def __init__(self, env_spec, conv_filters, conv_filter_sizes, conv_strides, conv_pad, name='CategoricalCNNPolicy', hidden_sizes=(32, 32), hidden_nonlinearity=tf.nn.relu, hidden_w_init=tf.initializers.glorot_uniform(), hidden_b_init=tf.zeros_initializer(), output_nonlinearity=tf.nn.softmax, output_w_init=tf.initializers.glorot_uniform(), output_b_init=tf.zeros_initializer(), layer_normalization=False): if not isinstance(env_spec.action_space, akro.Discrete): raise ValueError( 'CategoricalCNNPolicy only works with akro.Discrete action ' 'space.') if not isinstance(env_spec.observation_space, akro.Box) or \ not len(env_spec.observation_space.shape) in (2, 3): raise ValueError( '{} can only process 2D, 3D akro.Image or' ' akro.Box observations, but received an env_spec with ' 'observation_space of type {} and shape {}'.format( type(self).__name__, type(env_spec.observation_space).__name__, env_spec.observation_space.shape)) super().__init__(name, env_spec) self.obs_dim = env_spec.observation_space.shape self.action_dim = env_spec.action_space.n self.model = Sequential( CNNModel(filter_dims=conv_filter_sizes, num_filters=conv_filters, strides=conv_strides, padding=conv_pad, hidden_nonlinearity=hidden_nonlinearity, name='CNNModel'), MLPModel(output_dim=self.action_dim, hidden_sizes=hidden_sizes, hidden_nonlinearity=hidden_nonlinearity, hidden_w_init=hidden_w_init, hidden_b_init=hidden_b_init, output_nonlinearity=output_nonlinearity, output_w_init=output_w_init, output_b_init=output_b_init, layer_normalization=layer_normalization, name='MLPModel')) self._initialize()
def __init__(self, env_spec, conv_filters, conv_filter_sizes, conv_strides, conv_pad, name='CategoricalConvPolicy', hidden_sizes=[], hidden_nonlinearity=tf.nn.relu, hidden_w_init=tf.glorot_uniform_initializer(), hidden_b_init=tf.zeros_initializer(), output_nonlinearity=tf.nn.softmax, output_w_init=tf.glorot_uniform_initializer(), output_b_init=tf.zeros_initializer(), layer_normalization=False): assert isinstance(env_spec.action_space, akro.Discrete), ( 'CategoricalConvPolicy only works with akro.Discrete action ' 'space.') super().__init__(name, env_spec) self.obs_dim = env_spec.observation_space.shape self.action_dim = env_spec.action_space.n self.model = Sequential( CNNModel( filter_dims=conv_filter_sizes, num_filters=conv_filters, strides=conv_strides, padding=conv_pad, hidden_nonlinearity=hidden_nonlinearity, name='CNNModel'), MLPModel( output_dim=self.action_dim, hidden_sizes=hidden_sizes, hidden_nonlinearity=hidden_nonlinearity, hidden_w_init=hidden_w_init, hidden_b_init=hidden_b_init, output_nonlinearity=output_nonlinearity, output_w_init=output_w_init, output_b_init=output_b_init, layer_normalization=layer_normalization, name='MLPModel')) self._initialize()
class CategoricalConvPolicyWithModel(StochasticPolicy2): """ CategoricalConvPolicy with model. A policy that contains a CNN and a MLP to make prediction based on a categorical distribution. It only works with akro.tf.Discrete action space. Args: env_spec (garage.envs.env_spec.EnvSpec): Environment specification. conv_filter_sizes(tuple[int]): Dimension of the filters. For example, (3, 5) means there are two convolutional layers. The filter for first layer is of dimension (3 x 3) and the second one is of dimension (5 x 5). conv_filters(tuple[int]): Number of filters. For example, (3, 32) means there are two convolutional layers. The filter for the first layer has 3 channels and the second one with 32 channels. conv_strides(tuple[int]): The stride of the sliding window. For example, (1, 2) means there are two convolutional layers. The stride of the filter for first layer is 1 and that of the second layer is 2. conv_pad (str): The type of padding algorithm to use, either 'SAME' or 'VALID'. name (str): Policy name, also the variable scope of the policy. hidden_sizes (list[int]): Output dimension of dense layer(s). For example, (32, 32) means the MLP of this policy consists of two hidden layers, each with 32 hidden units. hidden_nonlinearity (callable): Activation function for intermediate dense layer(s). It should return a tf.Tensor. Set it to None to maintain a linear activation. hidden_w_init (callable): Initializer function for the weight of intermediate dense layer(s). The