def process(
self, sparse_data: StackedAssociativeArray
) -> Tuple[str, str, List[str]]:
lengths_blob = sparse_data.lengths
keys_blob = sparse_data.keys
values_blob = sparse_data.values
MISSING_SCALAR = C2.NextBlob("MISSING_SCALAR")
missing_value = 0.0 if self.set_missing_value_to_zero else MISSING_VALUE
workspace.FeedBlob(MISSING_SCALAR,
np.array([missing_value], dtype=np.float32))
C2.net().GivenTensorFill([], [MISSING_SCALAR],
shape=[],
values=[missing_value])
parameters: List[str] = [MISSING_SCALAR]
assert len(self.sorted_features) > 0, "Sorted features is empty"
dense_input = C2.NextBlob("dense_input")
dense_input_presence = C2.NextBlob("dense_input_presence")
C2.net().SparseToDenseMask(
[keys_blob, values_blob, MISSING_SCALAR, lengths_blob],
[dense_input, dense_input_presence],
mask=self.sorted_features,
return_presence_mask=True,
)
if self.set_missing_value_to_zero:
dense_input_presence = C2.And(
C2.GT(dense_input, -1e-4, broadcast=1),
C2.LT(dense_input, 1e-4, broadcast=1),
)
return dense_input, dense_input_presence, parameters
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 _forward_pass(
cls, model, trainer, normalized_dense_matrix, actions, qnet_output_blob
):
C2.set_model(model)
parameters = []
q_values = "q_values"
C2.net().Copy([qnet_output_blob], [q_values])
action_names = C2.NextBlob("action_names")
parameters.append(action_names)
workspace.FeedBlob(action_names, np.array(actions))
action_range = C2.NextBlob("action_range")
parameters.append(action_range)
workspace.FeedBlob(action_range, np.array(list(range(len(actions)))))
output_shape = C2.Shape(q_values)
output_shape_row_count = C2.Slice(output_shape, starts=[0], ends=[1])
output_row_shape = C2.Slice(q_values, starts=[0, 0], ends=[-1, 1])
output_feature_keys = "output/string_weighted_multi_categorical_features.keys"
workspace.FeedBlob(output_feature_keys, np.zeros(1, dtype=np.int64))
output_feature_keys_matrix = C2.ConstantFill(
output_row_shape, value=0, dtype=caffe2_pb2.TensorProto.INT64
)
# Note: sometimes we need to use an explicit output name, so we call
# C2.net().Fn(...)
C2.net().FlattenToVec([output_feature_keys_matrix], [output_feature_keys])
output_feature_lengths = (
"output/string_weighted_multi_categorical_features.lengths"
)
workspace.FeedBlob(output_feature_lengths, np.zeros(1, dtype=np.int32))
output_feature_lengths_matrix = C2.ConstantFill(
output_row_shape, value=1, dtype=caffe2_pb2.TensorProto.INT32
)
C2.net().FlattenToVec([output_feature_lengths_matrix], [output_feature_lengths])
output_keys = "output/string_weighted_multi_categorical_features.values.keys"
workspace.FeedBlob(output_keys, np.array(["a"]))
C2.net().Tile([action_names, output_shape_row_count], [output_keys], axis=0)
output_lengths_matrix = C2.ConstantFill(
output_row_shape, value=len(actions), dtype=caffe2_pb2.TensorProto.INT32
)
output_lengths = (
"output/string_weighted_multi_categorical_features.values.lengths"
)
workspace.FeedBlob(output_lengths, np.zeros(1, dtype=np.int32))
C2.net().FlattenToVec([output_lengths_matrix], [output_lengths])
output_values = (
"output/string_weighted_multi_categorical_features.values.values"
)
workspace.FeedBlob(output_values, np.array([1.0]))
C2.net().FlattenToVec([q_values], [output_values])
return parameters, q_values
def process(
self,
sorted_features: List[int],
sparse_data: StackedAssociativeArray,
set_missing_value_to_zero: bool = False,
) -> Tuple[str, List[str]]:
lengths_blob = sparse_data.lengths
keys_blob = sparse_data.keys
values_blob = sparse_data.values
MISSING_SCALAR = C2.NextBlob("MISSING_SCALAR")
missing_value = 0.0 if set_missing_value_to_zero else MISSING_VALUE
workspace.FeedBlob(MISSING_SCALAR,
np.array([missing_value], dtype=np.float32))
C2.net().GivenTensorFill([], [MISSING_SCALAR],
shape=[],
values=[missing_value])
parameters: List[str] = [MISSING_SCALAR]
assert len(sorted_features) > 0, "Sorted features is empty"
dense_input = C2.SparseToDenseMask(keys_blob,
values_blob,
MISSING_SCALAR,
lengths_blob,
mask=sorted_features)[0]
return dense_input, parameters
def _sum_deterministic_policy(self, model_names, path):
net = core.Net('DeterministicPolicy')
C2.set_net(net)
output = 'ActionProbabilities'
workspace.FeedBlob(output, np.array([1.0]))
model_outputs = []
for model in model_names:
model_output = '{}_Output'.format(model)
workspace.FeedBlob(model_output, np.array([1.0], dtype=np.float32))
model_outputs.append(model_output)
max_action = C2.FlattenToVec(
C2.ArgMax(C2.Transpose(C2.Sum(*model_outputs)))
)
one_blob = C2.NextBlob('one')
workspace.FeedBlob(one_blob, np.array([1.0], dtype=np.float32))
C2.net().SparseToDense(
[
max_action,
one_blob,
model_outputs[0],
],
[output],
)
meta = PredictorExportMeta(
net,
[one_blob],
model_outputs,
[output],
)
save_to_db('minidb', path, meta)
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 is a representation of the ith transition's 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)
q_values = C2.NextBlob("train_output")
state_action_pairs, _ = C2.Concat(states, actions, axis=1)
self.ml_trainer.make_forward_pass_ops(model, state_action_pairs,
q_values, False)
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 get_q_values(self, states: str, actions: str, use_target_network: bool) -> str:
state_action_pairs, _ = C2.Concat(states, actions, axis=1)
q_values = C2.NextBlob("q_values")
if use_target_network:
self.target_network.make_forward_pass_ops(
C2.model(), state_action_pairs, q_values, True
)
else:
self.ml_trainer.make_forward_pass_ops(
C2.model(), state_action_pairs, q_values, True
)
return q_values
def get_q_values_all_actions(self, states: str, use_target_network: bool) -> str:
"""
Takes in a set of states and runs the test Q Network on them.
Creates Q(states, actions), a blob with shape (batch_size, action_dim).
Q(states, actions)[i][j] is an approximation of Q*(states[i], action_j).
Note that action_j takes on every possible action (of which there are
self.action_dim_. Stores blob in self.output_blob and returns its value.
: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.
