def test_none_reduction():
pred = torch.rand(20, 10)
target = torch.rand(20, 10)
mae = MAE(reduction="none")(pred, target)
assert mae.size() == pred.size(
), "dimension should not change if reduction is none"
def from_dataset(
cls,
dataset: TimeSeriesDataSet,
allowed_encoder_known_variable_names: List[str] = None,
**kwargs,
):
"""
Create model from dataset.
Args:
dataset: timeseries dataset
allowed_encoder_known_variable_names: List of known variables that are allowed in encoder, defaults to all
**kwargs: additional arguments such as hyperparameters for model (see ``__init__()``)
Returns:
Recurrent network
"""
new_kwargs = copy(kwargs)
new_kwargs.update(
cls.deduce_default_output_parameters(dataset=dataset,
kwargs=kwargs,
default_loss=MAE()))
assert not isinstance(dataset.target_normalizer, NaNLabelEncoder) and (
not isinstance(dataset.target_normalizer, MultiNormalizer) or all([
not isinstance(normalizer, NaNLabelEncoder)
for normalizer in dataset.target_normalizer
])
), "target(s) should be continuous - categorical targets are not supported" # todo: remove this restriction
return super().from_dataset(dataset,
allowed_encoder_known_variable_names=
allowed_encoder_known_variable_names,
**new_kwargs)
def test_aggregation_metric(decoder_lengths, y):
y_pred = torch.tensor([[0.0, 2.0], [4.0, 3.0]])
if (decoder_lengths != y_pred.size(-1)).any():
y_packed = rnn.pack_padded_sequence(y, lengths=decoder_lengths, batch_first=True, enforce_sorted=False)
else:
y_packed = y
# metric
metric = AggregationMetric(MAE())
res = metric(y_pred, y_packed)
if (decoder_lengths == y_pred.size(-1)).all() and y.ndim == 2:
assert torch.isclose(res, (y.mean(0) - y_pred.mean(0)).abs().mean())
def test_composite_metric():
metric1 = SMAPE()
metric2 = MAE()
combined_metric = 1.0 * (0.3 * metric1 + 2.0 * metric2 + metric1)
assert isinstance(combined_metric, CompositeMetric), "combined metric should be composite metric"
# test repr()
repr(combined_metric)
# test results
y = torch.normal(0, 1, (10, 20)).abs()
y_pred = torch.normal(0, 1, (10, 20)).abs()
res1 = metric1(y_pred, y)
res2 = metric2(y_pred, y)
combined_res = combined_metric(y_pred, y)
assert torch.isclose(combined_res, res1 * 0.3 + res2 * 2.0 + res1)
# test quantiles and prediction
combined_metric.to_prediction(y_pred)
combined_metric.to_quantiles(y_pred)
def __init__(
self,
cell_type: str = "LSTM",
hidden_size: int = 10,
rnn_layers: int = 2,
dropout: float = 0.1,
static_categoricals: List[str] = [],
static_reals: List[str] = [],
time_varying_categoricals_encoder: List[str] = [],
time_varying_categoricals_decoder: List[str] = [],
categorical_groups: Dict[str, List[str]] = {},
time_varying_reals_encoder: List[str] = [],
time_varying_reals_decoder: List[str] = [],
embedding_sizes: Dict[str, Tuple[int, int]] = {},
embedding_paddings: List[str] = [],
embedding_labels: Dict[str, np.ndarray] = {},
x_reals: List[str] = [],
x_categoricals: List[str] = [],
output_size: Union[int, List[int]] = 1,
target: Union[str, List[str]] = None,
target_lags: Dict[str, List[int]] = {},
loss: MultiHorizonMetric = None,
logging_metrics: nn.ModuleList = None,
**kwargs,
):
"""
Recurrent Network.
