Feature/transformer refactorisation

darts/models/forecasting/transformer_model.py

@@ -8,6 +8,8 @@
 
 import torch
 import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import Transformer
 
 from darts.logging import get_logger, raise_if, raise_if_not, raise_log
 from darts.models.components import glu_variants, layer_norm_variants
@@ -186,6 +188,7 @@ def __init__(
         self.target_size = output_size
         self.nr_params = nr_params
         self.target_length = self.output_chunk_length
+        self.d_model = d_model
 
         self.encoder = nn.Linear(input_size, d_model)
         self.positional_encoding = _PositionalEncoding(
@@ -279,48 +282,105 @@ def __init__(
             custom_decoder=custom_decoder,
         )
 
-        self.decoder = nn.Linear(
-            d_model, self.target_length * self.target_size * self.nr_params
-        )
-
-    def _create_transformer_inputs(self, data):
-        # '_TimeSeriesSequentialDataset' stores time series in the
-        # (batch_size, input_chunk_length, input_size) format. PyTorch's nn.Transformer
-        # module needs it the (input_chunk_length, batch_size, input_size) format.
-        # Therefore, the first two dimensions need to be swapped.
-        src = data.permute(1, 0, 2)
-        tgt = src[-1:, :, :]
-
-        return src, tgt
+        self.decoder = nn.Linear(d_model, self.target_size * self.nr_params)
 
     @io_processor
     def forward(self, x_in: Tuple):
-        data, _ = x_in
-        # Here we create 'src' and 'tgt', the inputs for the encoder and decoder
-        # side of the Transformer architecture
-        src, tgt = self._create_transformer_inputs(data)
+        """
+        During training (teacher forcing) x_in = tuple(past_target + past_covariates, static_covariates, future_targets)
+        During inference x_in = tuple(past_target + past_covariates, static_covariates)
 
+        '_TimeSeriesSequentialDataset' stores time series in the
+        (batch_size, input_chunk_length, input_size) format. PyTorch's nn.Transformer
+        module needs it the (input_chunk_length, batch_size, input_size) format.
+        Therefore, the first two dimensions need to be swapped.
+        """
+        src = x_in[0].permute(1, 0, 2)
+        pad_size = (0, self.input_size - self.target_size)
+
+        # start token consists only of target series, past covariates are substituted with 0 padding
+        start_token = src[-1:, :, : self.target_size]
+        start_token_padded = F.pad(start_token, pad_size)
+
+        if len(x_in) == 3:
+            tgt = x_in[-1].permute(1, 0, 2)
+            tgt = F.pad(tgt, pad_size)
+            tgt = torch.cat([start_token_padded, tgt], dim=0)
+            return self._prediction_step(src, tgt)[:, :-1, :, :]
+
+        tgt = start_token_padded
+
+        predictions = []
+        for _ in range(self.target_length):
+            pred = self._prediction_step(src, tgt)[:, -1, :, :]
+            predictions.append(pred)
+            tgt = torch.cat(
+                [tgt, F.pad(pred.mean(dim=2).unsqueeze(dim=0), pad_size)],
+                dim=0,
+            )  # take average of samples
+        return torch.stack(predictions, dim=1)
+
+    def _prediction_step(self, src: torch.Tensor, tgt: torch.Tensor):
+        target_length = tgt.shape[0]
+        device, tensor_type = src.device, src.dtype
         # "math.sqrt(self.input_size)" is a normalization factor
         # see section 3.2.1 in 'Attention is All you Need' by Vaswani et al. (2017)
-        src = self.encoder(src) * math.sqrt(self.input_size)
-        src = self.positional_encoding(src)
+        src = self.encoder(src) * math.sqrt(self.d_model)
+        tgt = self.encoder(tgt) * math.sqrt(self.d_model)
 
-        tgt = self.encoder(tgt) * math.sqrt(self.input_size)
+        src = self.positional_encoding(src)
         tgt = self.positional_encoding(tgt)
 
-        x = self.transformer(src=src, tgt=tgt)
+        tgt_mask = Transformer.generate_square_subsequent_mask(
+            target_length, device
+        ).to(dtype=tensor_type)
+
+        x = self.transformer(src=src, tgt=tgt, tgt_mask=tgt_mask)
         out = self.decoder(x)
 
