def test_normalize_dense_matrix_enum(self):
normalization_parameters = {
1: NormalizationParameters(
identify_types.ENUM,
None,
None,
None,
None,
[12, 4, 2],
None,
None,
None,
),
2: NormalizationParameters(
identify_types.CONTINUOUS, None, 0, 0, 1, None, None, None, None
),
3: NormalizationParameters(
identify_types.ENUM, None, None, None, None, [15, 3], None, None, None
),
}
norm_net = core.Net("net")
C2.set_net(norm_net)
preprocessor = PreprocessorNet()
inputs = np.zeros([4, 3], dtype=np.float32)
feature_ids = [2, 1, 3] # Sorted according to feature type
inputs[:, feature_ids.index(1)] = [12, 4, 2, 2]
inputs[:, feature_ids.index(2)] = [1.0, 2.0, 3.0, 3.0]
inputs[:, feature_ids.index(3)] = [15, 3, 15, normalization.MISSING_VALUE]
input_blob = C2.NextBlob("input_blob")
workspace.FeedBlob(input_blob, np.array([0], dtype=np.float32))
normalized_output_blob, _ = preprocessor.normalize_dense_matrix(
input_blob, feature_ids, normalization_parameters, "", False
)
workspace.FeedBlob(input_blob, inputs)
workspace.RunNetOnce(norm_net)
normalized_feature_matrix = workspace.FetchBlob(normalized_output_blob)
np.testing.assert_allclose(
np.array(
[
[1.0, 1, 0, 0, 1, 0],
[2.0, 0, 1, 0, 0, 1],
[3.0, 0, 0, 1, 1, 0],
[3.0, 0, 0, 1, 0, 0], # Missing values should go to all 0
]
),
normalized_feature_matrix,
)
def 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
if isinstance(trainer.critic, DataParallel):
trainer.critic = trainer.critic.module
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)
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)
# Don't normalize actions, just go from sparse -> dense
action_dense_matrix, new_parameters = sparse_to_dense(
input_feature_lengths,
input_feature_keys,
input_feature_values,
list(action_normalization_parameters.keys()
), # TODO: Clean up in D10161240
)
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_int32 = "output_keys_int32"
output_keys = "output/float_features.keys"
workspace.FeedBlob(output_keys, np.zeros(1, dtype=np.int64))
C2.net().LengthsRangeFill([output_lengths], [output_keys_int32])
C2.net().Cast([output_keys_int32], [output_keys],
to=caffe2_pb2.TensorProto.INT64)
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, 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
if isinstance(trainer.actor, DataParallel):
trainer.actor = trainer.actor.module
buffer = PytorchCaffe2Converter.pytorch_net_to_buffer(
trainer.actor, input_dim, model_on_gpu)
actor_input_blob, actor_output_blob, caffe2_netdef = PytorchCaffe2Converter.buffer_to_caffe2_netdef(
buffer)
torch_workspace = caffe2_netdef.workspace
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))
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)
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)
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_int32 = "output_keys_int32"
output_keys = "output/float_features.keys"
workspace.FeedBlob(output_keys, np.zeros(1, dtype=np.int64))
C2.net().LengthsRangeFill([output_lengths], [output_keys_int32])
C2.net().Cast([output_keys_int32], [output_keys],
to=caffe2_pb2.TensorProto.INT64)
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, torch_init_net, parameters, int_features)
def export(
cls,
trainer,
state_normalization_parameters,
action_normalization_parameters,
int_features=False,
model_on_gpu=False,
normalize_actions=True,
):
"""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))
# Remove the input blob from parameters since it's not a real
# input (will be calculated by preprocessor)
parameters.remove(qnet_input_blob)
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()
sparse_to_dense_processor = Caffe2SparseToDenseProcessor()
sorted_state_features, _ = sort_features_by_normalization(
state_normalization_parameters
)
state_dense_matrix, new_parameters = sparse_to_dense_processor(
sorted_state_features,
StackedAssociativeArray(
input_feature_lengths, input_feature_keys, input_feature_values
),
)
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_processor(
sorted_action_features,
StackedAssociativeArray(
input_feature_lengths, input_feature_keys, input_feature_values
),
)
parameters.extend(new_parameters)
if normalize_actions:
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)
else:
action_normalized_dense_matrix = action_dense_matrix
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 preprocess_samples(self, samples: Samples,
minibatch_size: int) -> List[TrainingDataPage]:
samples.shuffle()
