master/rllib/models/torch/recurrent_net.py>`__),
or an attention (`GTrXL <https://arxiv.org/abs/1910.06764>`__) network
(`TF <https://github.com/ray-project/ray/blob/master/rllib/models/tf/attention_net.py>`__ or
It is not possible to use both auto-wrappers (lstm and attention) at the same time. Doing so will create an error.
Customizing Preprocessors and Models
------------------------------------
Custom Preprocessors and Environment Filters
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
..warning::
Custom preprocessors are deprecated, since they sometimes conflict with the built-in preprocessors for handling complex observation spaces.
Please use `wrapper classes <https://github.com/openai/gym/tree/master/gym/wrappers>`__ around your environment instead of preprocessors.
Note that the built-in **default** Preprocessors described above will still be used and won't be deprecated.
Instead of using the deprecated custom Preprocessors, you should use ``gym.Wrappers`` to preprocess your environment's output (observations and rewards),
but also your Model's computed actions before sending them back to the environment.
For example, for manipulating your env's observations or rewards, do:
If you would like to provide your own model logic (instead of using RLlib's built-in defaults), you
can sub-class either ``TFModelV2`` (for TensorFlow) or ``TorchModelV2`` (for PyTorch) and then
register and specify your sub-class in the config as follows:
Custom TensorFlow Models
````````````````````````
Custom TensorFlow models should subclass `TFModelV2 <https://github.com/ray-project/ray/blob/master/rllib/models/tf/tf_modelv2.py>`__ and implement the ``__init__()`` and ``forward()`` methods.
``forward()`` takes a dict of tensor inputs (mapping str to Tensor types), whose keys and values depend on
the `view requirements <rllib-sample-collection.html>`__ of the model.
Normally, this input dict contains only the current observation ``obs`` and an ``is_training`` boolean flag, as well as an optional list of RNN states.
``forward()`` should return the model output (of size ``self.num_outputs``) and - if applicable - a new list of internal
states (in case of RNNs or attention nets). You can also override extra methods of the model such as ``value_function`` to implement
a custom value branch.
Additional supervised/self-supervised losses can be added via the ``TFModelV2.custom_loss`` method:
See the `keras model example <https://github.com/ray-project/ray/blob/master/rllib/examples/custom_keras_model.py>`__ for a full example of a TF custom model.
More examples and explanations on how to implement custom Tuple/Dict processing models
(also check out `this test case here <https://github.com/ray-project/ray/blob/master/rllib/tests/test_nested_observation_spaces.py>`__),
custom RNNs, custom model APIs (on top of default models) follow further below.
Custom PyTorch Models
`````````````````````
Similarly, you can create and register custom PyTorch models by subclassing
`TorchModelV2 <https://github.com/ray-project/ray/blob/master/rllib/models/torch/torch_modelv2.py>`__ and implement the ``__init__()`` and ``forward()`` methods.
``forward()`` takes a dict of tensor inputs (mapping str to PyTorch tensor types), whose keys and values depend on
the `view requirements <rllib-sample-collection.html>`__ of the model.
Usually, the dict contains only the current observation ``obs`` and an ``is_training`` boolean flag, as well as an optional list of RNN states.
``forward()`` should return the model output (of size ``self.num_outputs``) and - if applicable - a new list of internal
states (in case of RNNs or attention nets). You can also override extra methods of the model such as ``value_function`` to implement
a custom value branch.
Additional supervised/self-supervised losses can be added via the ``TorchModelV2.custom_loss`` method:
See these examples of `fully connected <https://github.com/ray-project/ray/blob/master/rllib/models/torch/fcnet.py>`__, `convolutional <https://github.com/ray-project/ray/blob/master/rllib/models/torch/visionnet.py>`__, and `recurrent <https://github.com/ray-project/ray/blob/master/rllib/models/torch/recurrent_net.py>`__ torch models.
See the `torch model examples <https://github.com/ray-project/ray/blob/master/rllib/examples/models/>`__ for various examples on how to build a custom
PyTorch model (including recurrent ones).
