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,
)
def normalize_sparse_matrix(
self,
lengths_blob: str,
keys_blob: str,
values_blob: str,
normalization_parameters: Dict[int, NormalizationParameters],
blobname_prefix: str,
split_sparse_to_dense: bool,
split_expensive_feature_groups: bool,
normalize: bool = True,
sorted_features_override: List[int] = None,
) -> Tuple[str, List[str]]:
if sorted_features_override:
sorted_features = sorted_features_override
else:
sorted_features, _ = sort_features_by_normalization(
normalization_parameters)
int_features = [int(feature) for feature in sorted_features]
preprocess_num_batches = 8 if split_sparse_to_dense else 1
lengths_batch = []
keys_batch = []
values_batch = []
for _ in range(preprocess_num_batches):
lengths_batch.append(C2.NextBlob(blobname_prefix +
"_length_batch"))
keys_batch.append(C2.NextBlob(blobname_prefix + "_key_batch"))
values_batch.append(C2.NextBlob(blobname_prefix + "_value_batch"))
C2.net().Split([lengths_blob], lengths_batch, axis=0)
total_lengths_batch = []
for x in range(preprocess_num_batches):
total_lengths_batch.append(
C2.Reshape(C2.ReduceBackSum(lengths_batch[x],
num_reduce_dims=1),
shape=[1])[0])
total_lengths_batch_concat, _ = C2.Concat(*total_lengths_batch, axis=0)
C2.net().Split([keys_blob, total_lengths_batch_concat],
keys_batch,
axis=0)
C2.net().Split([values_blob, total_lengths_batch_concat],
values_batch,
axis=0)
dense_input_fragments = []
parameters: List[str] = []
MISSING_SCALAR = self._store_parameter(
parameters, "MISSING_SCALAR",
np.array([MISSING_VALUE], dtype=np.float32))
C2.net().GivenTensorFill([], [MISSING_SCALAR],
shape=[],
values=[MISSING_VALUE])
for preprocess_batch in range(preprocess_num_batches):
dense_input_fragment = C2.SparseToDenseMask(
keys_batch[preprocess_batch],
values_batch[preprocess_batch],
MISSING_SCALAR,
lengths_batch[preprocess_batch],
mask=int_features,
)[0]
if normalize:
normalized_fragment, p = self.normalize_dense_matrix(
dense_input_fragment,
sorted_features,
normalization_parameters,
blobname_prefix,
split_expensive_feature_groups,
)
dense_input_fragments.append(normalized_fragment)
parameters.extend(p)
else:
dense_input_fragments.append(dense_input_fragment)
dense_input = C2.NextBlob(blobname_prefix + "_dense_input")
dense_input_dims = C2.NextBlob(blobname_prefix + "_dense_input_dims")
C2.net().Concat(dense_input_fragments, [dense_input, dense_input_dims],
axis=0)
return dense_input, parameters