net = core.Net("gridworld_preprocessing")
C2.set_net(net)
preprocessor = PreprocessorNet(True)
saa = StackedAssociativeArray.from_dict_list(samples.states, "states")
state_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization,
"state_norm",
False,
False,
)
saa = StackedAssociativeArray.from_dict_list(samples.next_states,
"next_states")
next_state_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization,
"next_state_norm",
False,
False,
)
saa = StackedAssociativeArray.from_dict_list(samples.actions, "action")
action_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization_action,
"action_norm",
False,
False,
)
saa = StackedAssociativeArray.from_dict_list(samples.next_actions,
"next_action")
next_action_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization_action,
"next_action_norm",
False,
False,
)
propensities = np.array(samples.propensities,
dtype=np.float32).reshape(-1, 1)
rewards = np.array(samples.rewards, dtype=np.float32).reshape(-1, 1)
pnas_lengths_list = []
pnas_flat: List[List[str]] = []
for pnas in samples.possible_next_actions:
pnas_lengths_list.append(len(pnas))
pnas_flat.extend(pnas)
saa = StackedAssociativeArray.from_dict_list(pnas_flat,
"possible_next_actions")
pnas_lengths = np.array(pnas_lengths_list, dtype=np.int32)
pna_lens_blob = "pna_lens_blob"
workspace.FeedBlob(pna_lens_blob, pnas_lengths)
possible_next_actions_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization_action,
"possible_next_action_norm",
False,
False,
)
state_pnas_blob = preprocessor.concat_states_and_possible_actions(
next_state_matrix, possible_next_actions_matrix, pna_lens_blob)
workspace.RunNetOnce(net)
states_ndarray = workspace.FetchBlob(state_matrix)
actions_ndarray = workspace.FetchBlob(action_matrix)
next_states_ndarray = workspace.FetchBlob(next_state_matrix)
next_actions_ndarray = workspace.FetchBlob(next_action_matrix)
possible_next_actions_ndarray = workspace.FetchBlob(
possible_next_actions_matrix)
next_state_pnas_concat = workspace.FetchBlob(state_pnas_blob)
time_diffs = np.ones(len(states_ndarray))
episode_values = None
if samples.reward_timelines is not None:
episode_values = np.zeros(rewards.shape, dtype=np.float32)
for i, reward_timeline in enumerate(samples.reward_timelines):
for time_diff, reward in reward_timeline.items():
episode_values[i, 0] += reward * (DISCOUNT**time_diff)
tdps = []
pnas_start = 0
for start in range(0, states_ndarray.shape[0], minibatch_size):
end = start + minibatch_size
if end > states_ndarray.shape[0]:
break
pnas_end = pnas_start + np.sum(pnas_lengths[start:end])
pnas = possible_next_actions_ndarray[pnas_start:pnas_end]
pnas_start = pnas_end
tdps.append(
TrainingDataPage(
states=states_ndarray[start:end],
actions=actions_ndarray[start:end],
propensities=propensities[start:end],
rewards=rewards[start:end],
next_states=next_states_ndarray[start:end],
next_actions=next_actions_ndarray[start:end],
possible_next_actions=StackedArray(pnas_lengths[start:end],
pnas),
not_terminals=(pnas_lengths[start:end] > 0).reshape(-1, 1),
episode_values=episode_values[start:end]
if episode_values is not None else None,
time_diffs=time_diffs[start:end],
possible_next_actions_lengths=pnas_lengths[start:end],
next_state_pnas_concat=next_state_pnas_concat,
))
return tdps
def preprocess_samples(
self,
states: List[Dict[int, float]],
actions: List[Dict[int, float]],
rewards: List[float],
next_states: List[Dict[int, float]],
next_actions: List[Dict[int, float]],
is_terminals: List[bool],
possible_next_actions: List[List[Dict[int, float]]],
reward_timelines: List[Dict[int, float]],
minibatch_size: int,
) -> List[TrainingDataPage]:
# Shuffle
merged = list(
zip(states, actions, rewards, next_states, next_actions,
is_terminals, possible_next_actions, reward_timelines))
random.shuffle(merged)
states, actions, rewards, next_states, next_actions, is_terminals, \
possible_next_actions, reward_timelines = zip(*merged)
net = core.Net('gridworld_preprocessing')
C2.set_net(net)
preprocessor = PreprocessorNet(net, True)
saa = StackedAssociativeArray.from_dict_list(states, 'states')
state_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization,
'state_norm',
)
saa = StackedAssociativeArray.from_dict_list(next_states,
'next_states')
next_state_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization,
'next_state_norm',
)
saa = StackedAssociativeArray.from_dict_list(actions, 'action')
action_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization_action,
'action_norm',