Simple LSTM or GRU layer followed by output layer
Args:
cell_type (str, optional): Recurrent cell type ["LSTM", "GRU"]. Defaults to "LSTM".
hidden_size (int, optional): hidden recurrent size - the most important hyperparameter along with
``rnn_layers``. Defaults to 10.
rnn_layers (int, optional): Number of RNN layers - important hyperparameter. Defaults to 2.
dropout (float, optional): Dropout in RNN layers. Defaults to 0.1.
static_categoricals: integer of positions of static categorical variables
static_reals: integer of positions of static continuous variables
time_varying_categoricals_encoder: integer of positions of categorical variables for encoder
time_varying_categoricals_decoder: integer of positions of categorical variables for decoder
time_varying_reals_encoder: integer of positions of continuous variables for encoder
time_varying_reals_decoder: integer of positions of continuous variables for decoder
categorical_groups: dictionary where values
are list of categorical variables that are forming together a new categorical
variable which is the key in the dictionary
x_reals: order of continuous variables in tensor passed to forward function
x_categoricals: order of categorical variables in tensor passed to forward function
embedding_sizes: dictionary mapping (string) indices to tuple of number of categorical classes and
embedding size
embedding_paddings: list of indices for embeddings which transform the zero's embedding to a zero vector
embedding_labels: dictionary mapping (string) indices to list of categorical labels
output_size (Union[int, List[int]], optional): number of outputs (e.g. number of quantiles for
QuantileLoss and one target or list of output sizes).
target (str, optional): Target variable or list of target variables. Defaults to None.
target_lags (Dict[str, Dict[str, int]]): dictionary of target names mapped to list of time steps by
which the variable should be lagged.
Lags can be useful to indicate seasonality to the models. If you know the seasonalit(ies) of your data,
add at least the target variables with the corresponding lags to improve performance.
Defaults to no lags, i.e. an empty dictionary.
loss (MultiHorizonMetric, optional): loss: loss function taking prediction and targets.
logging_metrics (nn.ModuleList, optional): Metrics to log during training.
Defaults to nn.ModuleList([SMAPE(), MAE(), RMSE(), MAPE(), MASE()]).
"""
if loss is None:
loss = MAE()
if logging_metrics is None:
logging_metrics = nn.ModuleList(
[SMAPE(), MAE(), RMSE(),
MAPE(), MASE()])
self.save_hyperparameters()
# store loss function separately as it is a module
super().__init__(loss=loss, logging_metrics=logging_metrics, **kwargs)
self.embeddings = MultiEmbedding(
embedding_sizes=embedding_sizes,
embedding_paddings=embedding_paddings,
categorical_groups=categorical_groups,
x_categoricals=x_categoricals,
)
lagged_target_names = [
l for lags in target_lags.values() for l in lags
]
assert set(self.encoder_variables) - set(
to_list(target)
) - set(lagged_target_names) == set(self.decoder_variables) - set(
lagged_target_names
), "Encoder and decoder variables have to be the same apart from target variable"
for targeti in to_list(target):
assert (
targeti in time_varying_reals_encoder
), f"target {targeti} has to be real" # todo: remove this restriction
assert (
isinstance(target, str) and isinstance(loss, MultiHorizonMetric)
) or (
isinstance(target, (list, tuple)) and isinstance(loss, MultiLoss)
and len(loss) == len(target)
), "number of targets should be equivalent to number of loss metrics"
rnn_class = get_rnn(cell_type)
cont_size = len(self.reals)
cat_size = sum(
[size[1] for size in self.hparams.embedding_sizes.values()])
input_size = cont_size + cat_size
self.rnn = rnn_class(
input_size=input_size,
hidden_size=self.hparams.hidden_size,
num_layers=self.hparams.rnn_layers,
dropout=self.hparams.dropout if self.hparams.rnn_layers > 1 else 0,
batch_first=True,
)
# add linear layers for argument projects
if isinstance(target, str): # single target
self.output_projector = nn.Linear(self.hparams.hidden_size,
self.hparams.output_size)
assert not isinstance(
self.loss, QuantileLoss
), "QuantileLoss does not work with recurrent network"
else: # multi target
self.output_projector = nn.ModuleList([
nn.Linear(self.hparams.hidden_size, size)
for size in self.hparams.output_size
])
for l in self.loss:
assert not isinstance(
l, QuantileLoss
), "QuantileLoss does not work with recurrent network"
def __init__(
self,
stack_types: List[str] = ["trend", "seasonality"],
num_blocks=[3, 3],
num_block_layers=[3, 3],
widths=[32, 512],
sharing: List[int] = [True, True],
expansion_coefficient_lengths: List[int] = [3, 7],
prediction_length: int = 1,
context_length: int = 1,
dropout: float = 0.1,
learning_rate: float = 1e-2,
log_interval: int = -1,
log_gradient_flow: bool = False,
log_val_interval: int = None,
weight_decay: float = 1e-3,
loss: MultiHorizonMetric = None,
reduce_on_plateau_patience: int = 1000,
backcast_loss_ratio: float = 0.0,
logging_metrics: nn.ModuleList = None,
**kwargs,
):
"""
Initialize NBeats Model - use its :py:meth:`~from_dataset` method if possible.