function should return a tf.Tensor. hidden_b_init (callable): Initializer function for the bias of intermediate dense layer(s). The function should return a tf.Tensor. output_nonlinearity (callable): Activation function for output dense layer. It should return a tf.Tensor. Set it to None to maintain a linear activation. output_w_init (callable): Initializer function for the weight of output dense layer(s). The function should return a tf.Tensor. output_b_init (callable): Initializer function for the bias of output dense layer(s). The function should return a tf.Tensor. layer_normalization (bool): Bool for using layer normalization or not. """ def __init__(self, env_spec, conv_filters, conv_filter_sizes, conv_strides, conv_pad, name='CategoricalConvPolicy', hidden_sizes=[], hidden_nonlinearity=tf.nn.relu, hidden_w_init=tf.glorot_uniform_initializer(), hidden_b_init=tf.zeros_initializer(), output_nonlinearity=tf.nn.softmax, output_w_init=tf.glorot_uniform_initializer(), output_b_init=tf.zeros_initializer(), layer_normalization=False): assert isinstance(env_spec.action_space, Discrete), ( 'CategoricalConvPolicy only works with akro.tf.Discrete' 'action space.') super().__init__(name, env_spec) self.obs_dim = env_spec.observation_space.shape self.action_dim = env_spec.action_space.n self.model = Sequential( CNNModel(filter_dims=conv_filter_sizes, num_filters=conv_filters, strides=conv_strides, padding=conv_pad, hidden_nonlinearity=hidden_nonlinearity, name='CNNModel'), MLPModel(output_dim=self.action_dim, hidden_sizes=hidden_sizes, hidden_nonlinearity=hidden_nonlinearity, hidden_w_init=hidden_w_init, hidden_b_init=hidden_b_init, output_nonlinearity=output_nonlinearity, output_w_init=output_w_init, output_b_init=output_b_init, layer_normalization=layer_normalization, name='MLPModel')) self._initialize() def _initialize(self): state_input = tf.placeholder(tf.float32, shape=(None, ) + self.obs_dim) with tf.variable_scope(self._variable_scope): self.model.build(state_input) self._f_prob = tf.get_default_session().make_callable( self.model.outputs, feed_list=[self.model.input]) @property def vectorized(self): """Vectorized or not.""" return True @overrides def dist_info_sym(self, obs_var, state_info_vars=None, name=None): """Symbolic graph of the distribution.""" with tf.variable_scope(self._variable_scope): prob = self.model.build(obs_var, name=name) return dict(prob=prob) @overrides def dist_info(self, obs, state_infos=None): """Distribution info.""" prob = self._f_prob(obs) return dict(prob=prob) @overrides def get_action(self, observation): """Return a single action.""" # flat_obs = self.observation_space.flatten(observation) prob = self._f_prob([observation])[0] action = self.action_space.weighted_sample(prob) return action, dict(prob=prob) def get_actions(self, observations): """Return multiple actions.""" # flat_obs = self.observation_space.flatten_n(observations) probs = self._f_prob(observations) actions = list(map(self.action_space.weighted_sample, probs)) return actions, dict(prob=probs) @property def distribution(self): """Policy distribution.""" return Categorical(self.action_dim) def __getstate__(self): """Object.__getstate__.""" new_dict = self.__dict__.copy() del new_dict['_f_prob'] return new_dict def __setstate__(self, state): """Object.__setstate__.""" self.__dict__.update(state) self._initialize()
def __init__(self, env_spec, filters, strides, hidden_sizes=(256, ), name=None, padding='SAME', max_pooling=False, pool_strides=(2, 2), pool_shapes=(2, 2), cnn_hidden_nonlinearity=tf.nn.relu, hidden_nonlinearity=tf.nn.relu, hidden_w_init=tf.initializers.glorot_uniform(), hidden_b_init=tf.zeros_initializer(), output_nonlinearity=None, output_w_init=tf.initializers.glorot_uniform(), output_b_init=tf.zeros_initializer(), dueling=False, layer_normalization=False): if not isinstance(env_spec.observation_space, akro.Box) or \ not len(env_spec.observation_space.shape) in (2, 3): raise ValueError( '{} can only process 2D, 3D akro.Image or' ' akro.Box observations, but received an env_spec with ' 'observation_space of type {} and shape {}'.format( type(self).__name__, type(env_spec.observation_space).__name__, env_spec.observation_space.shape)) super().