"""
all_q_values = C2.NextBlob("all_q_values")
if use_target_network:
if self.conv_target_network is not None:
conv_output_blob = C2.NextBlob("conv_output")
self.conv_target_network.make_conv_pass_ops(
C2.model(), states, conv_output_blob
)
states = conv_output_blob
self.target_network.make_forward_pass_ops(
C2.model(), states, all_q_values, True
)
else:
if self.conv_ml_trainer is not None:
conv_output_blob = C2.NextBlob("conv_output")
self.conv_ml_trainer.make_conv_pass_ops(
C2.model(), states, conv_output_blob
)
states = conv_output_blob
self.ml_trainer.make_forward_pass_ops(
C2.model(), states, all_q_values, True
)
return all_q_values
def test_normalize_dense_matrix_enum(self):
normalization_parameters = {
1:
NormalizationParameters(
identify_types.ENUM,
None,
None,
None,
None,
[12, 4, 2],
None,
None,
None,
),
2:
NormalizationParameters(identify_types.CONTINUOUS, None, 0, 0, 1,
None, None, None, None),
3:
NormalizationParameters(identify_types.ENUM, None, None, None,
None, [15, 3], None, None, None),
}
norm_net = core.Net("net")
C2.set_net(norm_net)
preprocessor = PreprocessorNet()
inputs = np.zeros([4, 3], dtype=np.float32)
feature_ids = [2, 1, 3] # Sorted according to feature type
inputs[:, feature_ids.index(1)] = [12, 4, 2, 2]
inputs[:, feature_ids.index(2)] = [1.0, 2.0, 3.0, 3.0]
inputs[:, feature_ids.index(3)] = [
15, 3, 15, normalization.MISSING_VALUE
]
input_blob = C2.NextBlob("input_blob")
workspace.FeedBlob(input_blob, np.array([0], dtype=np.float32))
normalized_output_blob, _ = preprocessor.normalize_dense_matrix(
input_blob, feature_ids, normalization_parameters, "", False)
workspace.FeedBlob(input_blob, inputs)
workspace.RunNetOnce(norm_net)
normalized_feature_matrix = workspace.FetchBlob(normalized_output_blob)
np.testing.assert_allclose(
np.array([
[1.0, 1, 0, 0, 1, 0],
[2.0, 0, 1, 0, 0, 1],
[3.0, 0, 0, 1, 1, 0],
[3.0, 0, 0, 1, 0, 0], # Missing values should go to all 0
]),
normalized_feature_matrix,
)
def _store_parameter(self, parameters, name, value):
c2_name = C2.NextBlob(name)
if C2.init_net():
C2.init_net().GivenTensorFill(
[],
c2_name,
shape=value.shape,
values=value.flatten(),
dtype=schema.data_type_for_dtype(value.dtype),
)
C2.init_net().AddExternalOutput(c2_name)
else:
workspace.FeedBlob(c2_name, value)
parameters.append(c2_name)
return c2_name
def save_sum_deterministic_policy(model_names, path, db_type):
net = core.Net("DeterministicPolicy")
C2.set_net(net)
output = "ActionProbabilities"
workspace.FeedBlob(output, np.array([1.0]))
model_outputs = []
for model in model_names:
model_output = "{}_Output".format(model)
workspace.FeedBlob(model_output, np.array([[1.0]], dtype=np.float32))
model_outputs.append(model_output)
max_action = C2.FlattenToVec(C2.ArgMax(C2.Transpose(C2.Sum(*model_outputs))))
one_blob = C2.NextBlob("one")
workspace.FeedBlob(one_blob, np.array([1.0], dtype=np.float32))
C2.net().SparseToDense([max_action, one_blob, model_outputs[0]], [output])
meta = PredictorExportMeta(net, [one_blob], model_outputs, [output])
save_to_db(db_type, path, meta)
def sparse_to_dense(lengths_blob: str, keys_blob: str, values_blob: str,
sorted_features: List[int]) -> Tuple[str, List[str]]:
MISSING_SCALAR = C2.NextBlob("MISSING_SCALAR")
workspace.FeedBlob(MISSING_SCALAR,
np.array([MISSING_VALUE], dtype=np.float32))
C2.net().GivenTensorFill([], [MISSING_SCALAR],
shape=[],
values=[MISSING_VALUE])
parameters: List[str] = [MISSING_SCALAR]
assert len(sorted_features) > 0, "Sorted features is empty"
dense_input = C2.SparseToDenseMask(keys_blob,
values_blob,
MISSING_SCALAR,
lengths_blob,
mask=sorted_features)[0]
return dense_input, parameters
def target_values(self, input_blob: str) -> str:
""" Estimates the values for the given inputs using the target network
:param inputs The given inputs
"""
output_blob = C2.NextBlob("output_blob")
MakeForwardPassOps(
C2.model(),
self.tn_model_id,
input_blob,
output_blob,
self._weights,
self._biases,
self._trainer.activations,
self._trainer.layers,
self._trainer.dropout_ratio,
True,
)
return output_blob
def get_q_values_all_actions(
self,
states: str,
use_target_network: bool,
) -> str:
"""
Takes in a set of states and runs the test Q Network on them.
Creates Q(states, actions), a blob with shape (batch_size, action_dim).
Q(states, actions)[i][j] is an approximation of Q*(states[i], action_j).
Note that action_j takes on every possible action (of which there are
self.action_dim_. Stores blob in self.output_blob and returns its value.
: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.
"""
if use_target_network:
return self.target_network.target_values(states)
else:
all_q_values = C2.NextBlob("all_q_values")
MakeForwardPassOps(
C2.model(),
self.model_id + "_score",
states,
all_q_values,
self.weights,
self.biases,
self.activations,
self.layers,
self.dropout_ratio,
True,
)
return all_q_values
def get_q_values(
self,
states: str,
actions: str,
use_target_network: bool,
) -> str:
state_action_pairs, _ = C2.Concat(states, actions, axis=1)
if use_target_network:
return self.target_network.target_values(state_action_pairs)
else:
q_values = C2.NextBlob("q_values")
MakeForwardPassOps(
C2.model(),
self.model_id,
state_action_pairs,
q_values,
self.weights,
self.biases,
self.activations,
self.layers,
self.dropout_ratio,
True,
)
return q_values
def generate_train_net(
cls,
trainer,
model,
min_action_range_tensor_serving,
max_action_range_tensor_serving,
model_on_gpu,
):
input_dim = trainer.state_dim
if isinstance(trainer.actor, DataParallel):
trainer.actor = trainer.actor.module
buffer = PytorchCaffe2Converter.pytorch_net_to_buffer(
trainer.actor, input_dim, model_on_gpu
)
actor_input_blob, actor_output_blob, caffe2_netdef = PytorchCaffe2Converter.buffer_to_caffe2_netdef(
buffer
)
torch_workspace = caffe2_netdef.workspace
torch_parameters = torch_workspace.Blobs()
parameters = []
for blob_str in torch_parameters:
if str(blob_str) == str(actor_input_blob):
continue
workspace.FeedBlob(blob_str, torch_workspace.FetchBlob(blob_str))
parameters.append(str(blob_str))
torch_init_net = core.Net(caffe2_netdef.init_net)
torch_predict_net = core.Net(caffe2_netdef.predict_net)
# While converting to metanetdef, the external_input of predict_net
# will be recomputed. Add the real output of init_net to parameters
# to make sure they will be counted.
parameters.extend(
set(caffe2_netdef.init_net.external_output)
- set(caffe2_netdef.init_net.external_input)
)
# Feed action scaling tensors for serving
min_action_serving_blob = C2.NextBlob("min_action_range_tensor_serving")
workspace.FeedBlob(
min_action_serving_blob, min_action_range_tensor_serving.cpu().data.numpy()
)
parameters.append(str(min_action_serving_blob))
max_action_serving_blob = C2.NextBlob("max_action_range_tensor_serving")
workspace.FeedBlob(
max_action_serving_blob, max_action_range_tensor_serving.cpu().data.numpy()
)
parameters.append(str(max_action_serving_blob))
# Feed action scaling tensors for training [-1, 1] due to tanh actor
min_vals_training = trainer.min_action_range_tensor_training.cpu().data.numpy()
min_action_training_blob = C2.NextBlob("min_action_range_tensor_training")
workspace.FeedBlob(min_action_training_blob, min_vals_training)
parameters.append(str(min_action_training_blob))
max_vals_training = trainer.max_action_range_tensor_training.cpu().data.numpy()
max_action_training_blob = C2.NextBlob("max_action_range_tensor_training")
workspace.FeedBlob(max_action_training_blob, max_vals_training)
parameters.append(str(max_action_training_blob))
return (
torch_init_net,
torch_predict_net,
parameters,
actor_input_blob,
actor_output_blob,
min_action_training_blob,
max_action_training_blob,
min_action_serving_blob,
max_action_serving_blob,
)
def export(cls,
trainer,
actions,
state_normalization_parameters,
int_features=False):
""" Creates a DiscreteActionPredictor from a DiscreteActionTrainer.