Based on the article
`N-BEATS: Neural basis expansion analysis for interpretable time series
forecasting <http://arxiv.org/abs/1905.10437>`_. The network has (if used as ensemble) outperformed all
other methods
including ensembles of traditional statical methods in the M4 competition. The M4 competition is arguably
the most
important benchmark for univariate time series forecasting.
Args:
stack_types: One of the following values: “generic”, “seasonality" or “trend". A list of strings
of length 1 or ‘num_stacks’. Default and recommended value
for generic mode: [“generic”] Recommended value for interpretable mode: [“trend”,”seasonality”]
num_blocks: The number of blocks per stack. A list of ints of length 1 or ‘num_stacks’.
Default and recommended value for generic mode: [1] Recommended value for interpretable mode: [3]
num_block_layers: Number of fully connected layers with ReLu activation per block. A list of ints of length
1 or ‘num_stacks’.
Default and recommended value for generic mode: [4] Recommended value for interpretable mode: [4]
width: Widths of the fully connected layers with ReLu activation in the blocks.
A list of ints of length 1 or ‘num_stacks’. Default and recommended value for generic mode: [512]
Recommended value for interpretable mode: [256, 2048]
sharing: Whether the weights are shared with the other blocks per stack.
A list of ints of length 1 or ‘num_stacks’. Default and recommended value for generic mode: [False]
Recommended value for interpretable mode: [True]
expansion_coefficient_length: If the type is “G” (generic), then the length of the expansion
coefficient.
If type is “T” (trend), then it corresponds to the degree of the polynomial. If the type is “S”
(seasonal) then this is the minimum period allowed, e.g. 2 for changes every timestep.
A list of ints of length 1 or ‘num_stacks’. Default value for generic mode: [32] Recommended value for
interpretable mode: [3]
prediction_length: Length of the prediction. Also known as 'horizon'.
context_length: Number of time units that condition the predictions. Also known as 'lookback period'.
Should be between 1-10 times the prediction length.
backcast_loss_ratio: weight of backcast in comparison to forecast when calculating the loss.
A weight of 1.0 means that forecast and backcast loss is weighted the same (regardless of backcast and
forecast lengths). Defaults to 0.0, i.e. no weight.
loss: loss to optimize. Defaults to MASE().
log_gradient_flow: if to log gradient flow, this takes time and should be only done to diagnose training
failures
reduce_on_plateau_patience (int): patience after which learning rate is reduced by a factor of 10
logging_metrics (nn.ModuleList[MultiHorizonMetric]): list of metrics that are logged during training.
Defaults to nn.ModuleList([SMAPE(), MAE(), RMSE(), MAPE(), MASE()])
**kwargs: additional arguments to :py:class:`~BaseModel`.
"""
if logging_metrics is None:
logging_metrics = nn.ModuleList(
[SMAPE(), MAE(), RMSE(),
MAPE(), MASE()])
if loss is None:
loss = MASE()
self.save_hyperparameters()
super().__init__(loss=loss, logging_metrics=logging_metrics, **kwargs)
# setup stacks
self.net_blocks = nn.ModuleList()
for stack_id, stack_type in enumerate(stack_types):
for _ in range(num_blocks[stack_id]):
if stack_type == "generic":
net_block = NBEATSGenericBlock(
units=self.hparams.widths[stack_id],
thetas_dim=self.hparams.
expansion_coefficient_lengths[stack_id],
num_block_layers=self.hparams.
num_block_layers[stack_id],
backcast_length=context_length,
forecast_length=prediction_length,
dropout=self.hparams.dropout,
)
elif stack_type == "seasonality":
net_block = NBEATSSeasonalBlock(
units=self.hparams.widths[stack_id],
num_block_layers=self.hparams.