)
saa = StackedAssociativeArray.from_dict_list(next_actions,
'next_action')
next_action_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization_action,
'next_action_norm',
)
rewards = np.array(rewards, dtype=np.float32).reshape(-1, 1)
pnas_lengths_list = []
pnas_flat = []
for pnas in possible_next_actions:
pnas_lengths_list.append(len(pnas))
pnas_flat.extend(pnas)
saa = StackedAssociativeArray.from_dict_list(pnas_flat,
'possible_next_actions')
pnas_lengths = np.array(pnas_lengths_list, dtype=np.int32)
possible_next_actions_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization_action,
'possible_next_action_norm',
)
workspace.RunNetOnce(net)
states_ndarray = workspace.FetchBlob(state_matrix)
actions_ndarray = workspace.FetchBlob(action_matrix)
next_states_ndarray = workspace.FetchBlob(next_state_matrix)
next_actions_ndarray = workspace.FetchBlob(next_action_matrix)
possible_next_actions_ndarray = workspace.FetchBlob(
possible_next_actions_matrix)
tdps = []
pnas_start = 0
for start in range(0, states_ndarray.shape[0], minibatch_size):
end = start + minibatch_size
if end > states_ndarray.shape[0]:
break
pnas_end = pnas_start + np.sum(pnas_lengths[start:end])
pnas = possible_next_actions_ndarray[pnas_start:pnas_end]
pnas_start = pnas_end
tdps.append(
TrainingDataPage(
states=states_ndarray[start:end],
actions=actions_ndarray[start:end],
rewards=rewards[start:end],
next_states=next_states_ndarray[start:end],
next_actions=next_actions_ndarray[start:end],
possible_next_actions=StackedArray(pnas_lengths[start:end],
pnas),
not_terminals=(pnas_lengths[start:end] > 0).reshape(-1, 1),
reward_timelines=reward_timelines[start:end]
if reward_timelines else None,
))
return tdps
def test_prepare_normalization_and_normalize(self):
feature_value_map = read_data()
normalization_parameters = {}
for name, values in feature_value_map.items():
normalization_parameters[name] = normalization.identify_parameter(
name, values, 10, feature_type=self._feature_type_override(name)
)
for k, v in normalization_parameters.items():
if id_to_type(k) == CONTINUOUS:
self.assertEqual(v.feature_type, CONTINUOUS)
self.assertIs(v.boxcox_lambda, None)
self.assertIs(v.boxcox_shift, None)
elif id_to_type(k) == BOXCOX:
self.assertEqual(v.feature_type, BOXCOX)
self.assertIsNot(v.boxcox_lambda, None)
self.assertIsNot(v.boxcox_shift, None)
else:
assert v.feature_type == id_to_type(k)
sorted_features, _ = sort_features_by_normalization(normalization_parameters)
norm_net = core.Net("net")
C2.set_net(norm_net)
preprocessor = PreprocessorNet()
input_matrix = np.zeros([10000, len(sorted_features)], dtype=np.float32)
for i, feature in enumerate(sorted_features):
input_matrix[:, i] = feature_value_map[feature]
input_matrix_blob = "input_matrix_blob"
workspace.FeedBlob(input_matrix_blob, np.array([], dtype=np.float32))
output_blob, _ = preprocessor.normalize_dense_matrix(
input_matrix_blob, sorted_features, normalization_parameters, "", False
)
workspace.FeedBlob(input_matrix_blob, input_matrix)
workspace.RunNetOnce(norm_net)
normalized_feature_matrix = workspace.FetchBlob(output_blob)
normalized_features = {}
on_column = 0
for feature in sorted_features:
norm = normalization_parameters[feature]
if norm.feature_type == ENUM:
column_size = len(norm.possible_values)
else:
column_size = 1
normalized_features[feature] = normalized_feature_matrix[
:, on_column : (on_column + column_size)
]
on_column += column_size
self.assertTrue(
all(
[
np.isfinite(parameter.stddev) and np.isfinite(parameter.mean)
for parameter in normalization_parameters.values()
]
)
)
for k, v in six.iteritems(normalized_features):
self.assertTrue(np.all(np.isfinite(v)))
feature_type = normalization_parameters[k].feature_type
if feature_type == identify_types.PROBABILITY:
sigmoidv = special.expit(v)
self.assertTrue(
np.all(
np.logical_and(np.greater(sigmoidv, 0), np.less(sigmoidv, 1))
)
)
elif feature_type == identify_types.ENUM:
possible_values = normalization_parameters[k].possible_values
self.assertEqual(v.shape[0], len(feature_value_map[k]))
self.assertEqual(v.shape[1], len(possible_values))
possible_value_map = {}
for i, possible_value in enumerate(possible_values):
possible_value_map[possible_value] = i
for i, row in enumerate(v):
original_feature = feature_value_map[k][i]
self.assertEqual(
possible_value_map[original_feature], np.where(row == 1)[0][0]
)
elif feature_type == identify_types.QUANTILE:
for i, feature in enumerate(v[0]):
original_feature = feature_value_map[k][i]