__init__(name) self._env_spec = env_spec self._action_dim = env_spec.action_space.n self._filters = filters self._strides = strides self._hidden_sizes = hidden_sizes self._padding = padding self._max_pooling = max_pooling self._pool_strides = pool_strides self._pool_shapes = pool_shapes self._cnn_hidden_nonlinearity = cnn_hidden_nonlinearity self._hidden_nonlinearity = hidden_nonlinearity self._hidden_w_init = hidden_w_init self._hidden_b_init = hidden_b_init self._output_nonlinearity = output_nonlinearity self._output_w_init = output_w_init self._output_b_init = output_b_init self._layer_normalization = layer_normalization self._dueling = dueling self.obs_dim = self._env_spec.observation_space.shape action_dim = self._env_spec.action_space.flat_dim if not max_pooling: cnn_model = CNNModel(filters=filters, strides=strides, padding=padding, hidden_nonlinearity=cnn_hidden_nonlinearity) else: cnn_model = CNNModelWithMaxPooling( filters=filters, strides=strides, padding=padding, pool_strides=pool_strides, pool_shapes=pool_shapes, hidden_nonlinearity=cnn_hidden_nonlinearity) if not dueling: output_model = MLPModel(output_dim=action_dim, hidden_sizes=hidden_sizes, hidden_nonlinearity=hidden_nonlinearity, hidden_w_init=hidden_w_init, hidden_b_init=hidden_b_init, output_nonlinearity=output_nonlinearity, output_w_init=output_w_init, output_b_init=output_b_init, layer_normalization=layer_normalization) else: output_model = MLPDuelingModel( output_dim=action_dim, hidden_sizes=hidden_sizes, hidden_nonlinearity=hidden_nonlinearity, hidden_w_init=hidden_w_init, hidden_b_init=hidden_b_init, output_nonlinearity=output_nonlinearity, output_w_init=output_w_init, output_b_init=output_b_init, layer_normalization=layer_normalization) self.model = Sequential(cnn_model, output_model) self._initialize()
class DiscreteCNNQFunction(QFunction): """Q function based on a CNN-MLP structure for discrete action space. This class implements a Q value network to predict Q based on the input state and action. It uses an CNN and a MLP to fit the function of Q(s, a). Args: env_spec (garage.envs.env_spec.EnvSpec): Environment specification. filters (Tuple[Tuple[int, Tuple[int, int]], ...]): Number and dimension of filters. For example, ((3, (3, 5)), (32, (3, 3))) means there are two convolutional layers. The filter for the first layer have 3 channels and its shape is (3 x 5), while the filter for the second layer have 32 channels and its shape is (3 x 3). strides (tuple[int]): The stride of the sliding window. For example, (1, 2) means there are two convolutional layers. The stride of the filter for first layer is 1 and that of the second layer is 2. hidden_sizes (list[int]): Output dimension of dense layer(s). For example, (32, 32) means the MLP of this q-function consists of two hidden layers, each with 32 hidden units. name (str): Variable scope of the cnn. padding (str): The type of padding algorithm to use, either 'SAME' or 'VALID'. max_pooling (bool): Boolean for using max pooling layer or not. pool_shapes (tuple[int]): Dimension of the pooling layer(s). For example, (2, 2) means that all the pooling layers have shape (2, 2). pool_strides (tuple[int]): The strides of the pooling layer(s). For example, (2, 2) means that all the pooling layers have strides (2, 2). cnn_hidden_nonlinearity (callable): Activation function for intermediate dense layer(s) in the CNN. It should return a tf.Tensor. Set it to None to maintain a linear activation. hidden_nonlinearity (callable): Activation function for intermediate dense layer(s) in the MLP. It should return a tf.Tensor. Set it to None to maintain a linear activation. hidden_w_init (callable): Initializer function for the weight of intermediate dense layer(s) in the MLP. The function should return a tf.Tensor. hidden_b_init (callable): Initializer function for the bias of intermediate dense layer(s) in the MLP. The function should return a tf.Tensor. output_nonlinearity (callable): Activation function for output dense layer in the MLP. It should return a tf.Tensor. Set it to None to maintain a linear activation. output_w_init (callable): Initializer function for the weight of output dense layer(s) in the MLP. The function should return a tf.Tensor. output_b_init (callable): Initializer function for the bias of output dense