:param trainer DiscreteActionTrainer
:param actions list of action names
:param state_normalization_parameters state NormalizationParameters
:param int_features boolean indicating if int features blob will be present
"""
model = model_helper.ModelHelper(name="predictor")
net = model.net
C2.set_model(model)
workspace.FeedBlob('input/image', np.zeros([1, 1, 1, 1],
dtype=np.int32))
workspace.FeedBlob('input/float_features.lengths',
np.zeros(1, dtype=np.int32))
workspace.FeedBlob('input/float_features.keys',
np.zeros(1, dtype=np.int64))
workspace.FeedBlob('input/float_features.values',
np.zeros(1, dtype=np.float32))
input_feature_lengths = 'input_feature_lengths'
input_feature_keys = 'input_feature_keys'
input_feature_values = 'input_feature_values'
if int_features:
workspace.FeedBlob('input/int_features.lengths',
np.zeros(1, dtype=np.int32))
workspace.FeedBlob('input/int_features.keys',
np.zeros(1, dtype=np.int64))
workspace.FeedBlob('input/int_features.values',
np.zeros(1, dtype=np.int32))
C2.net().Cast(['input/int_features.values'],
['input/int_features.values_float'],
dtype=caffe2_pb2.TensorProto.FLOAT)
C2.net().MergeMultiScalarFeatureTensors([
'input/float_features.lengths', 'input/float_features.keys',
'input/float_features.values', 'input/int_features.lengths',
'input/int_features.keys', 'input/int_features.values_float'
], [
input_feature_lengths, input_feature_keys, input_feature_values
])
else:
C2.net().Copy(['input/float_features.lengths'],
[input_feature_lengths])
C2.net().Copy(['input/float_features.keys'], [input_feature_keys])
C2.net().Copy(['input/float_features.values'],
[input_feature_values])
parameters = []
if state_normalization_parameters is not None:
preprocessor = PreprocessorNet(net, True)
parameters.extend(preprocessor.parameters)
normalized_dense_matrix, new_parameters = \
preprocessor.normalize_sparse_matrix(
input_feature_lengths,
input_feature_keys,
input_feature_values,
state_normalization_parameters,
'state_norm',
)
parameters.extend(new_parameters)
else:
# Image input. Note: Currently this does the wrong thing if
# more than one image is passed at a time.
normalized_dense_matrix = 'input/image'
new_parameters, q_values = RLPredictor._forward_pass(
model,
trainer,
normalized_dense_matrix,
actions,
)
parameters.extend(new_parameters)
# Get 1 x n action index tensor under the max_q policy
max_q_act_idxs = 'max_q_policy_actions'
C2.net().Flatten([C2.ArgMax(q_values)], [max_q_act_idxs], axis=0)
shape_of_num_of_states = 'num_states_shape'
C2.net().FlattenToVec([max_q_act_idxs], [shape_of_num_of_states])
num_states, _ = C2.Reshape(C2.Size(shape_of_num_of_states), shape=[1])
# Get 1 x n action index tensor under the softmax policy
temperature = C2.NextBlob("temperature")
parameters.append(temperature)
workspace.FeedBlob(
temperature, np.array([trainer.rl_temperature], dtype=np.float32))
tempered_q_values = C2.Div(q_values, "temperature", broadcast=1)
softmax_values = C2.Softmax(tempered_q_values)
softmax_act_idxs_nested = 'softmax_act_idxs_nested'
C2.net().WeightedSample([softmax_values], [softmax_act_idxs_nested])
softmax_act_idxs = 'softmax_policy_actions'
C2.net().Flatten([softmax_act_idxs_nested], [softmax_act_idxs], axis=0)
# Concat action index tensors to get 2 x n tensor - [[max_q], [softmax]]
# transpose & flatten to get [a1_maxq, a1_softmax, a2_maxq, a2_softmax, ...]
max_q_act_blob = C2.Cast(max_q_act_idxs,
to=caffe2_pb2.TensorProto.INT32)
softmax_act_blob = C2.Cast(softmax_act_idxs,
to=caffe2_pb2.TensorProto.INT32)
C2.net().Append([max_q_act_blob, softmax_act_blob], [max_q_act_blob])
transposed_action_idxs = C2.Transpose(max_q_act_blob)
flat_transposed_action_idxs = C2.FlattenToVec(transposed_action_idxs)
output_values = 'output/string_single_categorical_features.values'
workspace.FeedBlob(output_values, np.zeros(1, dtype=np.int64))
C2.net().Gather(["action_names", flat_transposed_action_idxs],
[output_values])
output_lengths = 'output/string_single_categorical_features.lengths'
workspace.FeedBlob(output_lengths, np.zeros(1, dtype=np.int32))
C2.net().ConstantFill([shape_of_num_of_states], [output_lengths],
value=2,
dtype=caffe2_pb2.TensorProto.INT32)
output_keys = 'output/string_single_categorical_features.keys'
workspace.FeedBlob(output_keys, np.zeros(1, dtype=np.int64))
output_keys_tensor, _ = C2.Concat(
C2.ConstantFill(shape=[1, 1],
value=0,
dtype=caffe2_pb2.TensorProto.INT64),
C2.ConstantFill(shape=[1, 1],
value=1,
dtype=caffe2_pb2.TensorProto.INT64),
axis=0,
)
output_key_tile = C2.Tile(output_keys_tensor, num_states, axis=0)
C2.net().FlattenToVec([output_key_tile], [output_keys])
workspace.RunNetOnce(model.param_init_net)
workspace.CreateNet(net)
return DiscreteActionPredictor(net, parameters, int_features)
def preprocess_blob(self, blob, normalization_parameters):
"""
Takes in a blob and its normalization parameters. Outputs a tuple
whose first element is a blob containing the normalized input blob
and whose second element contains all the parameter blobs used to
create it.
Call this from a CPU context and ensure the input blob exists in it.
"""
parameters: List[str] = []
ZERO = self._store_parameter(parameters, "ZERO",
np.array([0], dtype=np.float32))
MISSING_U = self._store_parameter(
parameters, "MISSING_U",
np.array([MISSING_VALUE + 1e-4], dtype=np.float32))
MISSING_L = self._store_parameter(
parameters, "MISSING_L",
np.array([MISSING_VALUE - 1e-4], dtype=np.float32))
is_empty_l = C2.GT(blob, MISSING_L, broadcast=1)
is_empty_u = C2.LT(blob, MISSING_U, broadcast=1)
is_empty = C2.And(is_empty_l, is_empty_u)
for i in range(len(normalization_parameters) - 1):
if (normalization_parameters[i].feature_type !=
normalization_parameters[i + 1].feature_type):
raise Exception(
"Only one feature type is allowed per call to preprocess_blob!"
)
feature_type = normalization_parameters[0].feature_type
if feature_type == identify_types.BINARY:
TOLERANCE = self._store_parameter(parameters, "TOLERANCE",
np.array(1e-3, dtype=np.float32))
is_gt_zero = C2.GT(blob,
C2.Add(ZERO, TOLERANCE, broadcast=1),
broadcast=1)
is_lt_zero = C2.LT(blob,
C2.Sub(ZERO, TOLERANCE, broadcast=1),
broadcast=1)
bool_blob = C2.Or(is_gt_zero, is_lt_zero)
blob = C2.Cast(bool_blob, to=caffe2_pb2.TensorProto.FLOAT)
elif feature_type == identify_types.PROBABILITY:
blob = C2.Logit(C2.Clip(blob, min=0.01, max=0.99))
elif feature_type == identify_types.ENUM:
for parameter in normalization_parameters:
possible_values = parameter.possible_values
for x in possible_values:
if x < 0:
logger.fatal(
"Invalid enum possible value for feature: " +
str(x) + " " + str(parameter.possible_values))
raise Exception(
"Invalid enum possible value for feature " + blob +
": " + str(x) + " " +
str(parameter.possible_values))
int_blob = C2.Cast(blob, to=core.DataType.INT32)
# Batch one hot transform with MISSING_VALUE as a possible value
feature_lengths = [
len(p.possible_values) + 1 for p in normalization_parameters
]
feature_lengths_blob = self._store_parameter(
parameters,
"feature_lengths_blob",
np.array(feature_lengths, dtype=np.int32),
)
feature_values = [
x for p in normalization_parameters
for x in p.possible_values + [int(MISSING_VALUE)]
]
feature_values_blob = self._store_parameter(
parameters,
"feature_values_blob",
np.array(feature_values, dtype=np.int32),
)
one_hot_output = C2.BatchOneHot(int_blob, feature_lengths_blob,
feature_values_blob)
flattened_one_hot = C2.FlattenToVec(one_hot_output)
# Remove missing values with a mask
cols_to_include = [[1] * len(p.possible_values) + [0]
for p in normalization_parameters]
cols_to_include = [x for col in cols_to_include for x in col]
mask = self._store_parameter(
parameters, "mask", np.array(cols_to_include, dtype=np.int32))
zero_vec = C2.ConstantFill(one_hot_output,
value=0,
dtype=caffe2_pb2.TensorProto.INT32)
repeated_mask_bool = C2.Cast(C2.Add(zero_vec, mask, broadcast=1),
to=core.DataType.BOOL)
flattened_repeated_mask = C2.FlattenToVec(repeated_mask_bool)
flattened_one_hot_proc = C2.NextBlob("flattened_one_hot_proc")
flattened_one_hot_proc_indices = C2.NextBlob(
"flattened_one_hot_proc_indices")
C2.net().BooleanMask(
[flattened_one_hot, flattened_repeated_mask],
[flattened_one_hot_proc, flattened_one_hot_proc_indices],
)
one_hot_shape = C2.Shape(one_hot_output)
shape_delta = self._store_parameter(
parameters,
"shape_delta",
np.array([0, len(normalization_parameters)], dtype=np.int64),
)
target_shape = C2.Sub(one_hot_shape, shape_delta, broadcast=1)
output_int_blob = C2.NextBlob("output_int_blob")
output_int_blob_old_shape = C2.NextBlob(
"output_int_blob_old_shape")
C2.net().Reshape(
[flattened_one_hot_proc, target_shape],
[output_int_blob, output_int_blob_old_shape],
)
output_blob = C2.Cast(output_int_blob, to=core.DataType.FLOAT)
return output_blob, parameters
elif feature_type == identify_types.QUANTILE:
# This transformation replaces a set of values with their quantile.
# The quantile boundaries are provided in the normalization params.
quantile_sizes = [
len(norm.quantiles) for norm in normalization_parameters
]
num_boundaries_blob = self._store_parameter(
parameters,
"num_boundaries_blob",
np.array(quantile_sizes, dtype=np.int32),
)
quantile_values = np.array([], dtype=np.float32)
quantile_labels = np.array([], dtype=np.float32)
for norm in normalization_parameters:
quantile_values = np.append(
quantile_values, np.array(norm.quantiles,
dtype=np.float32))
# TODO: Fix this: the np.unique is making this part not true.
quantile_labels = np.append(
quantile_labels,
np.arange(len(norm.quantiles), dtype=np.float32) /
float(len(norm.quantiles)),
)
quantiles = np.vstack([quantile_values, quantile_labels]).T
quantiles_blob = self._store_parameter(parameters,
"quantiles_blob", quantiles)
quantile_blob = C2.Percentile(blob, quantiles_blob,
num_boundaries_blob)
blob = quantile_blob
elif (feature_type == identify_types.CONTINUOUS
or feature_type == identify_types.BOXCOX):
boxcox_shifts = []
boxcox_lambdas = []
means = []
stddevs = []
for norm in normalization_parameters:
if feature_type == identify_types.BOXCOX:
assert (norm.boxcox_shift is not None
and norm.boxcox_lambda is not None)
boxcox_shifts.append(norm.boxcox_shift)
boxcox_lambdas.append(norm.boxcox_lambda)
means.append(norm.mean)
stddevs.append(norm.stddev)
if feature_type == identify_types.BOXCOX:
boxcox_shift_blob = self._store_parameter(
parameters,
"boxcox_shift",
np.array(boxcox_shifts, dtype=np.float32),
)
boxcox_lambda_blob = self._store_parameter(
parameters,
"boxcox_shift",
np.array(boxcox_lambdas, dtype=np.float32),
)
blob = C2.BatchBoxCox(blob, boxcox_lambda_blob,
boxcox_shift_blob)
means_blob = self._store_parameter(
parameters, "means_blob", np.array([means], dtype=np.float32))
stddevs_blob = self._store_parameter(
parameters, "stddevs_blob",
np.array([stddevs], dtype=np.float32))
blob = C2.Sub(blob, means_blob, broadcast=1, axis=0)
blob = C2.Div(blob, stddevs_blob, broadcast=1, axis=0)
if self.clip_anomalies:
blob = C2.Clip(blob, min=-3.0, max=3.0)
else:
raise NotImplementedError(
"Invalid feature type: {}".format(feature_type))
zeros = C2.ConstantFill(blob, value=0.)
output_blob = C2.Where(is_empty, zeros, blob)
return output_blob, parameters
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
def _store_parameter(self, parameters, name, value):
c2_name = C2.NextBlob(name)
workspace.FeedBlob(c2_name, value)
parameters.append(c2_name)
return c2_name
def export(
cls,
trainer,
state_normalization_parameters,
action_normalization_parameters,
int_features=False,
model_on_gpu=False,
):
"""Export caffe2 preprocessor net and pytorch DQN forward pass as one
caffe2 net.
:param trainer ParametricDQNTrainer
:param state_normalization_parameters state NormalizationParameters
:param action_normalization_parameters action NormalizationParameters
:param int_features boolean indicating if int features blob will be present
:param model_on_gpu boolean indicating if the model is a GPU model or CPU model
"""
input_dim = trainer.num_features
if isinstance(trainer.q_network, DataParallel):
trainer.q_network = trainer.q_network.module
buffer = PytorchCaffe2Converter.pytorch_net_to_buffer(
trainer.q_network, input_dim, model_on_gpu
)
qnet_input_blob, qnet_output_blob, caffe2_netdef = PytorchCaffe2Converter.buffer_to_caffe2_netdef(
buffer
)
torch_workspace = caffe2_netdef.workspace
parameters = torch_workspace.Blobs()
for blob_str in parameters:
workspace.FeedBlob(blob_str, torch_workspace.FetchBlob(blob_str))
torch_init_net = core.Net(caffe2_netdef.init_net)
torch_predict_net = core.Net(caffe2_netdef.predict_net)
# While converting to metanetdef, the external_input of predict_net
# will be recomputed. Add the real output of init_net to parameters
# to make sure they will be counted.
parameters.extend(
set(caffe2_netdef.init_net.external_output)
- set(caffe2_netdef.init_net.external_input)
)
# ensure state and action IDs have no intersection
assert (
len(
set(state_normalization_parameters.keys())
& set(action_normalization_parameters.keys())
)
== 0
)
model = model_helper.ModelHelper(name="predictor")
net = model.net
C2.set_model(model)
workspace.FeedBlob("input/float_features.lengths", np.zeros(1, dtype=np.int32))
workspace.FeedBlob("input/float_features.keys", np.zeros(1, dtype=np.int64))
workspace.FeedBlob("input/float_features.values", np.zeros(1, dtype=np.float32))
input_feature_lengths = "input_feature_lengths"
input_feature_keys = "input_feature_keys"
input_feature_values = "input_feature_values"
if int_features:
workspace.FeedBlob(
"input/int_features.lengths", np.zeros(1, dtype=np.int32)
)
workspace.FeedBlob("input/int_features.keys", np.zeros(1, dtype=np.int64))
workspace.FeedBlob("input/int_features.values", np.zeros(1, dtype=np.int32))
C2.net().Cast(
["input/int_features.values"],
["input/int_features.values_float"],
dtype=caffe2_pb2.TensorProto.FLOAT,
)
C2.net().MergeMultiScalarFeatureTensors(
[
"input/float_features.lengths",
"input/float_features.keys",
"input/float_features.values",
"input/int_features.lengths",
"input/int_features.keys",
"input/int_features.values_float",
],
[input_feature_lengths, input_feature_keys, input_feature_values],
)
else:
C2.net().Copy(["input/float_features.lengths"], [input_feature_lengths])
C2.net().Copy(["input/float_features.keys"], [input_feature_keys])
C2.net().Copy(["input/float_features.values"], [input_feature_values])
preprocessor = PreprocessorNet(True)
sorted_state_features, _ = sort_features_by_normalization(
state_normalization_parameters
)
state_dense_matrix, new_parameters = sparse_to_dense(
input_feature_lengths,
input_feature_keys,
input_feature_values,
sorted_state_features,
)
parameters.extend(new_parameters)
state_normalized_dense_matrix, new_parameters = preprocessor.normalize_dense_matrix(
state_dense_matrix,
sorted_state_features,
state_normalization_parameters,
"state_norm",
False,
)
parameters.extend(new_parameters)
sorted_action_features, _ = sort_features_by_normalization(
action_normalization_parameters
)
action_dense_matrix, new_parameters = sparse_to_dense(
input_feature_lengths,
input_feature_keys,
input_feature_values,
sorted_action_features,
)
parameters.extend(new_parameters)
action_normalized_dense_matrix, new_parameters = preprocessor.normalize_dense_matrix(
action_dense_matrix,
sorted_action_features,
action_normalization_parameters,
"action_norm",
False,
)
parameters.extend(new_parameters)
state_action_normalized = "state_action_normalized"
state_action_normalized_dim = "state_action_normalized_dim"
net.Concat(
[state_normalized_dense_matrix, action_normalized_dense_matrix],
[state_action_normalized, state_action_normalized_dim],
axis=1,
)
net.Copy([state_action_normalized], [qnet_input_blob])
workspace.RunNetOnce(model.param_init_net)
workspace.RunNetOnce(torch_init_net)
net.AppendNet(torch_predict_net)
new_parameters, q_values = RLPredictor._forward_pass(
model, trainer, state_action_normalized, ["Q"], qnet_output_blob
)
parameters.extend(new_parameters)
flat_q_values_key = (
"output/string_weighted_multi_categorical_features.values.values"
)
num_examples, _ = C2.Reshape(C2.Size(flat_q_values_key), shape=[1])
q_value_blob, _ = C2.Reshape(flat_q_values_key, shape=[1, -1])
# Get 1 x n (number of examples) action index tensor under the max_q policy
max_q_act_idxs = "max_q_policy_actions"
C2.net().FlattenToVec([C2.ArgMax(q_value_blob)], [max_q_act_idxs])
max_q_act_blob = C2.Tile(max_q_act_idxs, num_examples, axis=0)
# Get 1 x n (number of examples) action index tensor under the softmax policy
temperature = C2.NextBlob("temperature")
parameters.append(temperature)
workspace.FeedBlob(
temperature, np.array([trainer.rl_temperature], dtype=np.float32)
)
tempered_q_values = C2.Div(q_value_blob, temperature, broadcast=1)
softmax_values = C2.Softmax(tempered_q_values)
softmax_act_idxs_nested = "softmax_act_idxs_nested"
C2.net().WeightedSample([softmax_values], [softmax_act_idxs_nested])
softmax_act_blob = C2.Tile(
C2.FlattenToVec(softmax_act_idxs_nested), num_examples, axis=0
)