num_block_layers[stack_id],
backcast_length=context_length,
forecast_length=prediction_length,
min_period=self.hparams.
expansion_coefficient_lengths[stack_id],
dropout=self.hparams.dropout,
)
elif stack_type == "trend":
net_block = NBEATSTrendBlock(
units=self.hparams.widths[stack_id],
thetas_dim=self.hparams.
expansion_coefficient_lengths[stack_id],
num_block_layers=self.hparams.
num_block_layers[stack_id],
backcast_length=context_length,
forecast_length=prediction_length,
dropout=self.hparams.dropout,
)
else:
raise ValueError(f"Unknown stack type {stack_type}")
self.net_blocks.append(net_block)
def __init__(
self,
hidden_size: int = 16,
lstm_layers: int = 1,
dropout: float = 0.1,
output_size: Union[int, List[int]] = 7,
loss: MultiHorizonMetric = None,
attention_head_size: int = 4,
max_encoder_length: int = 10,
static_categoricals: List[str] = [],
static_reals: List[str] = [],
time_varying_categoricals_encoder: List[str] = [],
time_varying_categoricals_decoder: List[str] = [],
categorical_groups: Dict[str, List[str]] = {},
time_varying_reals_encoder: List[str] = [],
time_varying_reals_decoder: List[str] = [],
x_reals: List[str] = [],
x_categoricals: List[str] = [],
hidden_continuous_size: int = 8,
hidden_continuous_sizes: Dict[str, int] = {},
embedding_sizes: Dict[str, Tuple[int, int]] = {},
embedding_paddings: List[str] = [],
embedding_labels: Dict[str, np.ndarray] = {},
learning_rate: float = 1e-3,
log_interval: Union[int, float] = -1,
log_val_interval: Union[int, float] = None,
log_gradient_flow: bool = False,
reduce_on_plateau_patience: int = 1000,
monotone_constaints: Dict[str, int] = {},
share_single_variable_networks: bool = False,
logging_metrics: nn.ModuleList = None,
**kwargs,
):
"""
Temporal Fusion Transformer for forecasting timeseries - use its :py:meth:`~from_dataset` method if possible.
Implementation of the article
`Temporal Fusion Transformers for Interpretable Multi-horizon Time Series
Forecasting <https://arxiv.org/pdf/1912.09363.pdf>`_. The network outperforms DeepAR by Amazon by 36-69%
in benchmarks.
Enhancements compared to the original implementation (apart from capabilities added through base model
such as monotone constraints):
* static variables can be continuous
* multiple categorical variables can be summarized with an EmbeddingBag
* variable encoder and decoder length by sample
* categorical embeddings are not transformed by variable selection network (because it is a redundant operation)
* variable dimension in variable selection network are scaled up via linear interpolation to reduce
number of parameters
* non-linear variable processing in variable selection network can be shared among decoder and encoder
(not shared by default)
Tune its hyperparameters with
:py:func:`~pytorch_forecasting.models.temporal_fusion_transformer.tuning.optimize_hyperparameters`.
Args:
hidden_size: hidden size of network which is its main hyperparameter and can range from 8 to 512
lstm_layers: number of LSTM layers (2 is mostly optimal)
dropout: dropout rate
output_size: number of outputs (e.g. number of quantiles for QuantileLoss and one target or list
of output sizes).
loss: loss function taking prediction and targets
attention_head_size: number of attention heads (4 is a good default)
max_encoder_length: length to encode (can be far longer than the decoder length but does not have to be)
static_categoricals: names of static categorical variables
static_reals: names of static continuous variables
time_varying_categoricals_encoder: names of categorical variables for encoder
time_varying_categoricals_decoder: names of categorical variables for decoder
time_varying_reals_encoder: names of continuous variables for encoder
time_varying_reals_decoder: names of continuous variables for decoder
categorical_groups: dictionary where values
are list of categorical variables that are forming together a new categorical
variable which is the key in the dictionary
x_reals: order of continuous variables in tensor passed to forward function
x_categoricals: order of categorical variables in tensor passed to forward function
hidden_continuous_size: default for hidden size for processing continous variables (similar to categorical
embedding size)
hidden_continuous_sizes: dictionary mapping continuous input indices to sizes for variable selection
(fallback to hidden_continuous_size if index is not in dictionary)
embedding_sizes: dictionary mapping (string) indices to tuple of number of categorical classes and
embedding size
embedding_paddings: list of indices for embeddings which transform the zero's embedding to a zero vector
embedding_labels: dictionary mapping (string) indices to list of categorical labels
learning_rate: learning rate
log_interval: log predictions every x batches, do not log if 0 or less, log interpretation if > 0. If < 1.0
, will log multiple entries per batch. Defaults to -1.
