Mqcnn rts (#668)

* Getting MQCNN up to date.

Co-authored-by: Jasper Schulz <schjaspe@amazon.de>
Co-authored-by: Aaron Spieler <aspiele@amazon.com>
Co-authored-by: Lorenzo Stella <lorenzostella@gmail.com>

src/gluonts/block/decoder.py

@@ -52,6 +52,7 @@ def hybrid_forward(
  pass
 
 
+# TODO: add support for static variables at some point
 class ForkingMLPDecoder(Seq2SeqDecoder):
  """
  Multilayer perceptron decoder for sequence-to-sequence models.
@@ -76,7 +77,7 @@ def __init__(
  self,
  dec_len: int,
  final_dim: int,
- hidden_dimension_sequence: List[int] = list([]),
+ hidden_dimension_sequence: List[int] = [],
  **kwargs,
  ) -> None:
  super().__init__(**kwargs)
@@ -104,6 +105,7 @@ def __init__(
  )
  self.model.add(layer)
 
+ # TODO: add support for static input at some point
  def hybrid_forward(
  self, F, dynamic_input: Tensor, static_input: Tensor = None
  ) -> Tensor:
@@ -115,11 +117,10 @@ def hybrid_forward(
  F
  A module that can either refer to the Symbol API or the NDArray
  API in MXNet.
-
  dynamic_input
- dynamic_features, shape (batch_size, sequence_length, num_features)
- or (N, T, C).
-
+ dynamic_features, shape (batch_size, sequence_length, num_features) or (N, T, C)
+ where sequence_length is equal to the encoder length, and num_features is equal
+ to channel_seq[-1] for the MQCNN for example.
  static_input
  not used in this decoder.
 

src/gluonts/block/enc2dec.py

@@ -38,7 +38,7 @@ def hybrid_forward(
  F,
  encoder_output_static: Tensor,
  encoder_output_dynamic: Tensor,
- future_features: Tensor,
+ future_features_dynamic: Tensor,
  ) -> Tuple[Tensor, Tensor, Tensor]:
  """
  Parameters
@@ -48,40 +48,36 @@ def hybrid_forward(
  shape (batch_size, num_features) or (N, C)
 
  encoder_output_dynamic
- shape (batch_size, context_length, num_features) or (N, T, C)
+ shape (batch_size, sequence_length, num_features) or (N, T, C)
 
- future_features
- shape (batch_size, prediction_length, num_features) or (N, T, C)
+ future_features_dynamic
+ shape (batch_size, sequence_length, prediction_length, num_features) or (N, T, P, C`)
 
 
  Returns
  -------
  Tensor
  shape (batch_size, num_features) or (N, C)
 
- Tensor
- shape (batch_size, prediction_length, num_features) or (N, T, C)
-
  Tensor
  shape (batch_size, sequence_length, num_features) or (N, T, C)
-
  """
  pass
 
 
 class PassThroughEnc2Dec(Seq2SeqEnc2Dec):
  """
  Simplest class for passing encoder tensors do decoder. Passes through
- tensors.
+ tensors, except that future_features_dynamic is dropped.
  """
 
  def hybrid_forward(
  self,
  F,
  encoder_output_static: Tensor,
  encoder_output_dynamic: Tensor,
- future_features: Tensor,
- ) -> Tuple[Tensor, Tensor, Tensor]:
+ future_features_dynamic: Tensor,
+ ) -> Tuple[Tensor, Tensor]:
  """
  Parameters
  ----------
@@ -90,10 +86,10 @@ def hybrid_forward(
  shape (batch_size, num_features) or (N, C)
 
  encoder_output_dynamic
- shape (batch_size, context_length, num_features) or (N, T, C)
+ shape (batch_size, sequence_length, num_features) or (N, T, C)
 
- future_features
- shape (batch_size, prediction_length, num_features) or (N, T, C)
+ future_features_dynamic
+ shape (batch_size, sequence_length, prediction_length, num_features) or (N, T, P, C`)
 
 
  Returns
@@ -102,10 +98,63 @@ def hybrid_forward(
  shape (batch_size, num_features) or (N, C)
 
  Tensor
- shape (batch_size, prediction_length, num_features) or (N, T, C)
+ shape (batch_size, prediction_length, num_features_02) or (N, T, C)
+ """
+ return encoder_output_static, encoder_output_dynamic
 
- Tensor
+
+class FutureFeatIntegratorEnc2Dec(Seq2SeqEnc2Dec):
+ """
+ Integrates the encoder_ouput_dynamic and future_features_dynamic into one
+ and passes them through as the dynamic input to the decoder.
+ """
+
+ def hybrid_forward(
+ self,
+ F,
+ encoder_output_static: Tensor,
+ encoder_output_dynamic: Tensor,
+ future_features_dynamic: Tensor,
+ ) -> Tuple[Tensor, Tensor]:
+ """
+ Parameters
+ ----------
+
+ encoder_output_static
+ shape (batch_size, num_features) or (N, C)
+
+ encoder_output_dynamic
  shape (batch_size, sequence_length, num_features) or (N, T, C)
 
+ future_features_dynamic
+ shape (batch_size, sequence_length, prediction_length, num_features) or (N, T, P, C`)
+
+
+ Returns
+ -------
+ Tensor
+ shape (batch_size, num_features) or (N, C)
+
+ Tensor
+ shape (batch_size, prediction_length, num_features_02) or (N, T, C)
+
+ Tensor
+ shape (1,)
  """
- return encoder_output_static, encoder_output_dynamic, future_features
+
+ # flatten the last two dimensions:
+ # => (batch_size, encoder_length, decoder_length * num_feature_dynamic)
+ future_features_dynamic = F.reshape(
+ future_features_dynamic, shape=(0, 0, -1)
+ )
+
+ # concatenate output of decoder and future_feat_dynamic covariates:
+ # => (batch_size, encoder_length, num_dec_input_dynamic + num_future_feat_dynamic)
+ total_dec_input_dynamic = F.concat(
+ encoder_output_dynamic, future_features_dynamic, dim=2
+ )
+
+ return (
+ encoder_output_static,
+ total_dec_input_dynamic,
+ )