layer(s) in the MLP. The function should return a tf.Tensor. dueling (bool): Bool for using dueling network or not. layer_normalization (bool): Bool for using layer normalization or not. """ def __init__(self, env_spec, filters, strides, hidden_sizes=(256, ), name=None, padding='SAME', max_pooling=False, pool_strides=(2, 2), pool_shapes=(2, 2), cnn_hidden_nonlinearity=tf.nn.relu, hidden_nonlinearity=tf.nn.relu, hidden_w_init=tf.initializers.glorot_uniform(), hidden_b_init=tf.zeros_initializer(), output_nonlinearity=None, output_w_init=tf.initializers.glorot_uniform(), output_b_init=tf.zeros_initializer(), dueling=False, layer_normalization=False): if not isinstance(env_spec.observation_space, akro.Box) or \ not len(env_spec.observation_space.shape) in (2, 3): raise ValueError( '{} can only process 2D, 3D akro.Image or' ' akro.Box observations, but received an env_spec with ' 'observation_space of type {} and shape {}'.format( type(self).__name__, type(env_spec.observation_space).__name__, env_spec.observation_space.shape)) super().__init__(name) self._env_spec = env_spec self._action_dim = env_spec.action_space.n self._filters = filters self._strides = strides self._hidden_sizes = hidden_sizes self._padding = padding self._max_pooling = max_pooling self._pool_strides = pool_strides self._pool_shapes = pool_shapes self._cnn_hidden_nonlinearity = cnn_hidden_nonlinearity self._hidden_nonlinearity = hidden_nonlinearity self._hidden_w_init = hidden_w_init self._hidden_b_init = hidden_b_init self._output_nonlinearity = output_nonlinearity self._output_w_init = output_w_init self._output_b_init = output_b_init self._layer_normalization = layer_normalization self._dueling = dueling self.obs_dim = self._env_spec.observation_space.shape action_dim = self._env_spec.action_space.flat_dim if not max_pooling: cnn_model = CNNModel(filters=filters, strides=strides, padding=padding, hidden_nonlinearity=cnn_hidden_nonlinearity) else: cnn_model = CNNModelWithMaxPooling( filters=filters, strides=strides, padding=padding, pool_strides=pool_strides, pool_shapes=pool_shapes, hidden_nonlinearity=cnn_hidden_nonlinearity) if not dueling: output_model = MLPModel(output_dim=action_dim, hidden_sizes=hidden_sizes, hidden_nonlinearity=hidden_nonlinearity, hidden_w_init=hidden_w_init, hidden_b_init=hidden_b_init, output_nonlinearity=output_nonlinearity, output_w_init=output_w_init, output_b_init=output_b_init, layer_normalization=layer_normalization) else: output_model = MLPDuelingModel( output_dim=action_dim, hidden_sizes=hidden_sizes, hidden_nonlinearity=hidden_nonlinearity, hidden_w_init=hidden_w_init, hidden_b_init=hidden_b_init, output_nonlinearity=output_nonlinearity, output_w_init=output_w_init, output_b_init=output_b_init, layer_normalization=layer_normalization) self.model = Sequential(cnn_model, output_model) self._initialize() def _initialize(self): """Initialize QFunction.""" if isinstance(self._env_spec.observation_space, akro.Image): obs_ph = tf.compat.v1.placeholder(tf.uint8, (None, ) + self.obs_dim, name='obs') augmented_obs_ph = tf.cast(obs_ph, tf.float32) / 255.0 else: obs_ph = tf.compat.v1.placeholder(tf.float32, (None, ) + self.obs_dim, name='obs') augmented_obs_ph = obs_ph with tf.compat.v1.variable_scope(self.name) as vs: self._variable_scope = vs self.model.build(augmented_obs_ph) self._obs_input = obs_ph @property def q_vals(self): """Return the Q values, the output of the network. Return: list[tf.Tensor]: Q values. """ return self.model.networks['default'].outputs @property def input(self): """Get input. Return: tf.Tensor: QFunction Input. """ return self._obs_input # pylint: disable=arguments-differ def get_qval_sym(self, state_input, name): """Symbolic graph for q-network. Args: state_input (tf.Tensor): The state input tf.Tensor to the network. name (str): Network variable scope. Return: tf.Tensor: The tf.Tensor output of Discrete CNN QFunction. """ with tf.compat.v1.variable_scope(self._variable_scope): augmented_state_input = state_input if isinstance(self._env_spec.observation_space, akro.Image): augmented_state_input = tf.cast(state_input, tf.float32) / 255.0 return self.model.build(augmented_state_input, name=name) def clone(self, name): """Return a clone of the Q-function. It only copies the configuration of the Q-function, not the parameters. Args: name(str) : Name of the newly created q-function. Returns: garage.tf.q_functions.DiscreteCNNQFunction: Clone of this object """ return self.