log_val_interval: frequency with which to log validation set metrics, defaults to log_interval
log_gradient_flow: if to log gradient flow, this takes time and should be only done to diagnose training
failures
reduce_on_plateau_patience (int): patience after which learning rate is reduced by a factor of 10
monotone_constaints (Dict[str, int]): dictionary of monotonicity constraints for continuous decoder
variables mapping
position (e.g. ``"0"`` for first position) to constraint (``-1`` for negative and ``+1`` for positive,
larger numbers add more weight to the constraint vs. the loss but are usually not necessary).
This constraint significantly slows down training. Defaults to {}.
share_single_variable_networks (bool): if to share the single variable networks between the encoder and
decoder. Defaults to False.
logging_metrics (nn.ModuleList[LightningMetric]): list of metrics that are logged during training.
Defaults to nn.ModuleList([SMAPE(), MAE(), RMSE(), MAPE()]).
**kwargs: additional arguments to :py:class:`~BaseModel`.
"""
if logging_metrics is None:
logging_metrics = nn.ModuleList([SMAPE(), MAE(), RMSE(), MAPE()])
if loss is None:
loss = QuantileLoss()
self.save_hyperparameters()
# store loss function separately as it is a module
assert isinstance(
loss,
LightningMetric), "Loss has to be a PyTorch Lightning `Metric`"
super().__init__(loss=loss, logging_metrics=logging_metrics, **kwargs)
# processing inputs
# embeddings
self.input_embeddings = MultiEmbedding(
embedding_sizes=self.hparams.embedding_sizes,
categorical_groups=self.hparams.categorical_groups,
embedding_paddings=self.hparams.embedding_paddings,
x_categoricals=self.hparams.x_categoricals,
max_embedding_size=self.hparams.hidden_size,
)
# continuous variable processing
self.prescalers = nn.ModuleDict({
name: nn.Linear(
1,
self.hparams.hidden_continuous_sizes.get(
name, self.hparams.hidden_continuous_size))
for name in self.reals
})
# variable selection
# variable selection for static variables
static_input_sizes = {
name: self.hparams.embedding_sizes[name][1]
for name in self.hparams.static_categoricals
}
static_input_sizes.update({
name: self.hparams.hidden_continuous_sizes.get(
name, self.hparams.hidden_continuous_size)
for name in self.hparams.static_reals
})
self.static_variable_selection = VariableSelectionNetwork(
input_sizes=static_input_sizes,
hidden_size=self.hparams.hidden_size,
input_embedding_flags={
name: True
for name in self.hparams.static_categoricals
},
dropout=self.hparams.dropout,
prescalers=self.prescalers,
)
# variable selection for encoder and decoder
encoder_input_sizes = {
name: self.hparams.embedding_sizes[name][1]
for name in self.hparams.time_varying_categoricals_encoder
}
encoder_input_sizes.update({
name: self.hparams.hidden_continuous_sizes.get(
name, self.hparams.hidden_continuous_size)
for name in self.hparams.time_varying_reals_encoder
})
decoder_input_sizes = {
name: self.hparams.embedding_sizes[name][1]
for name in self.hparams.time_varying_categoricals_decoder
}
decoder_input_sizes.update({
name: self.hparams.hidden_continuous_sizes.get(
name, self.hparams.hidden_continuous_size)
for name in self.hparams.time_varying_reals_decoder
})
# create single variable grns that are shared across decoder and encoder
if self.hparams.share_single_variable_networks:
self.shared_single_variable_grns = nn.ModuleDict()
for name, input_size in encoder_input_sizes.items():
self.shared_single_variable_grns[name] = GatedResidualNetwork(
input_size,
min(input_size, self.hparams.hidden_size),
self.hparams.hidden_size,
self.hparams.dropout,
)
for name, input_size in decoder_input_sizes.items():
if name not in self.shared_single_variable_grns:
self.shared_single_variable_grns[
name] = GatedResidualNetwork(
