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 normalize_dense_matrix(
self,
input_matrix: str,
features: List[str],
normalization_parameters: Dict[str, NormalizationParameters],
blobname_prefix: str,
) -> Tuple[str, List[str]]:
"""
Normalizes inputs according to parameters. Expects a dense matrix whose ith
column corresponds to feature i.
Note that the Caffe2 BatchBoxCox operator isn't implemented on CUDA GPU so
we need to use a CPU context.
:param input_matrix: Input matrix to normalize.
:param features: Array that maps feature ids to column indices.
:param normalization_parameters: Mapping from feature names to
NormalizationParameters.
:param blobname_prefix: Prefix for input blobs to norm_net.
:param num_output_features: The number of features in an output processed
datapoint. If set to None, this function will compute it.
"""
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CPU)):
feature_starts = self._get_type_boundaries(
features, normalization_parameters)
normalized_input_blobs = []
parameters: List[str] = []
for i, feature_type in enumerate(FEATURE_TYPES):
start_index = feature_starts[i]
if (i + 1) == len(FEATURE_TYPES):
end_index = len(normalization_parameters)
else:
end_index = feature_starts[i + 1]
if start_index == end_index:
continue # No features of this type
sliced_input_features = self._get_input_blob(
blobname_prefix, feature_type)
C2.net().Slice(
[input_matrix],
[sliced_input_features],
starts=[0, start_index],
ends=[-1, end_index],
)
normalized_input_blob, blob_parameters = self.preprocess_blob(
sliced_input_features,
[
normalization_parameters[x]
for x in features[start_index:end_index]
],
)
parameters.extend(blob_parameters)
normalized_input_blobs.append(normalized_input_blob)
for i, inp in enumerate(normalized_input_blobs):
logger.info("input# {}: {}".format(i, inp))
concatenated_input_blob, concatenated_input_blob_dim = C2.Concat(
*normalized_input_blobs, axis=1)
concatenated_input_blob = C2.NanCheck(concatenated_input_blob)
return concatenated_input_blob, 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 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)
MakeForwardPassOps(
model,
self.model_id,
state_action_pairs,
q_values,
self.weights,
self.biases,
self.activations,
self.layers,
self.dropout_ratio,
False,
)
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 concat_states_and_possible_next_actions(
self,
next_state_preprocessed_matrix_blob: str,
possible_next_actions_blob: str,
possible_next_actions_lengths_blob: str,
) -> str:
stacked_states = C2.LengthsTile(next_state_preprocessed_matrix_blob,
possible_next_actions_lengths_blob)
state_action_pairs, _ = C2.Concat(stacked_states,
possible_next_actions_blob,
axis=1)
return state_action_pairs
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(
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 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 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 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(
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)