expected = NumpyFeatureProcessor.value_to_quantile(
original_feature, normalization_parameters[k].quantiles
)
self.assertAlmostEqual(feature, expected, 2)
elif feature_type == identify_types.BINARY:
pass
elif (
feature_type == identify_types.CONTINUOUS
or feature_type == identify_types.BOXCOX
):
one_stddev = np.isclose(np.std(v, ddof=1), 1, atol=0.01)
zero_stddev = np.isclose(np.std(v, ddof=1), 0, atol=0.01)
zero_mean = np.isclose(np.mean(v), 0, atol=0.01)
self.assertTrue(
np.all(zero_mean),
"mean of feature {} is {}, not 0".format(k, np.mean(v)),
)
self.assertTrue(np.all(np.logical_or(one_stddev, zero_stddev)))
elif feature_type == identify_types.CONTINUOUS_ACTION:
less_than_max = v < 1
more_than_min = v > -1
self.assertTrue(
np.all(less_than_max),
"values are not less than 1: {}".format(v[less_than_max == False]),
)
self.assertTrue(
np.all(more_than_min),
"values are not more than -1: {}".format(v[more_than_min == False]),
)
else:
raise NotImplementedError()
def export_actor(
cls,
trainer,
state_normalization_parameters,
action_feature_ids,
min_action_range_tensor_serving,
max_action_range_tensor_serving,
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 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"
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()
sparse_to_dense_processor = Caffe2SparseToDenseProcessor()
sorted_features, _ = sort_features_by_normalization(
state_normalization_parameters)
state_dense_matrix, new_parameters = sparse_to_dense_processor(
sorted_features,
StackedAssociativeArray(input_feature_lengths, input_feature_keys,
input_feature_values),
)
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=0)
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)
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 = C2.NextBlob("input_feature_lengths")
input_feature_keys = C2.NextBlob("input_feature_keys")
input_feature_values = C2.NextBlob("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(True)
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)
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)
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)
output_values = "output/string_single_categorical_features.values"
workspace.FeedBlob(output_values, np.zeros(1, dtype=np.int64))
C2.net().Gather([action_names, flat_transposed_action_idxs],
[output_values])
output_lengths = "output/string_single_categorical_features.lengths"
workspace.FeedBlob(output_lengths, np.zeros(1, dtype=np.int32))
C2.net().ConstantFill(
[shape_of_num_of_states],
[output_lengths],
value=2,
dtype=caffe2_pb2.TensorProto.INT32,
)
output_keys = "output/string_single_categorical_features.keys"
workspace.FeedBlob(output_keys, np.zeros(1, dtype=np.int64))
output_keys_tensor, _ = C2.Concat(
C2.ConstantFill(shape=[1, 1],
value=0,
dtype=caffe2_pb2.TensorProto.INT64),
C2.ConstantFill(shape=[1, 1],
value=1,
dtype=caffe2_pb2.TensorProto.INT64),
axis=0,
)
output_key_tile = C2.Tile(output_keys_tensor, num_states, axis=0)
C2.net().FlattenToVec([output_key_tile], [output_keys])
workspace.RunNetOnce(model.param_init_net)
workspace.CreateNet(net)
return DiscreteActionPredictor(net, parameters, int_features)
def preprocess_samples_discrete(
self, samples: Samples, minibatch_size: int
) -> List[TrainingDataPage]:
samples.shuffle()
net = core.Net("gridworld_preprocessing")
C2.set_net(net)
preprocessor = PreprocessorNet(True)
saa = StackedAssociativeArray.from_dict_list(samples.states, "states")
state_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization,
"state_norm",
False,
False,
)
saa = StackedAssociativeArray.from_dict_list(samples.next_states, "next_states")
next_state_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization,
"next_state_norm",
False,
False,
)
workspace.RunNetOnce(net)
actions_one_hot = np.zeros(
[len(samples.actions), len(self.ACTIONS)], dtype=np.float32
)
for i, action in enumerate(samples.actions):
actions_one_hot[i, self.action_to_index(action)] = 1
rewards = np.array(samples.rewards, dtype=np.float32).reshape(-1, 1)
propensities = np.array(samples.propensities, dtype=np.float32).reshape(-1, 1)
next_actions_one_hot = np.zeros(
[len(samples.next_actions), len(self.ACTIONS)], dtype=np.float32
)
for i, action in enumerate(samples.next_actions):
if action == "":
continue
next_actions_one_hot[i, self.action_to_index(action)] = 1
possible_next_actions_mask = []
for pna in samples.possible_next_actions:
pna_mask = [0] * self.num_actions
for action in pna:
pna_mask[self.action_to_index(action)] = 1
possible_next_actions_mask.append(pna_mask)
possible_next_actions_mask = np.array(
possible_next_actions_mask, dtype=np.float32
)
is_terminals = np.array(samples.is_terminal, dtype=np.bool).reshape(-1, 1)
not_terminals = np.logical_not(is_terminals)
episode_values = None
if samples.reward_timelines is not None:
episode_values = np.zeros(rewards.shape, dtype=np.float32)
for i, reward_timeline in enumerate(samples.reward_timelines):
for time_diff, reward in reward_timeline.items():
episode_values[i, 0] += reward * (DISCOUNT ** time_diff)
states_ndarray = workspace.FetchBlob(state_matrix)
next_states_ndarray = workspace.FetchBlob(next_state_matrix)
time_diffs = np.ones(len(states_ndarray))
tdps = []
for start in range(0, states_ndarray.shape[0], minibatch_size):
end = start + minibatch_size
if end > states_ndarray.shape[0]:
break
tdps.append(
TrainingDataPage(
states=states_ndarray[start:end],
actions=actions_one_hot[start:end],
propensities=propensities[start:end],
rewards=rewards[start:end],
next_states=next_states_ndarray[start:end],
not_terminals=not_terminals[start:end],
next_actions=next_actions_one_hot[start:end],
possible_next_actions=possible_next_actions_mask[start:end],
episode_values=episode_values[start:end]
if episode_values is not None
else None,
time_diffs=time_diffs[start:end],
)
)
return tdps
def export_critic(cls,
trainer,
state_normalization_parameters,
action_normalization_parameters,
int_features=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 =\
len(state_normalization_parameters) + len(action_normalization_parameters)
buffer = PytorchCaffe2Converter.pytorch_net_to_buffer(
trainer.critic, input_dim)
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(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)
action_normalized_dense_matrix, new_parameters = \
preprocessor.normalize_sparse_matrix(
input_feature_lengths,
input_feature_keys,
input_feature_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)
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 from_trainers(cls, trainer, state_features, action_features,
state_normalization_parameters,
action_normalization_parameters):
""" Creates DiscreteActionPredictor from a list of action trainers
:param trainer DiscreteActionTrainer
: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_features) & set(action_features)) == 0)
normalization_parameters = dict(
list(state_normalization_parameters.items()) +
list(action_normalization_parameters.items()))
features = state_features + action_features
int_features = [int(feature) for feature in features]
inputs = [
'input/float_features.lengths', 'input/float_features.keys',
'input/float_features.values'
]
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))
model = model_helper.ModelHelper(name="predictor")
net = model.net
dense_input = net.NextBlob('dense_input')
default_input_value = net.NextBlob('default_input_value')
net.GivenTensorFill([], [default_input_value],
shape=[],
values=[MISSING_VALUE])
net.SparseToDenseMask([
'input/float_features.keys',
'input/float_features.values',
default_input_value,
'input/float_features.lengths',
], [dense_input],
mask=int_features)
for i, feature in enumerate(features):
net.Slice(
[dense_input],
[feature],
starts=[0, i],
ends=[-1, (i + 1)],
)
normalizer = PreprocessorNet(net, True)
parameters = list(normalizer.parameters[:])
normalized_input_blobs = []
zero = "ZERO_from_trainers"
workspace.FeedBlob(zero, np.array(0))
parameters.append(zero)
for feature in features:
normalized_input_blob, blob_parameters = normalizer.preprocess_blob(
feature,
normalization_parameters[feature],
)
parameters.extend(blob_parameters)
normalized_input_blobs.append(normalized_input_blob)
concatenated_input_blob = "PredictorInput"
output_dim = "PredictorOutputDim"
for i, inp in enumerate(normalized_input_blobs):
logger.info("input# {}: {}".format(i, inp))
net.Concat(normalized_input_blobs,
[concatenated_input_blob, output_dim],
axis=1)
net.NanCheck(concatenated_input_blob, concatenated_input_blob)
q_lengths = "output/string_weighted_multi_categorical_features.values.lengths"
workspace.FeedBlob(q_lengths, np.array([1], dtype=np.int32))
q_keys = "output/string_weighted_multi_categorical_features.values.keys"
workspace.FeedBlob(q_keys, np.array(['a']))
q_values_matrix = net.NextBlob('q_values_matrix')
trainer.build_predictor(model, concatenated_input_blob,
q_values_matrix)
parameters.extend(model.GetAllParams())