More examples and explanations on how to implement custom Tuple/Dict processing models (also check out `this test case here <https://github.com/ray-project/ray/blob/master/rllib/tests/test_nested_observation_spaces.py>`__),
custom RNNs, custom model APIs (on top of default models) follow further below.
Instead of using the ``use_lstm: True`` option, it may be preferable to use a custom recurrent model.
This provides more control over postprocessing the LSTM's output and can also allow the use of multiple LSTM cells to process different portions of the input.
For an RNN model it is recommended to subclass ``RecurrentNetwork`` (either the `TF <https://github.com/ray-project/ray/blob/master/rllib/models/tf/recurrent_net.py>`__
or `PyTorch <https://github.com/ray-project/ray/blob/master/rllib/models/torch/recurrent_net.py>`__ versions) and then implement ``__init__()``,
Note that the ``inputs`` arg entering ``forward_rnn`` is already a time-ranked single tensor (not an ``input_dict``!) with shape ``(B x T x ...)``.
If you further want to customize and need more direct access to the complete (non time-ranked) ``input_dict``, you can also override
your Model's ``forward`` method directly (as you would do with a non-RNN ModelV2). In that case, though, you are responsible for changing your inputs
and add the time rank to the incoming data (usually you just have to reshape).
You can check out the `rnn_model.py <https://github.com/ray-project/ray/blob/master/rllib/examples/models/rnn_model.py>`__ models as examples to implement
Similar to the RNN case described above, you could also implement your own attention-based networks, instead of using the
``use_attention: True`` flag in your model config.
Check out RLlib's `GTrXL (Attention Net) <https://arxiv.org/abs/1910.06764>`__ implementations
(for `TF <https://github.com/ray-project/ray/blob/master/rllib/models/tf/attention_net.py>`__ and `PyTorch <https://github.com/ray-project/ray/blob/master/rllib/models/torch/attention_net.py>`__)
to get a better idea on how to write your own models of this type. These are the models we use
as wrappers when ``use_attention=True``.
You can run `this example script <https://github.com/ray-project/ray/blob/master/rllib/examples/attention_net.py>`__ to run these nets within some of our algorithms.
`There is also a test case <https://github.com/ray-project/ray/blob/master/rllib/tests/test_attention_net_learning.py>`__, which confirms their learning capabilities in PPO and IMPALA.
In case RLlib does not properly detect the update ops for your custom model, you can override the ``update_ops()`` method to return the list of ops to run for updates.
A PyTorch version of the above model is also `given in the same file <https://github.com/ray-project/ray/blob/master/rllib/examples/models/trajectory_view_utilizing_models.py>`__.
Similar to custom models and preprocessors, you can also specify a custom action distribution class as follows. The action dist class is passed a reference to the ``model``, which you can use to access ``model.model_config`` or other attributes of the model. This is commonly used to implement `autoregressive action outputs <#autoregressive-action-distributions>`__.
..code-block:: python
import ray
import ray.rllib.agents.ppo as ppo
from ray.rllib.models import ModelCatalog
from ray.rllib.models.preprocessors import Preprocessor
You can mix supervised losses into any RLlib algorithm through custom models. For example, you can add an imitation learning loss on expert experiences, or a self-supervised autoencoder loss within the model. These losses can be defined over either policy evaluation inputs, or data read from `offline storage <rllib-offline.html#input-pipeline-for-supervised-losses>`__.
**TensorFlow**: To add a supervised loss to a custom TF model, you need to override the ``custom_loss()`` method. This method takes in the existing policy loss for the algorithm, which you can add your own supervised loss to before returning. For debugging, you can also return a dictionary of scalar tensors in the ``metrics()`` method. Here is a `runnable example <https://github.com/ray-project/ray/blob/master/rllib/examples/custom_loss.py>`__ of adding an imitation loss to CartPole training that is defined over a `offline dataset <rllib-offline.html#input-pipeline-for-supervised-losses>`__.
**PyTorch**: There is no explicit API for adding losses to custom torch models. However, you can modify the loss in the policy definition directly. Like for TF models, offline datasets can be incorporated by creating an input reader and calling ``reader.next()`` in the loss forward pass.