# Concat action idx vecs to get 2 x n tensor [[a_maxq, ..], [a_softmax, ..]]
# transpose & flatten to get [a_maxq, a_softmax, a_maxq, a_softmax, ...]
max_q_act_blob = C2.Cast(max_q_act_blob, to=caffe2_pb2.TensorProto.INT64)
softmax_act_blob = C2.Cast(softmax_act_blob, to=caffe2_pb2.TensorProto.INT64)
max_q_act_blob_nested, _ = C2.Reshape(max_q_act_blob, shape=[1, -1])
softmax_act_blob_nested, _ = C2.Reshape(softmax_act_blob, shape=[1, -1])
C2.net().Append(
[max_q_act_blob_nested, softmax_act_blob_nested], [max_q_act_blob_nested]
)
transposed_action_idxs = C2.Transpose(max_q_act_blob_nested)
flat_transposed_action_idxs = C2.FlattenToVec(transposed_action_idxs)
output_values = "output/int_single_categorical_features.values"
workspace.FeedBlob(output_values, np.zeros(1, dtype=np.int64))
C2.net().Copy([flat_transposed_action_idxs], [output_values])
output_lengths = "output/int_single_categorical_features.lengths"
workspace.FeedBlob(output_lengths, np.zeros(1, dtype=np.int32))
C2.net().ConstantFill(
[flat_q_values_key],
[output_lengths],
value=2,
dtype=caffe2_pb2.TensorProto.INT32,
)
output_keys = "output/int_single_categorical_features.keys"
workspace.FeedBlob(output_keys, np.zeros(1, dtype=np.int64))
output_keys_tensor, _ = C2.Concat(
C2.ConstantFill(shape=[1, 1], value=0, dtype=caffe2_pb2.TensorProto.INT64),
C2.ConstantFill(shape=[1, 1], value=1, dtype=caffe2_pb2.TensorProto.INT64),
axis=0,
)
output_key_tile = C2.Tile(output_keys_tensor, num_examples, axis=0)
C2.net().FlattenToVec([output_key_tile], [output_keys])
workspace.CreateNet(net)
return ParametricDQNPredictor(net, torch_init_net, parameters, int_features)
def export(
cls,
trainer,
state_normalization_parameters,
action_normalization_parameters,
):
""" Creates ContinuousActionDQNPredictor from a list of action trainers
:param trainer ContinuousActionDQNPredictor
:param state_features list of state feature names
:param action_features list of action feature names
"""
# ensure state and action IDs have no intersection
assert (len(
set(state_normalization_parameters.keys())
& set(action_normalization_parameters.keys())) == 0)
model = model_helper.ModelHelper(name="predictor")
net = model.net
C2.set_model(model)
workspace.FeedBlob('input/float_features.lengths',
np.zeros(1, dtype=np.int32))
workspace.FeedBlob('input/float_features.keys',
np.zeros(1, dtype=np.int32))
workspace.FeedBlob('input/float_features.values',
np.zeros(1, dtype=np.float32))
preprocessor = PreprocessorNet(net, True)
parameters = []
parameters.extend(preprocessor.parameters)
state_normalized_dense_matrix, new_parameters = \
preprocessor.normalize_sparse_matrix(
'input/float_features.lengths',
'input/float_features.keys',
'input/float_features.values',
state_normalization_parameters,
'state_norm',
)
parameters.extend(new_parameters)
action_normalized_dense_matrix, new_parameters = \
preprocessor.normalize_sparse_matrix(
'input/float_features.lengths',
'input/float_features.keys',
'input/float_features.values',
action_normalization_parameters,
'action_norm',
)
parameters.extend(new_parameters)
state_action_normalized = 'state_action_normalized'
state_action_normalized_dim = 'state_action_normalized_dim'
net.Concat(
[state_normalized_dense_matrix, action_normalized_dense_matrix],
[state_action_normalized, state_action_normalized_dim],
axis=1)
new_parameters, q_values = RLPredictor._forward_pass(
model,
trainer,
state_action_normalized,
['Q'],
)
parameters.extend(new_parameters)
flat_q_values_key = \
'output/string_weighted_multi_categorical_features.values.values'
num_examples, _ = C2.Reshape(C2.Size(flat_q_values_key), shape=[1])
q_value_blob, _ = C2.Reshape(flat_q_values_key, shape=[1, -1])
# Get 1 x n (number of examples) action index tensor under the max_q policy
max_q_act_idxs = 'max_q_policy_actions'
C2.net().FlattenToVec([C2.ArgMax(q_value_blob)], [max_q_act_idxs])
max_q_act_blob = C2.Tile(max_q_act_idxs, num_examples, axis=0)
# Get 1 x n (number of examples) action index tensor under the softmax policy
temperature = C2.NextBlob("temperature")
parameters.append(temperature)
workspace.FeedBlob(
temperature, np.array([trainer.rl_temperature], dtype=np.float32))
tempered_q_values = C2.Div(q_value_blob, "temperature", broadcast=1)
softmax_values = C2.Softmax(tempered_q_values)
softmax_act_idxs_nested = 'softmax_act_idxs_nested'
C2.net().WeightedSample([softmax_values], [softmax_act_idxs_nested])
softmax_act_blob = C2.Tile(C2.FlattenToVec(softmax_act_idxs_nested),
num_examples,
axis=0)
# Concat action idx vecs to get 2 x n tensor [[a_maxq, ..], [a_softmax, ..]]
# transpose & flatten to get [a_maxq, a_softmax, a_maxq, a_softmax, ...]
max_q_act_blob = C2.Cast(max_q_act_blob,
to=caffe2_pb2.TensorProto.INT64)
softmax_act_blob = C2.Cast(softmax_act_blob,
to=caffe2_pb2.TensorProto.INT64)
max_q_act_blob_nested, _ = C2.Reshape(max_q_act_blob, shape=[1, -1])
softmax_act_blob_nested, _ = C2.Reshape(softmax_act_blob,
shape=[1, -1])
C2.net().Append([max_q_act_blob_nested, softmax_act_blob_nested],
[max_q_act_blob_nested])
transposed_action_idxs = C2.Transpose(max_q_act_blob_nested)
flat_transposed_action_idxs = C2.FlattenToVec(transposed_action_idxs)
output_values = 'output/int_single_categorical_features.values'
workspace.FeedBlob(output_values, np.zeros(1, dtype=np.int64))
C2.net().Copy([flat_transposed_action_idxs], [output_values])
output_lengths = 'output/int_single_categorical_features.lengths'
workspace.FeedBlob(output_lengths, np.zeros(1, dtype=np.int32))
C2.net().ConstantFill([flat_q_values_key], [output_lengths],
value=2,
dtype=caffe2_pb2.TensorProto.INT32)
output_keys = 'output/int_single_categorical_features.keys'
workspace.FeedBlob(output_keys, np.zeros(1, dtype=np.int64))
output_keys_tensor, _ = C2.Concat(
C2.ConstantFill(shape=[1, 1],
value=0,
dtype=caffe2_pb2.TensorProto.INT64),
C2.ConstantFill(shape=[1, 1],
value=1,
dtype=caffe2_pb2.TensorProto.INT64),
axis=0,
)
output_key_tile = C2.Tile(output_keys_tensor, num_examples, axis=0)
C2.net().FlattenToVec([output_key_tile], [output_keys])
workspace.RunNetOnce(model.param_init_net)
workspace.CreateNet(net)
return ContinuousActionDQNPredictor(net, parameters)
def export_actor(
cls,
trainer,
state_normalization_parameters,
min_action_range_tensor_serving,
max_action_range_tensor_serving,
int_features=False,
model_on_gpu=False,
):
"""Export caffe2 preprocessor net and pytorch actor forward pass as one
caffe2 net.