__class__(name=name, env_spec=self._env_spec, filters=self._filters, strides=self._strides, hidden_sizes=self._hidden_sizes, padding=self._padding, max_pooling=self._max_pooling, pool_shapes=self._pool_shapes, pool_strides=self._pool_strides, hidden_nonlinearity=self._hidden_nonlinearity, hidden_w_init=self._hidden_w_init, hidden_b_init=self._hidden_b_init, output_nonlinearity=self._output_nonlinearity, output_w_init=self._output_w_init, output_b_init=self._output_b_init, dueling=self._dueling, layer_normalization=self._layer_normalization) def __setstate__(self, state): """Object.__setstate__. Args: state (dict): Unpickled state. """ self.__dict__.update(state) self._initialize() def __getstate__(self): """Object.__getstate__. Returns: dict: The state. """ new_dict = self.__dict__.copy() del new_dict['_obs_input'] return new_dict
def __init__(self, env_spec, filter_dims, num_filters, strides, hidden_sizes=[256], name=None, padding='SAME', max_pooling=False, pool_strides=(2, 2), pool_shapes=(2, 2), cnn_hidden_nonlinearity=tf.nn.relu, hidden_nonlinearity=tf.nn.relu, hidden_w_init=tf.glorot_uniform_initializer(), hidden_b_init=tf.zeros_initializer(), output_nonlinearity=None, output_w_init=tf.glorot_uniform_initializer(), output_b_init=tf.zeros_initializer(), dueling=False, layer_normalization=False): super().__init__(name) self._env_spec = env_spec self._action_dim = env_spec.action_space.n self._filter_dims = filter_dims self._num_filters = num_filters self._strides = strides self._hidden_sizes = hidden_sizes self._padding = padding self._max_pooling = max_pooling self._pool_strides = pool_strides self._pool_shapes = pool_shapes self._cnn_hidden_nonlinearity = cnn_hidden_nonlinearity self._hidden_nonlinearity = hidden_nonlinearity self._hidden_w_init = hidden_w_init self._hidden_b_init = hidden_b_init self._output_nonlinearity = output_nonlinearity self._output_w_init = output_w_init self._output_b_init = output_b_init self._layer_normalization = layer_normalization self._dueling = dueling self.obs_dim = self._env_spec.observation_space.shape action_dim = self._env_spec.action_space.flat_dim if not max_pooling: cnn_model = CNNModel(filter_dims=filter_dims, num_filters=num_filters, strides=strides, padding=padding, hidden_nonlinearity=cnn_hidden_nonlinearity) else: cnn_model = CNNModelWithMaxPooling( filter_dims=filter_dims, num_filters=num_filters, strides=strides, padding=padding, pool_strides=pool_strides, pool_shapes=pool_shapes, hidden_nonlinearity=cnn_hidden_nonlinearity) if not dueling: output_model = MLPModel(output_dim=action_dim, hidden_sizes=hidden_sizes, hidden_nonlinearity=hidden_nonlinearity, hidden_w_init=hidden_w_init, hidden_b_init=hidden_b_init, output_nonlinearity=output_nonlinearity, output_w_init=output_w_init, output_b_init=output_b_init, layer_normalization=layer_normalization) else: output_model = MLPDuelingModel( output_dim=action_dim, hidden_sizes=hidden_sizes, hidden_nonlinearity=hidden_nonlinearity, hidden_w_init=hidden_w_init, hidden_b_init=hidden_b_init, output_nonlinearity=output_nonlinearity, output_w_init=output_w_init, output_b_init=output_b_init, layer_normalization=layer_normalization) self.model = Sequential(cnn_model, output_model) self._initialize()
class DiscreteCNNQFunction(QFunction): """Q function based on a CNN-MLP structure for discrete action space. This class implements a Q value network to predict Q based on the input state and action. It uses an CNN and a MLP to fit the function of Q(s, a). Args: env_spec (garage.envs.env_spec.EnvSpec): Environment specification. filter_dims (tuple[int]): Dimension of the filters. For example, (3, 5) means there are two convolutional layers. The filter for first layer is of dimension (3 x 3) and the second one is of dimension (5 x 5). num_filters (tuple[int]): Number of filters. For example, (3, 32) means there are two convolutional layers. The filter for the first layer has 3 channels and the second one with 32 channels. strides (tuple[int]): The stride of the sliding window. For example, (1, 2) means there are