input_size,
min(input_size, self.hparams.hidden_size),
self.hparams.hidden_size,
self.hparams.dropout,
)
self.encoder_variable_selection = VariableSelectionNetwork(
input_sizes=encoder_input_sizes,
hidden_size=self.hparams.hidden_size,
input_embedding_flags={
name: True
for name in self.hparams.time_varying_categoricals_encoder
},
dropout=self.hparams.dropout,
context_size=self.hparams.hidden_size,
prescalers=self.prescalers,
single_variable_grns={}
if not self.hparams.share_single_variable_networks else
self.shared_single_variable_grns,
)
self.decoder_variable_selection = VariableSelectionNetwork(
input_sizes=decoder_input_sizes,
hidden_size=self.hparams.hidden_size,
input_embedding_flags={
name: True
for name in self.hparams.time_varying_categoricals_decoder
},
dropout=self.hparams.dropout,
context_size=self.hparams.hidden_size,
prescalers=self.prescalers,
single_variable_grns={}
if not self.hparams.share_single_variable_networks else
self.shared_single_variable_grns,
)
# static encoders
# for variable selection
self.static_context_variable_selection = GatedResidualNetwork(
input_size=self.hparams.hidden_size,
hidden_size=self.hparams.hidden_size,
output_size=self.hparams.hidden_size,
dropout=self.hparams.dropout,
)
# for hidden state of the lstm
self.static_context_initial_hidden_lstm = GatedResidualNetwork(
input_size=self.hparams.hidden_size,
hidden_size=self.hparams.hidden_size,
output_size=self.hparams.hidden_size,
dropout=self.hparams.dropout,
)
# for cell state of the lstm
self.static_context_initial_cell_lstm = GatedResidualNetwork(
input_size=self.hparams.hidden_size,
hidden_size=self.hparams.hidden_size,
output_size=self.hparams.hidden_size,
dropout=self.hparams.dropout,
)
# for post lstm static enrichment
self.static_context_enrichment = GatedResidualNetwork(
self.hparams.hidden_size, self.hparams.hidden_size,
self.hparams.hidden_size, self.hparams.dropout)
# lstm encoder (history) and decoder (future) for local processing
self.lstm_encoder = LSTM(
input_size=self.hparams.hidden_size,
hidden_size=self.hparams.hidden_size,
num_layers=self.hparams.lstm_layers,
dropout=self.hparams.dropout
if self.hparams.lstm_layers > 1 else 0,
batch_first=True,
)
self.lstm_decoder = LSTM(
input_size=self.hparams.hidden_size,
hidden_size=self.hparams.hidden_size,
num_layers=self.hparams.lstm_layers,
dropout=self.hparams.dropout
if self.hparams.lstm_layers > 1 else 0,
batch_first=True,
)
# skip connection for lstm
self.post_lstm_gate_encoder = GatedLinearUnit(
self.hparams.hidden_size, dropout=self.hparams.dropout)
self.post_lstm_gate_decoder = self.post_lstm_gate_encoder
# self.post_lstm_gate_decoder = GatedLinearUnit(self.hparams.hidden_size, dropout=self.hparams.dropout)
self.post_lstm_add_norm_encoder = AddNorm(self.hparams.hidden_size,
trainable_add=False)
# self.post_lstm_add_norm_decoder = AddNorm(self.hparams.hidden_size, trainable_add=True)
self.post_lstm_add_norm_decoder = self.post_lstm_add_norm_encoder
# static enrichment and processing past LSTM
self.static_enrichment = GatedResidualNetwork(
input_size=self.hparams.hidden_size,
hidden_size=self.hparams.hidden_size,
output_size=self.hparams.hidden_size,
dropout=self.hparams.dropout,
context_size=self.hparams.hidden_size,
)
# attention for long-range processing
self.multihead_attn = InterpretableMultiHeadAttention(
d_model=self.hparams.hidden_size,
n_head=self.hparams.attention_head_size,
dropout=self.hparams.dropout)
self.post_attn_gate_norm = GateAddNorm(self.hparams.hidden_size,
dropout=self.hparams.dropout,
trainable_add=False)
self.pos_wise_ff = GatedResidualNetwork(self.hparams.hidden_size,
self.hparams.hidden_size,
self.hparams.hidden_size,
dropout=self.hparams.dropout)
