class MaskedActionsMLP(DistributionalQModel, TFModelV2):
"""Tensorflow model for Envs that provide action masks with observations."""
def __init__(self, obs_space, action_space, num_outputs, model_config,
name, **kwargs):
super().__init__(obs_space, action_space, num_outputs, model_config,
name, **kwargs)
# DictFlatteningPreprocessor, combines all obs components together
# obs.shape for MLP should be a flattened game board obs
original_space = obs_space.original_space['board']
flat_obs_space = spaces.Box(low=np.min(original_space.low),
high=np.max(original_space.high),
shape=(np.prod(original_space.shape), ))
self.mlp = FullyConnectedNetwork(flat_obs_space, action_space,
num_outputs, model_config, name)
self.register_variables(self.mlp.variables())
def forward(self, input_dict, state, seq_lens):
obs = flatten(input_dict['obs']['board'])
action_mask = tf.maximum(tf.log(input_dict['obs']['action_mask']),
tf.float32.min)
model_out, _ = self.mlp({'obs': obs})
return action_mask + model_out, state
def value_function(self):
return self.mlp.value_function()
class CustomTFRPGModel(TFModelV2):
"""Example of interpreting repeated observations."""
def __init__(self, obs_space, action_space, num_outputs, model_config,
name):
super().__init__(obs_space, action_space, num_outputs, model_config,
name)
self.model = TFFCNet(obs_space, action_space, num_outputs,
model_config, name)
self.register_variables(self.model.variables())
def forward(self, input_dict, state, seq_lens):
# The unpacked input tensors, where M=MAX_PLAYERS, N=MAX_ITEMS:
# {
# 'items', <tf.Tensor shape=(?, M, N, 5)>,
# 'location', <tf.Tensor shape=(?, M, 2)>,
# 'status', <tf.Tensor shape=(?, M, 10)>,
# }
print("The unpacked input tensors:", input_dict["obs"])
print()
print("Unbatched repeat dim", input_dict["obs"].unbatch_repeat_dim())
print()
if tf.executing_eagerly():
print("Fully unbatched", input_dict["obs"].unbatch_all())
print()
return self.model.forward(input_dict, state, seq_lens)
def value_function(self):
return self.model.value_function()
class VMActionMaskModel(TFModelV2):
def __init__(self,
obs_space,
action_space,
num_outputs,
model_config,
name,
true_obs_shape=(51, 3),
action_embed_size=50,
*args,
**kwargs):
super(VMActionMaskModel,
self).__init__(obs_space, action_space, num_outputs,
model_config, name, *args, **kwargs)
self.action_embed_model = FullyConnectedNetwork(
spaces.Box(0, 1, shape=true_obs_shape), action_space,
action_embed_size, model_config, name + "_action_embedding")
self.register_variables(self.action_embed_model.variables())
def forward(self, input_dict, state, seq_lens):
avail_actions = input_dict["obs"]["avail_actions"]
action_mask = input_dict["obs"]["action_mask"]
action_embedding, _ = self.action_embed_model(
{"obs": input_dict["obs"]["state"]})
intent_vector = tf.expand_dims(action_embedding, 1)
action_logits = tf.reduce_sum(avail_actions * intent_vector, axis=1)
inf_mask = tf.maximum(tf.log(action_mask), tf.float32.min)
return action_logits + inf_mask, state
def value_function(self):
return self.action_embed_model.value_function()
class ParametricActionsModel(DistributionalQTFModel):
"""Parametric action model that handles the dot product and masking.
This assumes the outputs are logits for a single Categorical action dist.
Getting this to work with a more complex output (e.g., if the action space
is a tuple of several distributions) is also possible but left as an
exercise to the reader.
"""
def __init__(self,
obs_space,
action_space,
num_outputs,
model_config,
name,
true_obs_shape=(4, ),
action_embed_size=6,
**kw):
super(ParametricActionsModel, self).__init__(
obs_space, action_space, num_outputs, model_config, name, **kw)
if model_config['custom_options']['spy']:
true_obs_space = make_spy_space(model_config['custom_options']['parties'], model_config['custom_options']['blocks'])
else:
true_obs_space = make_blind_space(model_config['custom_options']['parties'], model_config['custom_options']['blocks'])
if model_config['custom_options']['extended']:
action_embed_size = 6
else:
action_embed_size = 4
total_dim = 0
for space in true_obs_space:
total_dim += get_preprocessor(space)(space).size
self.action_embed_model = FullyConnectedNetwork(
Box(-1, 1, shape = (total_dim,)), action_space, action_embed_size,
model_config, name + "_action_embed")
self.register_variables(self.action_embed_model.variables())
def forward(self, input_dict, state, seq_lens):