two convolutional layers. The stride of the filter for first layer is 1 and that of the second layer is 2. hidden_sizes (list[int]): Output dimension of dense layer(s). For example, (32, 32) means the MLP of this q-function consists of two hidden layers, each with 32 hidden units. name (str): Variable scope of the cnn. padding (str): The type of padding algorithm to use, either 'SAME' or 'VALID'. max_pooling (bool): Boolean for using max pooling layer or not. pool_shapes (tuple[int]): Dimension of the pooling layer(s). For example, (2, 2) means that all the pooling layers have shape (2, 2). pool_strides (tuple[int]): The strides of the pooling layer(s). For example, (2, 2) means that all the pooling layers have strides (2, 2). cnn_hidden_nonlinearity (callable): Activation function for intermediate dense layer(s) in the CNN. It should return a tf.Tensor. Set it to None to maintain a linear activation. hidden_nonlinearity (callable): Activation function for intermediate dense layer(s) in the MLP. It should return a tf.Tensor. Set it to None to maintain a linear activation. hidden_w_init (callable): Initializer function for the weight of intermediate dense layer(s) in the MLP. The function should return a tf.Tensor. hidden_b_init (callable): Initializer function for the bias of intermediate dense layer(s) in the MLP. The function should return a tf.Tensor. output_nonlinearity (callable): Activation function for output dense layer in the MLP. It should return a tf.Tensor. Set it to None to maintain a linear activation. output_w_init (callable): Initializer function for the weight of output dense layer(s) in the MLP. The function should return a tf.Tensor. output_b_init (callable): Initializer function for the bias of output dense layer(s) in the MLP. The function should return a tf.Tensor. dueling (bool): Bool for using dueling network or not. layer_normalization (bool): Bool for using layer normalization or not. """ def __init__(self, env_spec, filter_dims, num_filters, strides, hidden_sizes=[256], name=None, padding='SAME', max_pooling=False, pool_strides=(2, 2), pool_shapes=(2, 2), cnn_hidden_nonlinearity=tf.nn.relu, hidden_nonlinearity=tf.nn.relu, hidden_w_init=tf.glorot_uniform_initializer(), hidden_b_init=tf.zeros_initializer(), output_nonlinearity=None, output_w_init=tf.glorot_uniform_initializer(), output_b_init=tf.zeros_initializer(), dueling=False, layer_normalization=False): super().__init__(name) self._env_spec = env_spec self._action_dim = env_spec.action_space.n self._filter_dims = filter_dims self._num_filters = num_filters self._strides = strides self._hidden_sizes = hidden_sizes self._padding = padding self._max_pooling = max_pooling self._pool_strides = pool_strides self._pool_shapes = pool_shapes self._cnn_hidden_nonlinearity = cnn_hidden_nonlinearity self._hidden_nonlinearity = hidden_nonlinearity self._hidden_w_init = hidden_w_init self._hidden_b_init = hidden_b_init self._output_nonlinearity = output_nonlinearity self._output_w_init = output_w_init self._output_b_init = output_b_init self._layer_normalization = layer_normalization self._dueling = dueling self.obs_dim = self._env_spec.observation_space.shape action_dim = self._env_spec.action_space.flat_dim if not max_pooling: cnn_model = CNNModel(filter_dims=filter_dims, num_filters=num_filters, strides=strides, padding=padding, hidden_nonlinearity=cnn_hidden_nonlinearity) else: cnn_model = CNNModelWithMaxPooling( filter_dims=filter_dims, num_filters=num_filters, strides=strides, padding=padding, pool_strides=pool_strides, pool_shapes=pool_shapes, hidden_nonlinearity=cnn_hidden_nonlinearity) if not dueling: output_model = MLPModel(output_dim=action_dim, hidden_sizes=hidden_sizes, hidden_nonlinearity=hidden_nonlinearity, hidden_w_init=hidden_w_init, hidden_b_init=hidden_b_init, output_nonlinearity=output_nonlinearity, output_w_init=output_w_init, output_b_init=output_b_init, layer_normalization=layer_normalization) else: output_model = MLPDuelingModel( output_dim=action_dim, hidden_sizes=hidden_sizes, hidden_nonlinearity=hidden_nonlinearity, hidden_w_init=hidden_w_init, hidden_b_init=hidden_b_init, output_nonlinearity=output_nonlinearity, output_w_init=output_w_init, output_b_init=output_b_init, layer_normalization=layer_normalization) self.model = Sequential(cnn_model, output_model) self._initialize() def _initialize(self): obs_ph = tf.compat.v1.placeholder(tf.float32, (None, ) + self.obs_dim, name='obs') with tf.compat.v1.variable_scope(self.name) as vs: self._variable_scope = vs self.model.build(obs_ph) @property def q_vals(self): """Q values.""" return self.model.networks['default'].outputs @property def input(self): """Input tf.Tensor of the Q-function.""" return self.model.networks['default'].input def get_qval_sym(self, state_input, name): """Symbolic graph for q-network. Args: state_input (tf.Tensor): The state input tf.Tensor to the network. name (str): Network variable scope. Return: The tf.Tensor output of Discrete CNN QFunction. """ with tf.compat.v1.variable_scope(self._variable_scope): return self.model.build(state_input, name=name) def clone(self, name): """Return a clone of the Q-function. It only copies the configuration of the Q-function, not the parameters. Args: name: Name of the newly created q-function. """ return self.__class__(name=name, env_spec=self._env_spec, filter_dims=self._filter_dims, num_filters=self._num_filters, strides=self._strides, hidden_sizes=self._hidden_sizes, padding=self._padding, max_pooling=self._max_pooling, pool_shapes=self._pool_shapes, pool_strides=self._pool_strides, hidden_nonlinearity=self._hidden_nonlinearity, hidden_w_init=self._hidden_w_init, hidden_b_init=self._hidden_b_init, output_nonlinearity=self._output_nonlinearity, output_w_init=self._output_w_init, output_b_init=self._output_b_init, dueling=self._dueling, layer_normalization=self._layer_normalization) def __setstate__(self, state): """Object.__setstate__.""" self.__dict__.update(state) self._initialize()
class CategoricalCNNPolicy(StochasticPolicy): """CategoricalCNNPolicy. A policy that contains a CNN and a MLP to make prediction based on a categorical distribution. It only works with akro.Discrete action space. Args: env_spec (garage.envs.env_spec.EnvSpec): Environment specification. conv_filter_sizes(tuple[int]): Dimension of the filters. For example, (3, 5) means there are two convolutional layers. The filter for first layer is of dimension (3 x 3) and the second one is of dimension (5 x 5). conv_filters(tuple[int]): Number of filters. For example, (3, 32) means there are two convolutional layers. The filter for the first layer has 3 channels and the second one with 32 channels. conv_strides(tuple[int]): The stride of the sliding window. For example, (1, 2) means there are two convolutional layers. The stride of the filter for first layer is 1 and that of the second layer is 2. conv_pad (str): The type of padding algorithm to use, either 'SAME' or 'VALID'. name (str): Policy name, also the variable scope of the policy. hidden_sizes (list[int]): Output dimension of dense layer(s). For example, (32, 32) means the MLP of this policy consists of two hidden layers, each with 32 hidden units. hidden_nonlinearity (callable): Activation function for intermediate dense layer(s). It should return a tf.Tensor. Set it to None to maintain a linear activation. hidden_w_init (callable): Initializer function for the weight of intermediate dense layer(s). The function should return a tf.Tensor. hidden_b_init (callable): Initializer function for the bias of intermediate dense layer(s). The function should return a tf.Tensor. output_nonlinearity (callable): Activation function for output dense layer. It should return a tf.Tensor. Set it to None to maintain a linear activation. output_w_init (callable): Initializer function for the weight of output dense layer(s). The function should return a tf.Tensor. output_b_init (callable): Initializer function for the bias of output dense layer(s). The function should return a tf.Tensor. layer_normalization (bool): Bool for using layer normalization or not. """ def __init__(self, env_spec, conv_filters, conv_filter_sizes, conv_strides, conv_pad, name='CategoricalCNNPolicy', hidden_sizes=(32, 32), hidden_nonlinearity=tf.nn.relu, hidden_w_init=tf.initializers.glorot_uniform(), hidden_b_init=tf.zeros_initializer(), output_nonlinearity=tf.nn.softmax, output_w_init=tf.initializers.glorot_uniform(), output_b_init=tf.zeros_initializer(), layer_normalization=False): if not isinstance(env_spec.action_space, akro.Discrete): raise ValueError( 'CategoricalCNNPolicy only works with akro.Discrete action ' 'space.') if not isinstance(env_spec.observation_space, akro.Box) or \ not len(env_spec.observation_space.shape) in (2, 3): raise ValueError( '{} can only process 2D, 3D akro.Image or' ' akro.Box observations, but received an env_spec with ' 'observation_space of type {} and shape {}'.format( type(self).