q_values = 'output/string_weighted_multi_categorical_features.values.values'
workspace.FeedBlob(q_values, np.array([1.0]))
net.FlattenToVec([q_values_matrix], [q_values])
net.ConstantFill([q_values], [q_keys],
value="Q",
dtype=caffe2_pb2.TensorProto.STRING)
net.ConstantFill([q_values], [q_lengths],
value=1,
dtype=caffe2_pb2.TensorProto.INT32)
q_feature_lengths = "output/string_weighted_multi_categorical_features.lengths"
workspace.FeedBlob(q_feature_lengths, np.array([1], dtype=np.int32))
net.ConstantFill([q_values], [q_feature_lengths],
value=1,
dtype=caffe2_pb2.TensorProto.INT32)
q_feature_keys = "output/string_weighted_multi_categorical_features.keys"
workspace.FeedBlob(q_feature_keys, np.array([0], dtype=np.int64))
net.ConstantFill([q_values], [q_feature_keys],
value=0,
dtype=caffe2_pb2.TensorProto.INT64)
output_blobs = [
q_feature_lengths,
q_feature_keys,
q_lengths,
q_keys,
q_values,
]
workspace.RunNetOnce(model.param_init_net)
workspace.CreateNet(net)
predictor = cls(net, inputs, output_blobs, parameters,
workspace.CurrentWorkspace())
return predictor
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,
actions,
state_normalization_parameters,
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 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))
# Remove the input blob from parameters since it's not a real
# input (will be calculated by preprocessor)
parameters.remove(qnet_input_blob)
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"
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]
sparse_to_dense_processor = Caffe2SparseToDenseProcessor()
dense_matrix, new_parameters = sparse_to_dense_processor(
sorted_feature_ids,
StackedAssociativeArray(input_feature_lengths,
input_feature_keys,
input_feature_values),
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)
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])
preprocessor = PreprocessorNet(net, True)
parameters = []
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)
action_normalized_dense_matrix, new_parameters = \
preprocessor.normalize_sparse_matrix(
input_feature_lengths,
input_feature_keys,
input_feature_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, int_features)
def benchmark(num_forward_passes):
"""
Benchmark preprocessor speeds:
1 - PyTorch
2 - PyTorch -> ONNX -> C2
3 - C2
"""
feature_value_map = gen_data(
num_binary_features=10,
num_boxcox_features=10,
num_continuous_features=10,
num_enum_features=10,
num_prob_features=10,
num_quantile_features=10,
)
normalization_parameters = {}
for name, values in feature_value_map.items():
normalization_parameters[name] = normalization.identify_parameter(
name, values, 10
)
sorted_features, _ = sort_features_by_normalization(normalization_parameters)
# Dummy input
input_matrix = np.zeros([10000, len(sorted_features)], dtype=np.float32)
# PyTorch Preprocessor
pytorch_preprocessor = Preprocessor(normalization_parameters, False)
for i, feature in enumerate(sorted_features):
input_matrix[:, i] = feature_value_map[feature]
#################### time pytorch ############################
start = time.time()
for _ in range(NUM_FORWARD_PASSES):
_ = pytorch_preprocessor.forward(input_matrix)
end = time.time()
logger.info(
"PyTorch: {} forward passes done in {} seconds".format(
NUM_FORWARD_PASSES, end - start
)
)
################ time pytorch -> ONNX -> caffe2 ####################
buffer = PytorchCaffe2Converter.pytorch_net_to_buffer(
pytorch_preprocessor, len(sorted_features), False
)
input_blob, 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)
input_matrix_blob = "input_matrix_blob"
workspace.FeedBlob(input_blob, input_matrix)
workspace.RunNetOnce(torch_init_net)
start = time.time()
for _ in range(NUM_FORWARD_PASSES):
workspace.RunNetOnce(torch_predict_net)
_ = workspace.FetchBlob(output_blob)
end = time.time()
logger.info(
"PyTorch -> ONNX -> Caffe2: {} forward passes done in {} seconds".format(
NUM_FORWARD_PASSES, end - start
)
)
#################### time caffe2 ############################
norm_net = core.Net("net")
C2.set_net(norm_net)
preprocessor = PreprocessorNet()
input_matrix_blob = "input_matrix_blob"
workspace.FeedBlob(input_matrix_blob, np.array([], dtype=np.float32))
output_blob, _ = preprocessor.normalize_dense_matrix(
input_matrix_blob, sorted_features, normalization_parameters, "", False
)
workspace.FeedBlob(input_matrix_blob, input_matrix)
start = time.time()
for _ in range(NUM_FORWARD_PASSES):
workspace.RunNetOnce(norm_net)
