def _create_q_score_net(self) -> None:
self.q_score_model = ModelHelper(name="q_score_" + self.model_id)
C2.set_model(self.q_score_model)
self.q_score_output = self.get_q_values('states', 'actions', True)
workspace.RunNetOnce(self.q_score_model.param_init_net)
workspace.CreateNet(self.q_score_model.net)
C2.set_model(None)
def _create_rl_train_net(self) -> None:
self.rl_train_model = ModelHelper(name="rl_train_" + self.model_id)
C2.set_model(self.rl_train_model)
if self.maxq_learning:
next_q_values = self.get_max_q_values(
'next_states',
self.get_possible_next_actions(),
True,
)
else:
next_q_values = self.get_q_values('next_states', 'next_actions',
True)
q_vals_target = C2.Add(
'rewards',
C2.Mul(
C2.Mul(
C2.Cast('not_terminals',
to=caffe2_pb2.TensorProto.FLOAT), # type: ignore
self.rl_discount_rate,
broadcast=1,
),
next_q_values))
self.update_model('states', 'actions', q_vals_target)
workspace.RunNetOnce(self.rl_train_model.param_init_net)
workspace.CreateNet(self.rl_train_model.net)
C2.set_model(None)
def _create_reward_train_net(self) -> None:
self.reward_train_model = ModelHelper(name="reward_train_" +
self.model_id)
C2.set_model(self.reward_train_model)
self.update_model('states', 'actions', 'rewards')
workspace.RunNetOnce(self.reward_train_model.param_init_net)
workspace.CreateNet(self.reward_train_model.net)
C2.set_model(None)
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=1)
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 _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 _create_q_score_net(self) -> None:
self.q_score_model = ModelHelper(name="q_score_" + self.model_id)
C2.set_model(self.q_score_model)
self.q_score_output = self.get_q_values("states", "actions", True)
workspace.RunNetOnce(self.q_score_model.param_init_net)
self.q_score_model.net.Proto().num_workers = \
RLTrainer.DEFAULT_TRAINING_NUM_WORKERS
workspace.CreateNet(self.q_score_model.net)
C2.set_model(None)
def _create_internal_policy_net(self) -> None:
self.internal_policy_model = ModelHelper(name="q_score_" +
self.model_id)
C2.set_model(self.internal_policy_model)
self.internal_policy_output = C2.FlattenToVec(
self.get_q_values('states', 'actions', False))
workspace.RunNetOnce(self.internal_policy_model.param_init_net)
workspace.CreateNet(self.internal_policy_model.net)
C2.set_model(None)
def _create_internal_policy_net(self) -> None:
self.internal_policy_model = ModelHelper(name="internal_policy_" +
self.model_id)
C2.set_model(self.internal_policy_model)
self.internal_policy_output = self.get_q_values_all_actions(
"states", False)
workspace.RunNetOnce(self.internal_policy_model.param_init_net)
workspace.CreateNet(self.internal_policy_model.net)
C2.set_model(None)
def _create_reward_train_net(self) -> None:
self.reward_train_model = ModelHelper(name="reward_train_" +
self.model_id)
C2.set_model(self.reward_train_model)
self.update_model('states', 'actions', 'rewards')
workspace.RunNetOnce(self.reward_train_model.param_init_net)
self.reward_train_model.net.Proto().num_workers = \
RLTrainer.DEFAULT_TRAINING_NUM_WORKERS
workspace.CreateNet(self.reward_train_model.net)
C2.set_model(None)
def _create_reward_train_net(self) -> None:
self.reward_train_model = ModelHelper(name="reward_train_" +
self.model_id)
C2.set_model(self.reward_train_model)
self.update_model("states", "actions", "rewards")
workspace.RunNetOnce(self.reward_train_model.param_init_net)
self.reward_train_model.net.Proto().num_workers = (
RLTrainer.DEFAULT_TRAINING_NUM_WORKERS)
self.reward_train_model.net.Proto().type = "async_scheduling"
workspace.CreateNet(self.reward_train_model.net)
C2.set_model(None)
def _create_internal_policy_net(self) -> None:
self.internal_policy_model = ModelHelper(name="internal_policy_" +
self.model_id)
C2.set_model(self.internal_policy_model)
self.internal_policy_output = self.get_q_values_all_actions(
"states", False)
workspace.RunNetOnce(self.internal_policy_model.param_init_net)
self.internal_policy_model.net.Proto().num_workers = \
RLTrainer.DEFAULT_TRAINING_NUM_WORKERS
workspace.CreateNet(self.internal_policy_model.net)
C2.set_model(None)
def _create_internal_policy_net(self) -> None:
self.internal_policy_model = ModelHelper(name="q_score_" +
self.model_id)
C2.set_model(self.internal_policy_model)
self.internal_policy_output = C2.FlattenToVec(
self.get_q_values('states', 'actions', False))
workspace.RunNetOnce(self.internal_policy_model.param_init_net)
