def get_max_q_values(self, states: str, possible_actions: str,
use_target_network: bool) -> str:
"""
Takes in an array of states and outputs an array of the same shape
whose ith entry = max_{pna} Q(state_i, pna).
:param states: Numpy array with shape (batch_size, state_dim). Each
row contains a representation of a state.
:param possible_next_actions: Numpy array with shape (batch_size, action_dim).
possible_next_actions[i][j] = 1 iff the agent can take action j from
state i.
:use_target_network: Boolean that indicates whether or not to use this
trainer's TargetNetwork to compute Q values.
"""
q_values = self.get_q_values_all_actions(states, use_target_network)
# Set the q values of impossible actions to a very large negative
# number.
inverse_pna = C2.ConstantFill(possible_actions, value=1.0)
possible_actions_float = C2.Cast(possible_actions,
to=core.DataType.FLOAT)
inverse_pna = C2.Sub(inverse_pna, possible_actions_float)
inverse_pna = C2.Mul(inverse_pna,
self.ACTION_NOT_POSSIBLE_VAL,
broadcast=1)
q_values = C2.Add(q_values, inverse_pna)
q_values_max = C2.ReduceBackMax(q_values, num_reduce_dims=1)
return C2.ExpandDims(q_values_max, dims=[1])
def update_model(self, states: str, actions: str, q_vals_target: str) -> None:
"""
Takes in states, actions, and target q values. Updates the model:
Runs the forward pass, computing Q(states, actions).
Q(states, actions)[i][j] is an approximation of Q*(states[i], action_j).
Comptutes Loss of Q(states, actions) with respect to q_vals_targets
Updates Q Network's weights according to loss and optimizer
:param states: Numpy array with shape (batch_size, state_dim). The ith
row is a representation of the ith transition's state.
:param actions: Numpy array with shape (batch_size, action_dim). The ith
row contains the one-hotted representation of the ith action.
:param q_vals_targets: Numpy array with shape (batch_size, 1). The ith
row is the label to train against for the data from the ith transition.
"""
model = C2.model()
q_vals_target = C2.StopGradient(q_vals_target)
output_blob = C2.NextBlob("train_output")
if self.conv_ml_trainer is not None:
conv_output_blob = C2.NextBlob("conv_output")
self.conv_ml_trainer.make_conv_pass_ops(model, states, conv_output_blob)
states = conv_output_blob
self.ml_trainer.make_forward_pass_ops(model, states, output_blob, False)
q_val_select = C2.ReduceBackSum(C2.Mul(output_blob, actions))
q_values = C2.ExpandDims(q_val_select, dims=[1])
self.loss_blob = self.ml_trainer.generateLossOps(model, q_values, q_vals_target)
model.AddGradientOperators([self.loss_blob])
for param in model.params:
if param in model.param_to_grad:
param_grad = model.param_to_grad[param]
param_grad = C2.NanCheck(param_grad)
self.ml_trainer.addParameterUpdateOps(model)
def update_model(
self,
states: str,
actions: str,
q_vals_target: str,
) -> None:
"""
Takes in states, actions, and target q values. Updates the model:
Runs the forward pass, computing Q(states, actions).
Q(states, actions)[i][j] is an approximation of Q*(states[i], action_j).
Comptutes Loss of Q(states, actions) with respect to q_vals_targets
Updates Q Network's weights according to loss and optimizer
:param states: Numpy array with shape (batch_size, state_dim). The ith
row is a representation of the ith transition's state.
:param actions: Numpy array with shape (batch_size, action_dim). The ith
row contains the one-hotted representation of the ith action.
:param q_vals_targets: Numpy array with shape (batch_size, 1). The ith
row is the label to train against for the data from the ith transition.
"""
model = C2.model()
q_vals_target = C2.StopGradient(q_vals_target)
output_blob = C2.NextBlob("train_output")
MakeForwardPassOps(
model,
self.model_id,
states,
output_blob,
self.weights,
self.biases,
self.activations,
self.layers,
self.dropout_ratio,
False,
)
q_val_select = C2.ReduceBackSum(C2.Mul(output_blob, actions))
q_values = C2.ExpandDims(q_val_select, dims=[1])
self.loss_blob = GenerateLossOps(
model,
q_values,
q_vals_target,
)
model.AddGradientOperators([self.loss_blob])
for param in model.params:
if param in model.param_to_grad:
param_grad = model.param_to_grad[param]
param_grad = C2.NanCheck(param_grad)
AddParameterUpdateOps(
model,
optimizer_input=self.optimizer,
base_learning_rate=self.learning_rate,
gamma=self.gamma,
policy=self.lr_policy,
)