Mqcnn rts

src/gluonts/block/decoder.py

@@ -52,6 +52,7 @@ def hybrid_forward(
  pass
 
 
+# TODO: add support for static variables
 class ForkingMLPDecoder(Seq2SeqDecoder):
  """
  Multilayer perceptron decoder for sequence-to-sequence models.
@@ -104,6 +105,7 @@ def __init__(
  )
  self.model.add(layer)
 
+ # TODO: add support for static input
  def hybrid_forward(
  self, F, dynamic_input: Tensor, static_input: Tensor = None
  ) -> Tensor:

src/gluonts/block/encoder.py

@@ -32,10 +32,6 @@ class Seq2SeqEncoder(nn.HybridBlock):
  a dynamic latent code with the same length as the `target` sequence.
  """
 
- @validated()
- def __init__(self, **kwargs):
- super().__init__(**kwargs)
-
  # noinspection PyMethodOverriding
  def hybrid_forward(
  self,
@@ -77,9 +73,12 @@ def hybrid_forward(
  """
  raise NotImplementedError
 
- @staticmethod
  def _assemble_inputs(
- F, target: Tensor, static_features: Tensor, dynamic_features: Tensor
+ self,
+ F,
+ target: Tensor,
+ static_features: Tensor,
+ dynamic_features: Tensor,
  ) -> Tensor:
  """
  Assemble features from target, static features, and the dynamic
@@ -93,7 +92,7 @@ def _assemble_inputs(
 
  target
  target time series,
- shape (batch_size, sequence_length)
+ shape (batch_size, sequence_length, 1)
 
  static_features
  static features,
@@ -111,7 +110,6 @@ def _assemble_inputs(
  num_static_features + num_dynamic_features + 1)
 
  """
- target = target.expand_dims(axis=-1) # (N, T, 1)
 
  helper_ones = F.ones_like(target) # Ones of (N, T, 1)
  tiled_static_features = F.batch_dot(
@@ -156,7 +154,8 @@ def __init__(
  kernel_size_seq: List[int],
  channels_seq: List[int],
  use_residual: bool = False,
- use_covariates: bool = False,
+ use_static_feat: bool = False,
+ use_dynamic_feat: bool = False,
  **kwargs,
  ) -> None:
  assert all(
@@ -172,7 +171,8 @@ def __init__(
  super().__init__(**kwargs)
 
  self.use_residual = use_residual
- self.use_covariates = use_covariates
+ self.use_static_feat = use_static_feat
+ self.use_dynamic_feat = use_dynamic_feat
  self.cnn = nn.HybridSequential()
 
  it = zip(channels_seq, kernel_size_seq, dilation_seq)
@@ -203,7 +203,7 @@ def hybrid_forward(
 
  target
  target time series,
- shape (batch_size, sequence_length)
+ shape (batch_size, sequence_length, 1)
 
  static_features
  static features,
@@ -224,13 +224,15 @@ def hybrid_forward(
  shape (batch_size, sequence_length, num_dynamic_features)
  """
 
- if self.use_covariates:
- inputs = Seq2SeqEncoder._assemble_inputs(
+ if self.use_dynamic_feat and self.use_static_feat:
+ inputs = self._assemble_inputs(
  F,
  target=target,
  static_features=static_features,
  dynamic_features=dynamic_features,
  )
+ elif self.use_dynamic_feat:
+ inputs = F.concat(target, dynamic_features, dim=2) # (N, T, C)
  else:
  inputs = target
 
@@ -302,7 +304,7 @@ def hybrid_forward(
 
  target
  target time series,
- shape (batch_size, sequence_length)
+ shape (batch_size, sequence_length, 1)
 
  static_features
  static features,
@@ -380,7 +382,7 @@ def hybrid_forward(
  shape (batch_size, sequence_length, num_dynamic_features)
  """
 
- inputs = Seq2SeqEncoder._assemble_inputs(
+ inputs = self._assemble_inputs(
  F, target, static_features, dynamic_features
  )
  static_code = self.model(inputs)
@@ -442,7 +444,7 @@ def hybrid_forward(
 
  target
  target time series,
- shape (batch_size, sequence_length)
+ shape (batch_size, sequence_length, 1)
 
  static_features
  static features,
@@ -462,7 +464,7 @@ def hybrid_forward(
  dynamic code,
  shape (batch_size, sequence_length, num_dynamic_features)
  """
- inputs = Seq2SeqEncoder._assemble_inputs(
+ inputs = self._assemble_inputs(
  F, target, static_features, dynamic_features
  )
  dynamic_code = self.rnn(inputs)