__name__, type(env_spec.observation_space).__name__, env_spec.observation_space.shape)) super().__init__(name, env_spec) self.obs_dim = env_spec.observation_space.shape self.action_dim = env_spec.action_space.n self.model = Sequential( CNNModel(filter_dims=conv_filter_sizes, num_filters=conv_filters, strides=conv_strides, padding=conv_pad, hidden_nonlinearity=hidden_nonlinearity, name='CNNModel'), MLPModel(output_dim=self.action_dim, hidden_sizes=hidden_sizes, hidden_nonlinearity=hidden_nonlinearity, hidden_w_init=hidden_w_init, hidden_b_init=hidden_b_init, output_nonlinearity=output_nonlinearity, output_w_init=output_w_init, output_b_init=output_b_init, layer_normalization=layer_normalization, name='MLPModel')) self._initialize() def _initialize(self): if isinstance(self.env_spec.observation_space, akro.Image): state_input = tf.compat.v1.placeholder(tf.uint8, shape=(None, ) + self.obs_dim) state_input = tf.cast(state_input, tf.float32) state_input /= 255.0 else: state_input = tf.compat.v1.placeholder(tf.float32, shape=(None, ) + self.obs_dim) with tf.compat.v1.variable_scope(self.name) as vs: self._variable_scope = vs self.model.build(state_input) self._f_prob = tf.compat.v1.get_default_session().make_callable( self.model.outputs, feed_list=[self.model.input]) @property def vectorized(self): """Vectorized or not. Returns: bool: True if primitive supports vectorized operations. """ return True def dist_info_sym(self, obs_var, state_info_vars=None, name=None): """Build a symbolic graph of the distribution parameters. Args: obs_var (tf.Tensor): Tensor input for symbolic graph. state_info_vars (dict[np.ndarray]): Extra state information, e.g. previous action. name (str): Name for symbolic graph. Returns: dict[tf.Tensor]: Outputs of the symbolic graph of distribution parameters. """ with tf.compat.v1.variable_scope(self._variable_scope): if isinstance(self.env_spec.observation_space, akro.Image): obs_var = tf.cast(obs_var, tf.float32) / 255.0 prob = self.model.build(obs_var, name=name) return dict(prob=prob) def dist_info(self, obs, state_infos=None): """Get distribution parameters. Args: obs (np.ndarray): Observation input. state_infos (dict[np.ndarray]): Extra state information, e.g. previous action. Returns: dict[np.ndarray]: Distribution parameters. """ prob = self._f_prob(obs) return dict(prob=prob) def get_action(self, observation): """Get single action from this policy for the input observation. Args: observation (numpy.ndarray): Observation from environment. Returns: numpy.ndarray: Predicted action. dict[str: np.ndarray]: Action distribution. """ if len(observation.shape) < len(self.obs_dim): observation = self.env_spec.observation_space.unflatten( observation) prob = self._f_prob([observation])[0] action = self.action_space.weighted_sample(prob) return action, dict(prob=prob) def get_actions(self, observations): """Get multiple actions from this policy for the input observations. Args: observations (numpy.ndarray): Observations from environment. Returns: numpy.ndarray: Predicted actions. dict[str: np.ndarray]: Action distributions. """ if len(observations[0].shape) < len(self.obs_dim): observations = self.env_spec.observation_space.unflatten_n( observations) probs = self._f_prob(observations) actions = list(map(self.action_space.weighted_sample, probs)) return actions, dict(prob=probs) @property def distribution(self): """Policy distribution. Returns: garage.tf.distributions.Categorical: Policy distribution. """ return Categorical(self.action_dim) def __getstate__(self): """Object.__getstate__. Returns: dict: The state to be pickled for the instance. """ new_dict = super().__getstate__() del new_dict['_f_prob'] return new_dict def __setstate__(self, state): """Object.__setstate__. Args: state (dict): Unpickled state. """ super().__setstate__(state) self._initialize()