# Extract the available actions tensor from the observation.
avail_actions = input_dict["obs"]["avail_actions"]
action_mask = input_dict["obs"]["action_mask"]
# Compute the predicted action embedding
action_embed, _ = self.action_embed_model({
"obs": input_dict["obs"]["bitcoin"]
})
# Expand the model output to [BATCH, 1, EMBED_SIZE]. Note that the
# avail actions tensor is of shape [BATCH, MAX_ACTIONS, EMBED_SIZE].
intent_vector = tf.expand_dims(action_embed, 1)
# Batch dot product => shape of logits is [BATCH, MAX_ACTIONS].
action_logits = tf.reduce_sum(avail_actions * intent_vector, axis=2)
# Mask out invalid actions (use tf.float32.min for stability)
inf_mask = tf.maximum(tf.log(action_mask), tf.float32.min)
return action_logits + inf_mask, state
def value_function(self):
return self.action_embed_model.value_function()
class CustomModel(TFModelV2):
"""Example of a custom model that just delegates to a fc-net."""
def __init__(self, obs_space, action_space, num_outputs, model_config,
name):
super(CustomModel, self).__init__(obs_space, action_space, num_outputs,
model_config, name)
self.model = FullyConnectedNetwork(obs_space, action_space,
num_outputs, model_config, name)
self.register_variables(self.model.variables())
def forward(self, input_dict, state, seq_lens):
return self.model.forward(input_dict, state, seq_lens)
def value_function(self):
return self.model.value_function()
class FCModel(TFModelV2):
'''Fully Connected Model'''
def __init__(self, obs_space, action_space, num_outputs, model_config,
name):
super(FCModel, self).__init__(obs_space, action_space, num_outputs,
model_config, name)
self.model = FullyConnectedNetwork(obs_space, action_space,
num_outputs, model_config, name)
self.register_variables(self.model.variables())
def forward(self, input_dict, state, seq_lens):
return self.model.forward(input_dict, state, seq_lens)
def value_function(self):
return self.model.value_function()
class ParametricActionsModel(TFModelV2):
""" Parametric model that handles varying action spaces"""
def __init__(self,
obs_space,
action_space,
num_outputs,
model_config,
name,
true_obs_shape=(24, ),
action_embed_size=None):
super(ParametricActionsModel,
self).__init__(obs_space, action_space, num_outputs,
model_config, name)
if action_embed_size is None:
action_embed_size = action_space.n # this works for Discrete() action
# we get the size of the output of the preprocessor automatically chosen by rllib for the real_obs space
real_obs = obs_space.original_space['real_obs']
true_obs_shape = get_preprocessor(real_obs)(
real_obs).size # this will we an integer
# true_obs_shape = obs_space.original_space['real_obs']
self.action_embed_model = FullyConnectedNetwork(
obs_space=Box(-1, 1, shape=(true_obs_shape, )),
action_space=action_space,
num_outputs=action_embed_size,
model_config=model_config,
name=name + "_action_embed")
self.base_model = self.action_embed_model.base_model
self.register_variables(self.action_embed_model.variables())
def forward(self, input_dict, state, seq_lens):
# Compute the predicted action probabilties
# input_dict["obs"]["real_obs"] is a list of 1d tensors if the observation space is a Tuple while
# it should be a tensor. When it is a list we concatenate the various 1d tensors
obs_concat = input_dict["obs"]["real_obs"]
if isinstance(obs_concat, list):
obs_concat = tf.concat(values=flatten_list(obs_concat), axis=1)
action_embed, _ = self.action_embed_model({"obs": obs_concat})
# Mask out invalid actions (use tf.float32.min for stability)
action_mask = input_dict["obs"]["action_mask"]
inf_mask = tf.maximum(tf.math.log(action_mask), tf.float32.min)
return action_embed + inf_mask, state
def value_function(self):
return self.action_embed_model.value_function()
class CentralizedCriticModel(TFModelV2):
"""Multi-agent model that implements a centralized value function."""
