garage.tf.models.model module

Base model classes.

class BaseModel

Bases: abc.ABC

Interface-only abstract class for models.

A Model contains the structure/configuration of a set of computation graphs, or can be understood as a set of networks. Using a model requires calling build() with given input placeholder, which can be either tf.compat.v1.placeholder, or the output from another model. This makes composition of complex models with simple models much easier.

Examples

model = SimpleModel(output_dim=2) # To use a model, first create a placeholder. # In the case of TensorFlow, we create a tf.compat.v1.placeholder. input_ph = tf.compat.v1.placeholder(tf.float32, shape=(None, 2))

# Building the model output = model.build(input_ph)

# We can also pass the output of a model to another model. # Here we pass the output from the above SimpleModel object. model_2 = ComplexModel(output_dim=2) output_2 = model_2.build(output)

build(*inputs, name=None)

Output of model with the given input placeholder(s).

This function is implemented by subclasses to create their computation graphs, which will be managed by Model. Generally, subclasses should implement build() directly.

Parameters:

• inputs (object) – Input(s) for the model.
• name (str) – Name of the model.

Returns:

Output(s) of the model.

Return type:

list[tf.Tensor]

name

Name for this Model.

parameters

Parameters of the Model.

The output of a model is determined by its parameter. It could be the weights of a neural network model or parameters of a loss function model.

Returns:Parameters. Return type:list[tf.Tensor]

class Model(name)

Bases: garage.tf.models.model.BaseModel

Model class for TensorFlow.

A TfModel only contains the structure/configuration of the underlying computation graphs. Connectivity information are all in Network class. A TfModel contains zero or more Network.

When a Network is created, it reuses the parameter from the model. If a Network is built without given a name, the name “default” will be used.

* Do not call tf.global_variable_initializers() after building a model as it will reassign random weights to the model. The parameters inside a model will be initialized when calling build(). *

Pickling is handled automatcailly. The target weights should be assigned to self._default_parameters before pickling, so that the newly created model can check if target weights exist or not. When unpickled, the unserialized model will load the weights from self._default_parameters.

The design is illustrated as the following:

input_1 input_2

|

============== Model (TfModel)=================== | | | | | | Parameters | | | ============= / ============ | | | default | / | Network2 | | | | (Network) |/ |(Network) | | | ============= ============ | | | | | =================================================

|

|

(outputs from ‘default’ networks) |

outputs from [‘Network2’] network

Examples are also available in tests/garage/tf/models/test_model.

Parameters:name (str) – Name of the model. It will also become the variable scope of the model. Every model should have a unique name.

build(*inputs, name=None)

Build a Network with the given input(s).

* Do not call tf.global_variable_initializers() after building a model as it will reassign random weights to the model. The parameters inside a model will be initialized when calling build(). *

It uses the same, fixed variable scope for all Networks, to ensure parameter sharing. Different Networks must have an unique name.

Parameters:

• inputs (list[tf.Tensor]) – Tensor input(s), recommended to be positional arguments, for example, def build(self, state_input, action_input, name=None).
• name (str) – Name of the model, which is also the name scope of the model.

Raises:

ValueError – When a Network with the same name is already built.

Returns:

Output tensors of the model with the given

inputs.

Return type:

list[tf.Tensor]

input

Default input of the model.

When the model is built the first time, by default it creates the ‘default’ network. This property creates a reference to the input of the network.

Returns:Default input of the model. Return type:tf.Tensor

inputs

Default inputs of the model.

When the model is built the first time, by default it creates the ‘default’ network. This property creates a reference to the inputs of the network.

Returns:Default inputs of the model. Return type:list[tf.Tensor]

name

Name (str) of the model.

This is also the variable scope of the model.

Returns:Name of the model. Return type:str

network_input_spec()

Network input spec.

Returns:List of key(str) for the network inputs. Return type:list[str]

network_output_spec()

Network output spec.

Returns:List of key(str) for the network outputs. Return type:list[str]

output

Default output of the model.

When the model is built the first time, by default it creates the ‘default’ network. This property creates a reference to the output of the network.

Returns:Default output of the model. Return type:tf.Tensor

outputs

Default outputs of the model.

When the model is built the first time, by default it creates the ‘default’ network. This property creates a reference to the outputs of the network.

Returns:Default outputs of the model. Return type:list[tf.Tensor]

parameters

Parameters of the model.

Returns:Parameters Return type:np.ndarray

class Network

Bases: object

Network class For TensorFlow.

A Network contains connectivity information by inputs/outputs. When a Network is built, it appears as a subgraph in the computation graphs, scoped by the Network name. All Networks built with the same model share the same parameters, i.e same inputs yield to same outputs.

input

Tensor input of the Network.

Returns:Input. Return type:tf.Tensor

inputs

Tensor inputs of the Network.

Returns:Inputs. Return type:list[tf.Tensor]

output

Tensor output of the Network.

Returns:Output. Return type:tf.Tensor

outputs

Tensor outputs of the Network.

Returns:Outputs. Return type:list[tf.Tensor]