self.internal_policy_model.net.Proto().num_workers = \
RLTrainer.DEFAULT_TRAINING_NUM_WORKERS
workspace.CreateNet(self.internal_policy_model.net)
C2.set_model(None)
def _create_all_q_score_net(self) -> None:
self.all_q_score_model = ModelHelper(name="all_q_score_" +
self.model_id)
C2.set_model(self.all_q_score_model)
self.all_q_score_output = self.get_q_values_all_actions(
"states", False)
workspace.RunNetOnce(self.all_q_score_model.param_init_net)
self.all_q_score_model.net.Proto().num_workers = (
RLTrainer.DEFAULT_TRAINING_NUM_WORKERS)
self.all_q_score_model.net.Proto().type = "async_scheduling"
workspace.CreateNet(self.all_q_score_model.net)
C2.set_model(None)
def _create_rl_train_net(self) -> None:
self.rl_train_model = ModelHelper(name="rl_train_" + self.model_id)
C2.set_model(self.rl_train_model)
if self.reward_shape is not None:
for action_index, boost in self.reward_shape.items():
action_boost = C2.Mul(
C2.Slice(
"actions", starts=[0, action_index], ends=[-1, action_index + 1]
),
boost,
broadcast=1,
)
C2.net().Sum(["rewards", action_boost], ["rewards"])
if self.maxq_learning:
next_q_values = self.get_max_q_values(
"next_states", self.get_possible_next_actions(), True
)
else:
next_q_values = self.get_q_values("next_states", "next_actions", True)
discount_blob = C2.ConstantFill("time_diff", value=self.rl_discount_rate)
if self.use_seq_num_diff_as_time_diff:
time_diff_adjusted_discount_blob = C2.Pow(
discount_blob, C2.Cast("time_diff", to=caffe2_pb2.TensorProto.FLOAT)
)
else:
time_diff_adjusted_discount_blob = discount_blob
q_vals_target = C2.Add(
"rewards",
C2.Mul(
C2.Mul(
C2.Cast(
"not_terminals", to=caffe2_pb2.TensorProto.FLOAT
), # type: ignore
time_diff_adjusted_discount_blob,
broadcast=1,
),
next_q_values,
),
)
self.update_model("states", "actions", q_vals_target)
workspace.RunNetOnce(self.rl_train_model.param_init_net)
self.rl_train_model.net.Proto().num_workers = (
RLTrainer.DEFAULT_TRAINING_NUM_WORKERS
)
self.rl_train_model.net.Proto().type = "async_scheduling"
workspace.CreateNet(self.rl_train_model.net)
C2.set_model(None)
def _create_reward_train_net(self) -> None:
self.reward_train_model = ModelHelper(name="reward_train_" +
self.model_id)
C2.set_model(self.reward_train_model)
if self.reward_shape is not None:
for action_index, boost in self.reward_shape.items():
action_boost = C2.Mul(
C2.Slice("actions",
starts=[0, action_index],
ends=[-1, action_index + 1]),
boost,
broadcast=1,
)
C2.net().Sum(["rewards", action_boost], ["rewards"])
self.update_model("states", "actions", "rewards")
workspace.RunNetOnce(self.reward_train_model.param_init_net)
workspace.CreateNet(self.reward_train_model.net)
C2.set_model(None)
def _create_rl_train_net(self) -> None:
self.rl_train_model = ModelHelper(name="rl_train_" + self.model_id)
C2.set_model(self.rl_train_model)
if self.reward_shape is not None:
for action_index, boost in self.reward_shape.items():
action_boost = C2.Mul(
C2.Slice(
'actions',
starts=[0, action_index],
ends=[-1, action_index + 1],
),
boost,
broadcast=1,
)
C2.net().Sum(['rewards', action_boost], ['rewards'])
if self.maxq_learning:
next_q_values = self.get_max_q_values(
'next_states',
self.get_possible_next_actions(),
True,
)
else:
next_q_values = self.get_q_values('next_states', 'next_actions',
True)
q_vals_target = C2.Add(
'rewards',
C2.Mul(
C2.Mul(
C2.Cast('not_terminals',
to=caffe2_pb2.TensorProto.FLOAT), # type: ignore
self.rl_discount_rate,
broadcast=1,
),
next_q_values))
self.update_model('states', 'actions', q_vals_target)
workspace.RunNetOnce(self.rl_train_model.param_init_net)
workspace.CreateNet(self.rl_train_model.net)
C2.set_model(None)
def _create_reward_train_net(self) -> None:
self.reward_train_model = ModelHelper(name="reward_train_" +
self.model_id)
C2.set_model(self.reward_train_model)
if self.reward_shape is not None:
for action_index, boost in self.reward_shape.items():
action_boost = C2.Mul(
C2.Slice("actions",
starts=[0, action_index],
ends=[-1, action_index + 1]),
boost,
broadcast=1,
)
C2.net().Sum(["rewards", action_boost], ["rewards"])
self.update_model("states", "actions", "rewards")
workspace.RunNetOnce(self.reward_train_model.param_init_net)
self.reward_train_model.net.Proto().num_workers = (
RLTrainer.DEFAULT_TRAINING_NUM_WORKERS)
self.reward_train_model.net.Proto().type = "async_scheduling"
workspace.CreateNet(self.reward_train_model.net)
C2.set_model(None)
def _create_rl_train_net(self) -> None:
self.rl_train_model = ModelHelper(name="rl_train_" + self.model_id)
C2.set_model(self.rl_train_model)
if self.maxq_learning:
next_q_values = self.get_max_q_values(
'next_states',
self.get_possible_next_actions(),
True,
)
else:
next_q_values = self.get_q_values('next_states', 'next_actions',
True)
discount_blob = C2.ConstantFill("time_diff",
value=self.rl_discount_rate)
time_diff_adjusted_discount_blob = C2.Pow(
discount_blob, C2.Cast("time_diff",
to=caffe2_pb2.TensorProto.FLOAT))
q_vals_target = C2.Add(
"rewards",
C2.Mul(
C2.Mul(
C2.Cast("not_terminals",
to=caffe2_pb2.TensorProto.FLOAT), # type: ignore
time_diff_adjusted_discount_blob,
broadcast=1,
),
next_q_values,
),
)
self.update_model('states', 'actions', q_vals_target)
workspace.RunNetOnce(self.rl_train_model.param_init_net)
self.rl_train_model.net.Proto().num_workers = \
RLTrainer.DEFAULT_TRAINING_NUM_WORKERS
workspace.CreateNet(self.rl_train_model.net)
C2.set_model(None)
def export_critic(
cls,
trainer,
state_normalization_parameters,
action_normalization_parameters,
int_features=False,
model_on_gpu=False,
):
"""Export caffe2 preprocessor net and pytorch critic forward pass as one
caffe2 net.
:param trainer DDPGTrainer
:param state_normalization_parameters state NormalizationParameters
:param action_normalization_parameters action NormalizationParameters
:param int_features boolean indicating if int features blob will be present
"""
input_dim = trainer.state_dim + trainer.action_dim
buffer = PytorchCaffe2Converter.pytorch_net_to_buffer(
trainer.critic, input_dim, model_on_gpu
)
critic_input_blob, critic_output_blob, caffe2_netdef = PytorchCaffe2Converter.buffer_to_caffe2_netdef(
buffer
)
torch_workspace = caffe2_netdef.workspace
parameters = []
for blob_str in torch_workspace.Blobs():
workspace.FeedBlob(blob_str, torch_workspace.FetchBlob(blob_str))
parameters.append(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)
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)
# Don't normalize actions, just go from sparse -> dense
action_dense_matrix, new_parameters = preprocessor.normalize_sparse_matrix(
input_feature_lengths,
input_feature_keys,
input_feature_values,
action_normalization_parameters,
"action_norm",
False,
False,
normalize=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_dense_matrix],
[state_action_normalized, state_action_normalized_dim],
axis=1,
)
net.Copy([state_action_normalized], [critic_input_blob])
workspace.RunNetOnce(model.param_init_net)
workspace.RunNetOnce(torch_init_net)
net.AppendNet(torch_init_net)
net.AppendNet(torch_predict_net)
C2.FlattenToVec(C2.ArgMax(critic_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(critic_output_blob))],
[output_lengths],
value=trainer.critic.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))
C2.net().FlattenToVec([critic_output_blob], [output_values])
workspace.CreateNet(net)
return DDPGPredictor(net, parameters, int_features)
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 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_actor(cls,
trainer,
state_normalization_parameters,
int_features=False):
"""Export caffe2 preprocessor net and pytorch actor forward pass as one
caffe2 net.
:param trainer DDPGTrainer
:param state_normalization_parameters state NormalizationParameters
:param int_features boolean indicating if int features blob will be present
"""
input_dim = len(state_normalization_parameters)
buffer = PytorchCaffe2Converter.pytorch_net_to_buffer(
trainer.actor, input_dim)
actor_input_blob, actor_output_blob, caffe2_netdef =\
PytorchCaffe2Converter.buffer_to_caffe2_netdef(buffer)
torch_workspace = caffe2_netdef.workspace
parameters = []
for blob_str in torch_workspace.Blobs():
workspace.FeedBlob(blob_str, torch_workspace.FetchBlob(blob_str))
parameters.append(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)
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(net, True)
parameters.extend(preprocessor.parameters)
state_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)
net.Copy([state_normalized_dense_matrix], [actor_input_blob])
workspace.RunNetOnce(model.param_init_net)
workspace.RunNetOnce(torch_init_net)
net.AppendNet(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))
C2.net().FlattenToVec([actor_output_blob], [output_values])
workspace.CreateNet(net)
return DDPGPredictor(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)
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)