:param trainer DDPGTrainer
:param state_normalization_parameters state NormalizationParameters
:param min_action_range_tensor_serving pytorch tensor that specifies
min action value for each dimension
:param max_action_range_tensor_serving pytorch tensor that specifies
min action value for each dimension
:param state_normalization_parameters state NormalizationParameters
:param int_features boolean indicating if int features blob will be present
:param model_on_gpu boolean indicating if the model is a GPU model or CPU model
"""
input_dim = trainer.state_dim
buffer = PytorchCaffe2Converter.pytorch_net_to_buffer(
trainer.actor, input_dim, model_on_gpu
)
actor_input_blob, actor_output_blob, caffe2_netdef = PytorchCaffe2Converter.buffer_to_caffe2_netdef(
buffer
)
torch_workspace = caffe2_netdef.workspace
parameters = torch_workspace.Blobs()
for blob_str in parameters:
workspace.FeedBlob(blob_str, torch_workspace.FetchBlob(blob_str))
torch_init_net = core.Net(caffe2_netdef.init_net)
torch_predict_net = core.Net(caffe2_netdef.predict_net)
model = model_helper.ModelHelper(name="predictor")
net = model.net
C2.set_model(model)
# Feed action scaling tensors for serving
min_action_serving_blob = C2.NextBlob("min_action_range_tensor_serving")
workspace.FeedBlob(
min_action_serving_blob, min_action_range_tensor_serving.cpu().data.numpy()
)
parameters.append(str(min_action_serving_blob))
max_action_serving_blob = C2.NextBlob("max_action_range_tensor_serving")
workspace.FeedBlob(
max_action_serving_blob, max_action_range_tensor_serving.cpu().data.numpy()
)
parameters.append(str(max_action_serving_blob))
# Feed action scaling tensors for training [-1, 1] due to tanh actor
min_vals_training = trainer.min_action_range_tensor_training.cpu().data.numpy()
min_action_training_blob = C2.NextBlob("min_action_range_tensor_training")
workspace.FeedBlob(min_action_training_blob, min_vals_training)
parameters.append(str(min_action_training_blob))
max_vals_training = trainer.max_action_range_tensor_training.cpu().data.numpy()
max_action_training_blob = C2.NextBlob("max_action_range_tensor_training")
workspace.FeedBlob(max_action_training_blob, max_vals_training)
parameters.append(str(max_action_training_blob))
workspace.FeedBlob("input/float_features.lengths", np.zeros(1, dtype=np.int32))
workspace.FeedBlob("input/float_features.keys", np.zeros(1, dtype=np.int64))
workspace.FeedBlob("input/float_features.values", np.zeros(1, dtype=np.float32))
input_feature_lengths = "input_feature_lengths"
input_feature_keys = "input_feature_keys"
input_feature_values = "input_feature_values"
if int_features:
workspace.FeedBlob(
"input/int_features.lengths", np.zeros(1, dtype=np.int32)
)
workspace.FeedBlob("input/int_features.keys", np.zeros(1, dtype=np.int64))
workspace.FeedBlob("input/int_features.values", np.zeros(1, dtype=np.int32))
C2.net().Cast(
["input/int_features.values"],
["input/int_features.values_float"],
dtype=caffe2_pb2.TensorProto.FLOAT,
)
C2.net().MergeMultiScalarFeatureTensors(
[
"input/float_features.lengths",
"input/float_features.keys",
"input/float_features.values",
"input/int_features.lengths",
"input/int_features.keys",
"input/int_features.values_float",
],
[input_feature_lengths, input_feature_keys, input_feature_values],
)
else:
C2.net().Copy(["input/float_features.lengths"], [input_feature_lengths])
C2.net().Copy(["input/float_features.keys"], [input_feature_keys])
C2.net().Copy(["input/float_features.values"], [input_feature_values])
preprocessor = PreprocessorNet(True)
state_normalized_dense_matrix, new_parameters = preprocessor.normalize_sparse_matrix(
input_feature_lengths,
input_feature_keys,
input_feature_values,
state_normalization_parameters,
"state_norm",
False,
False,
)
parameters.extend(new_parameters)
net.Copy([state_normalized_dense_matrix], [actor_input_blob])
workspace.RunNetOnce(model.param_init_net)
workspace.RunNetOnce(torch_init_net)
net.AppendNet(torch_predict_net)
C2.FlattenToVec(C2.ArgMax(actor_output_blob))
output_lengths = "output/float_features.lengths"
workspace.FeedBlob(output_lengths, np.zeros(1, dtype=np.int32))
C2.net().ConstantFill(
[C2.FlattenToVec(C2.ArgMax(actor_output_blob))],
[output_lengths],
value=trainer.actor.layers[-1].out_features,
dtype=caffe2_pb2.TensorProto.INT32,
)
output_keys = "output/float_features.keys"
workspace.FeedBlob(output_keys, np.zeros(1, dtype=np.int32))
C2.net().LengthsRangeFill([output_lengths], [output_keys])
output_values = "output/float_features.values"
workspace.FeedBlob(output_values, np.zeros(1, dtype=np.float32))
# Scale actors actions from [-1, 1] to serving range
prev_range = C2.Sub(max_action_training_blob, min_action_training_blob)
new_range = C2.Sub(max_action_serving_blob, min_action_serving_blob)
subtract_prev_min = C2.Sub(actor_output_blob, min_action_training_blob)
div_by_prev_range = C2.Div(subtract_prev_min, prev_range)
scaled_for_serving_actions = C2.Add(
C2.Mul(div_by_prev_range, new_range), min_action_serving_blob
)
C2.net().FlattenToVec([scaled_for_serving_actions], [output_values])
workspace.CreateNet(net)
return DDPGPredictor(net, parameters, int_features)
def export(
cls,
trainer,
actions,
state_normalization_parameters,
int_features=False,
model_on_gpu=False,
set_missing_value_to_zero=False,
):
"""Export caffe2 preprocessor net and pytorch DQN forward pass as one
caffe2 net.
:param trainer DQNTrainer
:param state_normalization_parameters state NormalizationParameters
:param int_features boolean indicating if int features blob will be present
:param model_on_gpu boolean indicating if the model is a GPU model or CPU model
"""
input_dim = trainer.num_features
q_network = (trainer.q_network.module if isinstance(
trainer.q_network, DataParallel) else trainer.q_network)
buffer = PytorchCaffe2Converter.pytorch_net_to_buffer(
q_network, input_dim, model_on_gpu)
qnet_input_blob, qnet_output_blob, caffe2_netdef = PytorchCaffe2Converter.buffer_to_caffe2_netdef(
buffer)
torch_workspace = caffe2_netdef.workspace
parameters = torch_workspace.Blobs()
for blob_str in parameters:
workspace.FeedBlob(blob_str, torch_workspace.FetchBlob(blob_str))
torch_init_net = core.Net(caffe2_netdef.init_net)
torch_predict_net = core.Net(caffe2_netdef.predict_net)
logger.info("Generated ONNX predict net:")
logger.info(str(torch_predict_net.Proto()))