RLlib supports complex and variable-length observation spaces, including ``gym.spaces.Tuple``, ``gym.spaces.Dict``, and ``rllib.utils.spaces.Repeated``. The handling of these spaces is transparent to the user. RLlib internally will insert preprocessors to insert padding for repeated elements, flatten complex observations into a fixed-size vector during transit, and unpack the vector into the structured tensor before sending it to the model. The flattened observation is available to the model as ``input_dict["obs_flat"]``, and the unpacked observation as ``input_dict["obs"]``.
To enable batching of struct observations, RLlib unpacks them in a `StructTensor-like format <https://github.com/tensorflow/community/blob/master/rfcs/20190910-struct-tensor.md>`__. In summary, repeated fields are "pushed down" and become the outer dimensions of tensor batches, as illustrated in this figure from the StructTensor RFC.
..image:: struct-tensor.png
For further information about complex observation spaces, see:
* A custom environment and model that uses `repeated struct fields <https://github.com/ray-project/ray/blob/master/rllib/examples/complex_struct_space.py>`__.
* The pydoc of the `Repeated space <https://github.com/ray-project/ray/blob/master/rllib/utils/spaces/repeated.py>`__.
* The pydoc of the batched `repeated values tensor <https://github.com/ray-project/ray/blob/master/rllib/models/repeated_values.py>`__.
Custom models can be used to work with environments where (1) the set of valid actions `varies per step <https://neuro.cs.ut.ee/the-use-of-embeddings-in-openai-five>`__, and/or (2) the number of valid actions is `very large <https://arxiv.org/abs/1811.00260>`__. The general idea is that the meaning of actions can be completely conditioned on the observation, i.e., the ``a`` in ``Q(s, a)`` becomes just a token in ``[0, MAX_AVAIL_ACTIONS)`` that only has meaning in the context of ``s``. This works with algorithms in the `DQN and policy-gradient families <rllib-env.html>`__ and can be implemented as follows:
1. The environment should return a mask and/or list of valid action embeddings as part of the observation for each step. To enable batching, the number of actions can be allowed to vary from 1 to some max number:
2. A custom model can be defined that can interpret the ``action_mask`` and ``avail_actions`` portions of the observation. Here the model computes the action logits via the dot product of some network output and each action embedding. Invalid actions can be masked out of the softmax by scaling the probability to zero:
Depending on your use case it may make sense to use just the masking, just action embeddings, or both. For a runnable example of this in code, check out `parametric_actions_cartpole.py <https://github.com/ray-project/ray/blob/master/rllib/examples/parametric_actions_cartpole.py>`__. Note that since masking introduces ``tf.float32.min`` values into the model output, this technique might not work with all algorithm options. For example, algorithms might crash if they incorrectly process the ``tf.float32.min`` values. The cartpole example has working configurations for DQN (must set ``hiddens=[]``), PPO (must disable running mean and set ``model.vf_share_layers=True``), and several other algorithms. Not all algorithms support parametric actions; see the `algorithm overview <rllib-algorithms.html#available-algorithms-overview>`__.
In an action space with multiple components (e.g., ``Tuple(a1, a2)``), you might want ``a2`` to be conditioned on the sampled value of ``a1``, i.e., ``a2_sampled ~ P(a2 | a1_sampled, obs)``. Normally, ``a1`` and ``a2`` would be sampled independently, reducing the expressivity of the policy.
To do this, you need both a custom model that implements the autoregressive pattern, and a custom action distribution class that leverages that model. The `autoregressive_action_dist.py <https://github.com/ray-project/ray/blob/master/rllib/examples/autoregressive_action_dist.py>`__ example shows how this can be implemented for a simple binary action space. For a more complex space, a more efficient architecture such as a `MADE <https://arxiv.org/abs/1502.03509>`__ is recommended. Note that sampling a `N-part` action requires `N` forward passes through the model, however computing the log probability of an action can be done in one pass:
Not all algorithms support autoregressive action distributions; see the `algorithm overview table <rllib-algorithms.html#available-algorithms-overview>`__ for more information.