src/gluonts/block/quantile_output.py

@@ -86,9 +86,12 @@ def hybrid_forward(
  Tensor
  weighted sum of the quantile losses, shape N1 x N1 x ... Nk
  """
- y_pred_all = F.split(
- y_pred, axis=-1, num_outputs=self.num_quantiles, squeeze_axis=1
- )
+ if self.num_quantiles > 1:
+ y_pred_all = F.split(
+ y_pred, axis=-1, num_outputs=self.num_quantiles, squeeze_axis=1
+ )
+ else:
+ y_pred_all = [F.squeeze(y_pred, axis=-1)]
 
  qt_loss = []
  for i, y_pred_q in enumerate(y_pred_all):

src/gluonts/model/deepar/_estimator.py

@@ -21,6 +21,7 @@
 # First-party imports
 from gluonts.core.component import DType, validated
 from gluonts.dataset.field_names import FieldName
+from gluonts.dataset.stat import calculate_dataset_statistics
 from gluonts.distribution import DistributionOutput, StudentTOutput
 from gluonts.model.estimator import GluonEstimator
 from gluonts.model.predictor import Predictor, RepresentableBlockPredictor
@@ -146,8 +147,8 @@ def __init__(
  assert num_layers > 0, "The value of `num_layers` should be > 0"
  assert num_cells > 0, "The value of `num_cells` should be > 0"
  assert dropout_rate >= 0, "The value of `dropout_rate` should be >= 0"
- assert (cardinality is not None and use_feat_static_cat) or (
- cardinality is None and not use_feat_static_cat
+ assert (cardinality and use_feat_static_cat) or (
+ not (cardinality or use_feat_static_cat)
  ), "You should set `cardinality` if and only if `use_feat_static_cat=True`"
  assert cardinality is None or all(
  [c > 0 for c in cardinality]
@@ -197,6 +198,16 @@ def __init__(
 
  self.num_parallel_samples = num_parallel_samples
 
+ @classmethod
+ def derive_auto_fields(cls, train_iter):
+ stats = calculate_dataset_statistics(train_iter)
+
+ return {
+ "use_feat_dynamic_real": stats.num_feat_dynamic_real > 0,
+ "use_feat_static_cat": bool(stats.feat_static_cat),
+ "cardinality": [len(cats) for cats in stats.feat_static_cat],
+ }
+
  def create_transformation(self) -> Transformation:
  remove_field_names = [FieldName.FEAT_DYNAMIC_CAT]
  if not self.use_feat_static_real:

src/gluonts/model/estimator.py

@@ -12,7 +12,7 @@
 # permissions and limitations under the License.
 
 # Standard library imports
-from typing import NamedTuple, Optional
+from typing import NamedTuple, Optional, Iterator
 
 # Third-party imports
 import numpy as np
@@ -69,6 +69,21 @@ def train(
  def from_hyperparameters(cls, **hyperparameters):
  return from_hyperparameters(cls, **hyperparameters)
 
+ @classmethod
+ def derive_auto_fields(cls, train_iter):
+ return {}
+
+ @classmethod
+ def from_inputs(cls, train_iter, **params):
+ # auto_params usually include `use_feat_dynamic_real`, `use_feat_static_cat` and `cardinality`
+ auto_params = cls.derive_auto_fields(train_iter)
+ # FIXME: probably params should take precedence over auto_params, since they were deliberately set,
+ # however, on that case this method does not make sense, since if params says `use_feat_dynamic_real`=True
+ # but `auto_params`=False, then this will lead to an error, since the appropriate data does not exist.
+ # This the only context in which this method makes sense is when auto_params take precedence, which could
+ # lead to overwriting of explicit parameters. In this case a warning should be issued.
+ return cls.from_hyperparameters(**auto_params, **params)
+
 
 class DummyEstimator(Estimator):
  """
@@ -126,7 +141,10 @@ def from_hyperparameters(cls, **hyperparameters) -> "GluonEstimator":
  )
 
  try:
- trainer = from_hyperparameters(Trainer, **hyperparameters)
+ trainer = hyperparameters.get("trainer")
+ if not isinstance(trainer, Trainer):
+ trainer = from_hyperparameters(Trainer, **hyperparameters)
+
  return cls(
  **Model(**{**hyperparameters, "trainer": trainer}).__dict__
  )

src/gluonts/model/predictor.py

@@ -156,6 +156,15 @@ def deserialize(
  def from_hyperparameters(cls, **hyperparameters):
  return from_hyperparameters(cls, **hyperparameters)
 
+ @classmethod
+ def derive_auto_fields(cls, train_iter):
+ return {}
+
+ @classmethod
+ def from_inputs(cls, train_iter, **params):
+ auto_params = cls.derive_auto_fields(train_iter)
+ return cls.from_hyperparameters(**auto_params, **params)
+
 
 class RepresentablePredictor(Predictor):
  """