# While converting to metanetdef, the external_input of predict_net
# will be recomputed. Add the real output of init_net to parameters
# to make sure they will be counted.
parameters.extend(
set(caffe2_netdef.init_net.external_output) -
set(caffe2_netdef.init_net.external_input))
model = model_helper.ModelHelper(name="predictor")
net = model.net
C2.set_model(model)
workspace.FeedBlob("input/image", np.zeros([1, 1, 1, 1],
dtype=np.int32))
workspace.FeedBlob("input/float_features.lengths",
np.zeros(1, dtype=np.int32))
workspace.FeedBlob("input/float_features.keys",
np.zeros(1, dtype=np.int64))
workspace.FeedBlob("input/float_features.values",
np.zeros(1, dtype=np.float32))
input_feature_lengths = "input_feature_lengths"
input_feature_keys = "input_feature_keys"
input_feature_values = "input_feature_values"
if int_features:
workspace.FeedBlob("input/int_features.lengths",
np.zeros(1, dtype=np.int32))
workspace.FeedBlob("input/int_features.keys",
np.zeros(1, dtype=np.int64))
workspace.FeedBlob("input/int_features.values",
np.zeros(1, dtype=np.int32))
C2.net().Cast(
["input/int_features.values"],
["input/int_features.values_float"],
dtype=caffe2_pb2.TensorProto.FLOAT,
)
C2.net().MergeMultiScalarFeatureTensors(
[
"input/float_features.lengths",
"input/float_features.keys",
"input/float_features.values",
"input/int_features.lengths",
"input/int_features.keys",
"input/int_features.values_float",
],
[
input_feature_lengths, input_feature_keys,
input_feature_values
],
)
else:
C2.net().Copy(["input/float_features.lengths"],
[input_feature_lengths])
C2.net().Copy(["input/float_features.keys"], [input_feature_keys])
C2.net().Copy(["input/float_features.values"],
[input_feature_values])
if state_normalization_parameters is not None:
sorted_feature_ids = sort_features_by_normalization(
state_normalization_parameters)[0]
dense_matrix, new_parameters = sparse_to_dense(
input_feature_lengths,
input_feature_keys,
input_feature_values,
sorted_feature_ids,
set_missing_value_to_zero=set_missing_value_to_zero,
)
parameters.extend(new_parameters)
preprocessor_net = PreprocessorNet()
state_normalized_dense_matrix, new_parameters = preprocessor_net.normalize_dense_matrix(
dense_matrix,
sorted_feature_ids,
state_normalization_parameters,
"state_norm_",
True,
)
parameters.extend(new_parameters)
else:
# Image input. Note: Currently this does the wrong thing if
# more than one image is passed at a time.
state_normalized_dense_matrix = "input/image"
net.Copy([state_normalized_dense_matrix], [qnet_input_blob])
workspace.RunNetOnce(model.param_init_net)
workspace.RunNetOnce(torch_init_net)
net.AppendNet(torch_predict_net)
new_parameters, q_values = RLPredictor._forward_pass(
model, trainer, state_normalized_dense_matrix, actions,
qnet_output_blob)
parameters.extend(new_parameters)
# Get 1 x n action index tensor under the max_q policy
max_q_act_idxs = "max_q_policy_actions"
C2.net().Flatten([C2.ArgMax(q_values)], [max_q_act_idxs], axis=0)
shape_of_num_of_states = "num_states_shape"
C2.net().FlattenToVec([max_q_act_idxs], [shape_of_num_of_states])
num_states, _ = C2.Reshape(C2.Size(shape_of_num_of_states), shape=[1])
# Get 1 x n action index tensor under the softmax policy
temperature = C2.NextBlob("temperature")
parameters.append(temperature)
workspace.FeedBlob(
temperature, np.array([trainer.rl_temperature], dtype=np.float32))
tempered_q_values = C2.Div(q_values, temperature, broadcast=1)
softmax_values = C2.Softmax(tempered_q_values)
softmax_act_idxs_nested = "softmax_act_idxs_nested"
C2.net().WeightedSample([softmax_values], [softmax_act_idxs_nested])
softmax_act_idxs = "softmax_policy_actions"
C2.net().Flatten([softmax_act_idxs_nested], [softmax_act_idxs], axis=0)
action_names = C2.NextBlob("action_names")
parameters.append(action_names)
workspace.FeedBlob(action_names, np.array(actions))
# Concat action index tensors to get 2 x n tensor - [[max_q], [softmax]]
# transpose & flatten to get [a1_maxq, a1_softmax, a2_maxq, a2_softmax, ...]
max_q_act_blob = C2.Cast(max_q_act_idxs,
to=caffe2_pb2.TensorProto.INT32)
softmax_act_blob = C2.Cast(softmax_act_idxs,
to=caffe2_pb2.TensorProto.INT32)
C2.net().Append([max_q_act_blob, softmax_act_blob], [max_q_act_blob])
transposed_action_idxs = C2.Transpose(max_q_act_blob)
flat_transposed_action_idxs = C2.FlattenToVec(transposed_action_idxs)
workspace.FeedBlob(OUTPUT_SINGLE_CAT_VALS_NAME,
np.zeros(1, dtype=np.int64))
C2.net().Gather([action_names, flat_transposed_action_idxs],
[OUTPUT_SINGLE_CAT_VALS_NAME])
workspace.FeedBlob(OUTPUT_SINGLE_CAT_LENGTHS_NAME,
np.zeros(1, dtype=np.int32))
C2.net().ConstantFill(
[shape_of_num_of_states],
[OUTPUT_SINGLE_CAT_LENGTHS_NAME],
value=2,
dtype=caffe2_pb2.TensorProto.INT32,
)
workspace.FeedBlob(OUTPUT_SINGLE_CAT_KEYS_NAME,
np.zeros(1, dtype=np.int64))
output_keys_tensor, _ = C2.Concat(
C2.ConstantFill(shape=[1, 1],
value=0,
dtype=caffe2_pb2.TensorProto.INT64),
C2.ConstantFill(shape=[1, 1],
value=1,
dtype=caffe2_pb2.TensorProto.INT64),
axis=0,
)
output_key_tile = C2.Tile(output_keys_tensor, num_states, axis=0)
C2.net().FlattenToVec([output_key_tile], [OUTPUT_SINGLE_CAT_KEYS_NAME])
workspace.CreateNet(net)
return DQNPredictor(net, torch_init_net, parameters, int_features)
def export(
cls,
trainer,
state_normalization_parameters,
action_normalization_parameters,
int_features=False,
):
""" Creates a ContinuousActionDQNPredictor from a ContinuousActionDQNTrainer.
:param trainer ContinuousActionDQNTrainer
:param state_normalization_parameters state NormalizationParameters
:param action_normalization_parameters action NormalizationParameters
:param int_features boolean indicating if int features blob will be present
"""
# ensure state and action IDs have no intersection
assert (
len(
set(state_normalization_parameters.keys())
& set(action_normalization_parameters.keys())
)
== 0
)
model = model_helper.ModelHelper(name="predictor")
net = model.net
C2.set_model(model)
workspace.FeedBlob("input/float_features.lengths", np.zeros(1, dtype=np.int32))
workspace.FeedBlob("input/float_features.keys", np.zeros(1, dtype=np.int64))
workspace.FeedBlob("input/float_features.values", np.zeros(1, dtype=np.float32))
input_feature_lengths = "input_feature_lengths"
input_feature_keys = "input_feature_keys"
input_feature_values = "input_feature_values"
if int_features:
workspace.FeedBlob(
"input/int_features.lengths", np.zeros(1, dtype=np.int32)
)
workspace.FeedBlob("input/int_features.keys", np.zeros(1, dtype=np.int64))
workspace.FeedBlob("input/int_features.values", np.zeros(1, dtype=np.int32))
C2.net().Cast(
["input/int_features.values"],
["input/int_features.values_float"],
dtype=caffe2_pb2.TensorProto.FLOAT,
)
C2.net().MergeMultiScalarFeatureTensors(
[
"input/float_features.lengths",
"input/float_features.keys",
"input/float_features.values",
"input/int_features.lengths",
"input/int_features.keys",
"input/int_features.values_float",
],
[input_feature_lengths, input_feature_keys, input_feature_values],
)
else:
C2.net().Copy(["input/float_features.lengths"], [input_feature_lengths])
C2.net().Copy(["input/float_features.keys"], [input_feature_keys])
C2.net().Copy(["input/float_features.values"], [input_feature_values])
parameters = []
state_normalized_dense_matrix, new_parameters = sparse_to_dense(
input_feature_lengths,
input_feature_keys,
input_feature_values,
state_normalization_parameters,
None,
)
parameters.extend(new_parameters)
action_normalized_dense_matrix, new_parameters = sparse_to_dense(
input_feature_lengths,
input_feature_keys,
input_feature_values,
action_normalization_parameters,
None,
)
parameters.extend(new_parameters)
state_action_normalized = "state_action_normalized"
state_action_normalized_dim = "state_action_normalized_dim"
net.Concat(
[state_normalized_dense_matrix, action_normalized_dense_matrix],
[state_action_normalized, state_action_normalized_dim],
axis=1,
)
new_parameters, q_values = RLPredictor._forward_pass(
model, trainer, state_action_normalized, ["Q"]
)
parameters.extend(new_parameters)
flat_q_values_key = (
"output/string_weighted_multi_categorical_features.values.values"
)
num_examples, _ = C2.Reshape(C2.Size(flat_q_values_key), shape=[1])
q_value_blob, _ = C2.Reshape(flat_q_values_key, shape=[1, -1])
# Get 1 x n (number of examples) action index tensor under the max_q policy
max_q_act_idxs = "max_q_policy_actions"
C2.net().FlattenToVec([C2.ArgMax(q_value_blob)], [max_q_act_idxs])
max_q_act_blob = C2.Tile(max_q_act_idxs, num_examples, axis=0)
# Get 1 x n (number of examples) action index tensor under the softmax policy
temperature = C2.NextBlob("temperature")
parameters.append(temperature)
workspace.FeedBlob(
temperature, np.array([trainer.rl_temperature], dtype=np.float32)
)
tempered_q_values = C2.Div(q_value_blob, temperature, broadcast=1)
softmax_values = C2.Softmax(tempered_q_values)
softmax_act_idxs_nested = "softmax_act_idxs_nested"
C2.net().WeightedSample([softmax_values], [softmax_act_idxs_nested])
softmax_act_blob = C2.Tile(
C2.FlattenToVec(softmax_act_idxs_nested), num_examples, axis=0
)