_ = workspace.FetchBlob(output_blob)
end = time.time()
logger.info(
"Caffe2: {} forward passes done in {} seconds".format(
NUM_FORWARD_PASSES, end - start
)
)
def export(
cls,
trainer,
actions,
state_normalization_parameters,
int_features=False,
model_on_gpu=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
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)
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:
preprocessor = PreprocessorNet(clip_anomalies=True)
state_normalized_dense_matrix, new_parameters = preprocessor.normalize_sparse_matrix(
input_feature_lengths,
input_feature_keys,
input_feature_values,
state_normalization_parameters,
blobname_prefix="state_norm",
split_sparse_to_dense=False,
split_expensive_feature_groups=False,
)
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, parameters, int_features)
def test_preprocessing_network(self):
feature_value_map = read_data()
normalization_parameters = {}
for name, values in feature_value_map.items():
normalization_parameters[name] = normalization.identify_parameter(
name, values, feature_type=self._feature_type_override(name)
)
test_features = NumpyFeatureProcessor.preprocess(
feature_value_map, normalization_parameters
)
net = core.Net("PreprocessingTestNet")
C2.set_net(net)
preprocessor = PreprocessorNet()
name_preprocessed_blob_map = {}
for feature_name in feature_value_map:
workspace.FeedBlob(str(feature_name), np.array([0], dtype=np.int32))
preprocessed_blob, _ = preprocessor.preprocess_blob(
str(feature_name), [normalization_parameters[feature_name]]
)
name_preprocessed_blob_map[feature_name] = preprocessed_blob
workspace.CreateNet(net)
for feature_name, feature_value in six.iteritems(feature_value_map):
feature_value = np.expand_dims(feature_value, -1)
workspace.FeedBlob(str(feature_name), feature_value)
workspace.RunNetOnce(net)
for feature_name in feature_value_map:
normalized_features = workspace.FetchBlob(
name_preprocessed_blob_map[feature_name]
)
if feature_name != ENUM_FEATURE_ID:
normalized_features = np.squeeze(normalized_features, -1)
tolerance = 0.01
if feature_name == BOXCOX_FEATURE_ID:
# At the limit, boxcox has some numerical instability
tolerance = 0.5
non_matching = np.where(
np.logical_not(
np.isclose(
normalized_features,
test_features[feature_name],
rtol=tolerance,
atol=tolerance,
)
)
)
self.assertTrue(
np.all(
np.isclose(
normalized_features,
test_features[feature_name],
rtol=tolerance,
atol=tolerance,
)
),
"{} does not match: {} {}".format(
feature_name,
normalized_features[non_matching].tolist(),
test_features[feature_name][non_matching].tolist(),
),
)
def test_prepare_normalization_and_normalize(self):
features, feature_value_map = preprocessing_util.read_data()
normalization_parameters = {}
for name, values in feature_value_map.items():
normalization_parameters[name] = normalization.identify_parameter(
values, 10)
for k, v in normalization_parameters.items():
if k == CONTINUOUS:
self.assertEqual(v.feature_type, CONTINUOUS)
self.assertIs(v.boxcox_lambda, None)
self.assertIs(v.boxcox_shift, None)
elif k == BOXCOX:
self.assertEqual(v.feature_type, BOXCOX)
self.assertIsNot(v.boxcox_lambda, None)
self.assertIsNot(v.boxcox_shift, None)
else:
assert v.feature_type == k or v.feature_type + "_2" + k
norm_net = core.Net("net")
C2.set_net(norm_net)
preprocessor = PreprocessorNet(norm_net, False)
input_matrix = np.zeros([10000, len(features)], dtype=np.float32)
for i, feature in enumerate(features):
input_matrix[:, i] = feature_value_map[feature]
input_matrix_blob = 'input_matrix_blob'
workspace.FeedBlob(input_matrix_blob, np.array([], dtype=np.float32))
output_blob, _ = preprocessor.normalize_dense_matrix(
input_matrix_blob, features, normalization_parameters, '')
workspace.FeedBlob(input_matrix_blob, input_matrix)
workspace.RunNetOnce(norm_net)
normalized_feature_matrix = workspace.FetchBlob(output_blob)
normalized_features = {}
on_column = 0
for feature in features:
norm = normalization_parameters[feature]
if norm.feature_type == ENUM:
column_size = len(norm.possible_values)
else:
column_size = 1
normalized_features[feature] = \
normalized_feature_matrix[:, on_column:(
on_column + column_size
)]
on_column += column_size
self.assertTrue(
all([
np.isfinite(parameter.stddev) and np.isfinite(parameter.mean)
for parameter in normalization_parameters.values()
]))
for k, v in six.iteritems(normalized_features):
self.assertTrue(np.all(np.isfinite(v)))
feature_type = normalization_parameters[k].feature_type