def __init__(self, obs_space, action_space, num_outputs, model_config,
name):
super(CentralizedCriticModel,
self).__init__(obs_space, action_space, num_outputs,
model_config, name)
# Base of the model
self.model = FullyConnectedNetwork(obs_space, action_space,
num_outputs, model_config, name)
self.register_variables(self.model.variables())
# Central VF maps (obs, opp_obs, opp_act) -> vf_pred
obs = tf.keras.layers.Input(shape=(6, ), name="obs")
opp_obs = tf.keras.layers.Input(shape=(6, ), name="opp_obs")
opp_act = tf.keras.layers.Input(shape=(2, ), name="opp_act")
concat_obs = tf.keras.layers.Concatenate(axis=1)(
[obs, opp_obs, opp_act])
central_vf_dense = tf.keras.layers.Dense(16,
activation=tf.nn.tanh,
name="c_vf_dense")(concat_obs)
central_vf_out = tf.keras.layers.Dense(
1, activation=None, name="c_vf_out")(central_vf_dense)
self.central_vf = tf.keras.Model(inputs=[obs, opp_obs, opp_act],
outputs=central_vf_out)
self.register_variables(self.central_vf.variables)
@override(ModelV2)
def forward(self, input_dict, state, seq_lens):
return self.model.forward(input_dict, state, seq_lens)
def central_value_function(self, obs, opponent_obs, opponent_actions):
return tf.reshape(
self.central_vf(
[obs, opponent_obs,
tf.one_hot(opponent_actions, 2)]), [-1])
@override(ModelV2)
def value_function(self):
return self.model.value_function() # not used
class CentralizedCriticModel(TFModelV2):
"""Multi-agent model that implements a centralized VF."""
# TODO(@evinitsky) make this work with more than boxes
def __init__(self, obs_space, action_space, num_outputs, model_config,
name):
super(CentralizedCriticModel,
self).__init__(obs_space, action_space, num_outputs,
model_config, name)
# Base of the model
self.model = FullyConnectedNetwork(obs_space, action_space,
num_outputs, model_config, name)
self.register_variables(self.model.variables())
# Central VF maps (obs, opp_ops, opp_act) -> vf_pred
self.max_num_agents = model_config['custom_options']['max_num_agents']
self.obs_space_shape = obs_space.shape[0]
other_obs = tf.keras.layers.Input(shape=(obs_space.shape[0] *
self.max_num_agents, ),
name="opp_obs")
central_vf_dense = tf.keras.layers.Dense(
model_config['custom_options']['central_vf_size'],
activation=tf.nn.tanh,
name="c_vf_dense")(other_obs)
central_vf_out = tf.keras.layers.Dense(
1, activation=None, name="c_vf_out")(central_vf_dense)
self.central_vf = tf.keras.Model(inputs=[other_obs],
outputs=central_vf_out)
self.register_variables(self.central_vf.variables)
def forward(self, input_dict, state, seq_lens):
return self.model.forward(input_dict, state, seq_lens)
def central_value_function(self, obs, opponent_obs):
return tf.reshape(self.central_vf([opponent_obs]), [-1])
def value_function(self):
return self.model.value_function() # not used
class ParametricActionsModel(TFModelV2):
"""
Parametric action model used to filter out invalid action from environment
"""
def import_from_h5(self, h5_file):
pass
def __init__(
self,
obs_space,
action_space,
num_outputs,
model_config,
name,
):
name = "Pa_model"
super(ParametricActionsModel,
self).__init__(obs_space, action_space, num_outputs,
model_config, name)
# get real obs space, discarding action mask
real_obs_space = obs_space.original_space.spaces['array_obs']
# define action embed model
self.action_embed_model = FullyConnectedNetwork(
real_obs_space, action_space, num_outputs, model_config,
name + "_action_embed")
self.register_variables(self.action_embed_model.variables())
def forward(self, input_dict, state, seq_lens):
"""
Override forward pass to mask out invalid actions
Arguments:
input_dict (dict): dictionary of input tensors, including "obs",
"obs_flat", "prev_action", "prev_reward", "is_training"
state (list): list of state tensors with sizes matching those
returned by get_initial_state + the batch dimension
seq_lens (Tensor): 1d tensor holding input sequence lengths
Returns:
(outputs, state): The model output tensor of size
[BATCH, num_outputs]
"""
obs = input_dict['obs']
# extract action mask [batch size, num players]
action_mask = obs['action_mask']
# extract original observations [batch size, obs size]
array_obs = obs['array_obs']
# Compute the predicted action embedding
# size [batch size, num players * num players]
action_embed, _ = self.action_embed_model({"obs": array_obs})
# Mask out invalid actions (use tf.float32.min for stability)
# size [batch size, num players * num players]
inf_mask = tf.maximum(tf.log(action_mask), tf.float32.min)
inf_mask = tf.cast(inf_mask, tf.float32)
masked_actions = action_embed + inf_mask
# return masked action embed and state
return masked_actions, state
def value_function(self):
return self.action_embed_model.value_function()