# Concat action idx vecs to get 2 x n tensor [[a_maxq, ..], [a_softmax, ..]]
# transpose & flatten to get [a_maxq, a_softmax, a_maxq, a_softmax, ...]
max_q_act_blob = C2.Cast(max_q_act_blob, to=caffe2_pb2.TensorProto.INT64)
softmax_act_blob = C2.Cast(softmax_act_blob, to=caffe2_pb2.TensorProto.INT64)
max_q_act_blob_nested, _ = C2.Reshape(max_q_act_blob, shape=[1, -1])
softmax_act_blob_nested, _ = C2.Reshape(softmax_act_blob, shape=[1, -1])
C2.net().Append(
[max_q_act_blob_nested, softmax_act_blob_nested], [max_q_act_blob_nested]
)
transposed_action_idxs = C2.Transpose(max_q_act_blob_nested)
flat_transposed_action_idxs = C2.FlattenToVec(transposed_action_idxs)
output_values = "output/int_single_categorical_features.values"
workspace.FeedBlob(output_values, np.zeros(1, dtype=np.int64))
C2.net().Copy([flat_transposed_action_idxs], [output_values])
output_lengths = "output/int_single_categorical_features.lengths"
workspace.FeedBlob(output_lengths, np.zeros(1, dtype=np.int32))
C2.net().ConstantFill(
[flat_q_values_key],
[output_lengths],
value=2,
dtype=caffe2_pb2.TensorProto.INT32,
)
output_keys = "output/int_single_categorical_features.keys"
workspace.FeedBlob(output_keys, np.zeros(1, dtype=np.int64))
output_keys_tensor, _ = C2.Concat(
C2.ConstantFill(shape=[1, 1], value=0, dtype=caffe2_pb2.TensorProto.INT64),
C2.ConstantFill(shape=[1, 1], value=1, dtype=caffe2_pb2.TensorProto.INT64),
axis=0,
)
output_key_tile = C2.Tile(output_keys_tensor, num_examples, axis=0)
C2.net().FlattenToVec([output_key_tile], [output_keys])
workspace.RunNetOnce(model.param_init_net)
workspace.CreateNet(net)
return ContinuousActionDQNPredictor(net, parameters, int_features)
def export_actor(
cls,
trainer,
state_normalization_parameters,
action_feature_ids,
min_action_range_tensor_serving,
max_action_range_tensor_serving,
int_features=False,
model_on_gpu=False,
):
"""Export caffe2 preprocessor net and pytorch actor forward pass as one
caffe2 net.
:param trainer DDPGTrainer
:param state_normalization_parameters state NormalizationParameters
:param min_action_range_tensor_serving pytorch tensor that specifies
min action value for each dimension
:param max_action_range_tensor_serving pytorch tensor that specifies
min action value for each dimension
:param state_normalization_parameters state NormalizationParameters
:param int_features boolean indicating if int features blob will be present
:param model_on_gpu boolean indicating if the model is a GPU model or CPU model
"""
model = model_helper.ModelHelper(name="predictor")
net = model.net
C2.set_model(model)
parameters: List[str] = []
workspace.FeedBlob("input/float_features.lengths", np.zeros(1, dtype=np.int32))
workspace.FeedBlob("input/float_features.keys", np.zeros(1, dtype=np.int64))
workspace.FeedBlob("input/float_features.values", np.zeros(1, dtype=np.float32))
input_feature_lengths = "input_feature_lengths"
input_feature_keys = "input_feature_keys"
input_feature_values = "input_feature_values"
if int_features:
workspace.FeedBlob(
"input/int_features.lengths", np.zeros(1, dtype=np.int32)
)
workspace.FeedBlob("input/int_features.keys", np.zeros(1, dtype=np.int64))
workspace.FeedBlob("input/int_features.values", np.zeros(1, dtype=np.int32))
C2.net().Cast(
["input/int_features.values"],
["input/int_features.values_float"],
dtype=caffe2_pb2.TensorProto.FLOAT,
)
C2.net().MergeMultiScalarFeatureTensors(
[
"input/float_features.lengths",
"input/float_features.keys",
"input/float_features.values",
"input/int_features.lengths",
"input/int_features.keys",
"input/int_features.values_float",
],
[input_feature_lengths, input_feature_keys, input_feature_values],
)
else:
C2.net().Copy(["input/float_features.lengths"], [input_feature_lengths])
C2.net().Copy(["input/float_features.keys"], [input_feature_keys])
C2.net().Copy(["input/float_features.values"], [input_feature_values])
preprocessor = PreprocessorNet()
sorted_features, _ = sort_features_by_normalization(
state_normalization_parameters
)
state_dense_matrix, new_parameters = sparse_to_dense(
input_feature_lengths,
input_feature_keys,
input_feature_values,
sorted_features,
)
parameters.extend(new_parameters)
state_normalized_dense_matrix, new_parameters = preprocessor.normalize_dense_matrix(
state_dense_matrix,
sorted_features,
state_normalization_parameters,
"state_norm",
False,
)
parameters.extend(new_parameters)
torch_init_net, torch_predict_net, new_parameters, actor_input_blob, actor_output_blob, min_action_training_blob, max_action_training_blob, min_action_serving_blob, max_action_serving_blob = DDPGPredictor.generate_train_net(
trainer,
model,
min_action_range_tensor_serving,
max_action_range_tensor_serving,
model_on_gpu,
)
parameters.extend(new_parameters)
net.Copy([state_normalized_dense_matrix], [actor_input_blob])
workspace.RunNetOnce(model.param_init_net)
workspace.RunNetOnce(torch_init_net)
net.AppendNet(torch_predict_net)
# Scale actors actions from [-1, 1] to serving range
prev_range = C2.Sub(max_action_training_blob, min_action_training_blob)
new_range = C2.Sub(max_action_serving_blob, min_action_serving_blob)
subtract_prev_min = C2.Sub(actor_output_blob, min_action_training_blob)
div_by_prev_range = C2.Div(subtract_prev_min, prev_range)
scaled_for_serving_actions = C2.Add(
C2.Mul(div_by_prev_range, new_range), min_action_serving_blob
)
output_lengths = "output/float_features.lengths"
workspace.FeedBlob(output_lengths, np.zeros(1, dtype=np.int32))
C2.net().ConstantFill(
[C2.FlattenToVec(C2.ArgMax(actor_output_blob))],
[output_lengths],
value=trainer.actor.layers[-1].out_features,
dtype=caffe2_pb2.TensorProto.INT32,
)
action_feature_ids_blob = C2.NextBlob("action_feature_ids")
workspace.FeedBlob(
action_feature_ids_blob, np.array(action_feature_ids, dtype=np.int64)
)
parameters.append(action_feature_ids_blob)
output_keys = "output/float_features.keys"
workspace.FeedBlob(output_keys, np.zeros(1, dtype=np.int64))
num_examples, _ = C2.Reshape(C2.Size("input/float_features.lengths"), shape=[1])
C2.net().Tile([action_feature_ids_blob, num_examples], [output_keys], axis=1)
output_values = "output/float_features.values"
workspace.FeedBlob(output_values, np.zeros(1, dtype=np.float32))
C2.net().FlattenToVec([scaled_for_serving_actions], [output_values])
workspace.CreateNet(net)
return DDPGPredictor(net, torch_init_net, parameters, int_features)