         # Here we change the data format
         # from (1, batch_size, output_chunk_length * output_size)
         # to (batch_size, output_chunk_length, output_size, nr_params)
-        predictions = out[0, :, :]
+        predictions = out.permute(1, 0, 2)
         predictions = predictions.view(
-            -1, self.target_length, self.target_size, self.nr_params
+            -1, target_length, self.target_size, self.nr_params
         )
 
         return predictions
 
+    def training_step(self, train_batch, batch_idx) -> torch.Tensor:
+        """performs the training step"""
+        train_batch = list(train_batch)
+        future_targets = train_batch[-1]
+        train_batch.append(future_targets)
+        return super().training_step(train_batch, batch_idx)
+
+    def _produce_train_output(self, input_batch: Tuple):
+        """
+        Feeds PastCovariatesTorchModel with input and output chunks of a PastCovariatesSequentialDataset for
+        training.
+
+        Parameters:
+        input_batch
+            ``(past_target, past_covariates, static_covariates, future_target)`` during training
+
+            ``(past_target, past_covariates, static_covariates)`` during validation (not teacher forced)
+        """
+
+        past_target, past_covariates, static_covariates = input_batch[:3]
+        # Currently all our PastCovariates models require past target and covariates concatenated
+        inpt = [
+            torch.cat([past_target, past_covariates], dim=2)
+            if past_covariates is not None
+            else past_target,
+            static_covariates,
+        ]
+
+        # add future targets when training (teacher forcing)
+        if len(input_batch) == 4:
+            inpt.append(input_batch[-1])
+        return self(inpt)
+
 
 class TransformerModel(PastCovariatesTorchModel):
     def __init__(
@@ -349,7 +409,7 @@ def __init__(
         The multi-head attention mechanism is highly parallelizable, which makes the transformer architecture
         very suitable to be trained with GPUs.
 
-        The transformer architecture implemented here is based on [1]_.
+        The transformer architecture implemented here is based on [1]_ abd uses teacher forcing [4]_.
 
         This model supports past covariates (known for `input_chunk_length` points before prediction time).
 
@@ -373,7 +433,7 @@ def __init__(
             Fraction of neurons affected by Dropout (default=0.1).
         activation
             The activation function of encoder/decoder intermediate layer, (default='relu').
-            can be one of the glu variant's FeedForward Network (FFN)[2]. A feedforward network is a
+            can be one of the glu variant's FeedForward Network (FFN)[3]. A feedforward network is a
             fully-connected layer with an activation. The glu variant's FeedForward Network are a series
             of FFNs designed to work better with Transformer based models. ["GLU", "Bilinear", "ReGLU", "GEGLU",
             "SwiGLU", "ReLU", "GELU"] or one the pytorch internal activations ["relu", "gelu"]
@@ -452,14 +512,11 @@ def __init__(
             .. highlight:: python
             .. code-block:: python
 
-                def encode_year(idx):
-                    return (idx.year - 1950) / 50
-
                 add_encoders={
                     'cyclic': {'future': ['month']},
                     'datetime_attribute': {'future': ['hour', 'dayofweek']},
                     'position': {'past': ['relative'], 'future': ['relative']},
-                    'custom': {'past': [encode_year]},
+                    'custom': {'past': [lambda idx: (idx.year - 1950) / 50]},
                     'transformer': Scaler()
                 }
             ..
@@ -526,17 +583,7 @@ def encode_year(idx):
         .. [2] Shazeer, Noam, "GLU Variants Improve Transformer", 2020. arVix https://arxiv.org/abs/2002.05202.
         .. [3] T. Kim et al. "Reversible Instance Normalization for Accurate Time-Series Forecasting against
                 Distribution Shift", https://openreview.net/forum?id=cGDAkQo1C0p
-
-        Notes
-        -----
-        Disclaimer:
-        This current implementation is fully functional and can already produce some good predictions. However,
-        it is still limited in how it uses the Transformer architecture because the `tgt` input of
-        `torch.nn.Transformer` is not utilized to its full extent. Currently, we simply pass the last value of the
-        `src` input to `tgt`. To get closer to the way the Transformer is usually used in language models, we
-        should allow the model to consume its own output as part of the `tgt` argument, such that when predicting
-        sequences of values, the input to the `tgt` argument would grow as outputs of the transformer model would be
-        added to it. Of course, the training of the model would have to be adapted accordingly.
+        .. [4] Teacher Forcing PyTorch tutorial: https://github.com/pytorch/examples/tree/main/word_language_model
 
         Examples
         --------