if feature_type == identify_types.PROBABILITY:
sigmoidv = special.expit(v)
self.assertTrue(
np.all(
np.logical_and(np.greater(sigmoidv, 0),
np.less(sigmoidv, 1))))
elif feature_type == identify_types.ENUM:
possible_values = normalization_parameters[k].possible_values
self.assertEqual(v.shape[0], len(feature_value_map[k]))
self.assertEqual(v.shape[1], len(possible_values))
possible_value_map = {}
for i, possible_value in enumerate(possible_values):
possible_value_map[possible_value] = i
for i, row in enumerate(v):
original_feature = feature_value_map[k][i]
self.assertEqual(possible_value_map[original_feature],
np.where(row == 1)[0][0])
elif feature_type == identify_types.QUANTILE:
for i, feature in enumerate(v[0]):
original_feature = feature_value_map[k][i]
expected = self._value_to_quantile(
original_feature,
normalization_parameters[k].quantiles)
self.assertAlmostEqual(feature, expected, 2)
elif feature_type == identify_types.BINARY:
pass
elif feature_type == identify_types.CONTINUOUS or \
feature_type == identify_types.BOXCOX:
one_stddev = np.isclose(np.std(v, ddof=1), 1, atol=0.01)
zero_stddev = np.isclose(np.std(v, ddof=1), 0, atol=0.01)
zero_mean = np.isclose(np.mean(v), 0, atol=0.01)
self.assertTrue(
np.all(zero_mean),
'mean of feature {} is {}, not 0'.format(k, np.mean(v)))
self.assertTrue(np.all(np.logical_or(one_stddev, zero_stddev)))
else:
raise NotImplementedError()
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 preprocess_samples_discrete(
self,
states: List[Dict[int, float]],
actions: List[str],
rewards: List[float],
next_states: List[Dict[int, float]],
next_actions: List[str],
is_terminals: List[bool],
possible_next_actions: List[List[str]],
reward_timelines: Optional[List[Dict[int, float]]],
minibatch_size: int,
) -> List[TrainingDataPage]:
# Shuffle
if reward_timelines is None:
merged = list(
zip(states, actions, rewards, next_states, next_actions,
is_terminals, possible_next_actions))
random.shuffle(merged)
states, actions, rewards, next_states, next_actions, \
is_terminals, possible_next_actions = zip(*merged)
else:
merged = list(
zip(states, actions, rewards, next_states, next_actions,
is_terminals, possible_next_actions, reward_timelines))
random.shuffle(merged)
states, actions, rewards, next_states, next_actions, \
is_terminals, possible_next_actions, reward_timelines = zip(*merged)
net = core.Net('gridworld_preprocessing')
C2.set_net(net)
preprocessor = PreprocessorNet(net, True)
saa = StackedAssociativeArray.from_dict_list(states, 'states')
state_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization,
'state_norm',
)
saa = StackedAssociativeArray.from_dict_list(next_states,
'next_states')
next_state_matrix, _ = preprocessor.normalize_sparse_matrix(
saa.lengths,
saa.keys,
saa.values,
self.normalization,
'next_state_norm',
)
workspace.RunNetOnce(net)
actions_one_hot = np.zeros(
[len(actions), len(self.ACTIONS)], dtype=np.float32)
for i, action in enumerate(actions):
actions_one_hot[i, self.ACTIONS.index(action)] = 1
rewards = np.array(rewards, dtype=np.float32).reshape(-1, 1)
next_actions_one_hot = np.zeros(
[len(next_actions), len(self.ACTIONS)], dtype=np.float32)
for i, action in enumerate(next_actions):
if action == '':
continue
next_actions_one_hot[i, self.ACTIONS.index(action)] = 1
possible_next_actions_mask = []
for pna in possible_next_actions:
pna_mask = [0] * self.num_actions
for action in pna:
pna_mask[self.ACTIONS.index(action)] = 1
possible_next_actions_mask.append(pna_mask)
possible_next_actions_mask = np.array(possible_next_actions_mask,
dtype=np.float32)
is_terminals = np.array(is_terminals, dtype=np.bool).reshape(-1, 1)
not_terminals = np.logical_not(is_terminals)
if reward_timelines is not None:
reward_timelines = np.array(reward_timelines, dtype=np.object)
states_ndarray = workspace.FetchBlob(state_matrix)
next_states_ndarray = workspace.FetchBlob(next_state_matrix)
tdps = []
for start in range(0, states_ndarray.shape[0], minibatch_size):
end = start + minibatch_size
if end > states_ndarray.shape[0]:
break
tdps.append(
TrainingDataPage(
states=states_ndarray[start:end],
actions=actions_one_hot[start:end],
rewards=rewards[start:end],
next_states=next_states_ndarray[start:end],
not_terminals=not_terminals[start:end],
next_actions=next_actions_one_hot[start:end],
possible_next_actions=possible_next_actions_mask[
start:end],
reward_timelines=reward_timelines[start:end]
if reward_timelines is not None else None,
))
return tdps