# output processing -> no dropout at this late stage
self.pre_output_gate_norm = GateAddNorm(self.hparams.hidden_size,
dropout=None,
trainable_add=False)
if self.n_targets > 1: # if to run with multiple targets
self.output_layer = nn.ModuleList([
nn.Linear(self.hparams.hidden_size, output_size)
for output_size in self.hparams.output_size
])
else:
self.output_layer = nn.Linear(self.hparams.hidden_size,
self.hparams.output_size)
def __init__(
self,
cell_type: str = "LSTM",
hidden_size: int = 10,
rnn_layers: int = 2,
dropout: float = 0.1,
static_categoricals: List[str] = [],
static_reals: List[str] = [],
time_varying_categoricals_encoder: List[str] = [],
time_varying_categoricals_decoder: List[str] = [],
categorical_groups: Dict[str, List[str]] = {},
time_varying_reals_encoder: List[str] = [],
time_varying_reals_decoder: List[str] = [],
embedding_sizes: Dict[str, Tuple[int, int]] = {},
embedding_paddings: List[str] = [],
embedding_labels: Dict[str, np.ndarray] = {},
x_reals: List[str] = [],
x_categoricals: List[str] = [],
n_validation_samples: int = None,
n_plotting_samples: int = None,
target: Union[str, List[str]] = None,
loss: DistributionLoss = None,
logging_metrics: nn.ModuleList = None,
**kwargs,
):
"""
DeepAR Network.
The code is based on the article `DeepAR: Probabilistic forecasting with autoregressive recurrent networks
<https://www.sciencedirect.com/science/article/pii/S0169207019301888>`_.
Args:
cell_type (str, optional): Recurrent cell type ["LSTM", "GRU"]. Defaults to "LSTM".
hidden_size (int, optional): hidden recurrent size - the most important hyperparameter along with
``rnn_layers``. Defaults to 10.
rnn_layers (int, optional): Number of RNN layers - important hyperparameter. Defaults to 2.
dropout (float, optional): Dropout in RNN layers. Defaults to 0.1.
static_categoricals: integer of positions of static categorical variables
static_reals: integer of positions of static continuous variables
time_varying_categoricals_encoder: integer of positions of categorical variables for encoder
time_varying_categoricals_decoder: integer of positions of categorical variables for decoder
time_varying_reals_encoder: integer of positions of continuous variables for encoder
time_varying_reals_decoder: integer of positions of continuous variables for decoder
categorical_groups: dictionary where values
are list of categorical variables that are forming together a new categorical
variable which is the key in the dictionary
x_reals: order of continuous variables in tensor passed to forward function
x_categoricals: order of categorical variables in tensor passed to forward function
embedding_sizes: dictionary mapping (string) indices to tuple of number of categorical classes and
embedding size
embedding_paddings: list of indices for embeddings which transform the zero's embedding to a zero vector
embedding_labels: dictionary mapping (string) indices to list of categorical labels
n_validation_samples (int, optional): Number of samples to use for calculating validation metrics.
Defaults to None, i.e. no sampling at validation stage and using "mean" of distribution for logging
metrics calculation.
n_plotting_samples (int, optional): Number of samples to generate for plotting predictions
during training. Defaults to ``n_validation_samples`` if not None or 100 otherwise.
target (str, optional): Target variable or list of target variables. Defaults to None.
loss (DistributionLoss, optional): Distribution loss function. Keep in mind that each distribution
loss function might have specific requirements for target normalization.
Defaults to :py:class:`~pytorch_forecasting.metrics.NormalDistributionLoss`.
logging_metrics (nn.ModuleList, optional): Metrics to log during training.
Defaults to nn.ModuleList([SMAPE(), MAE(), RMSE(), MAPE(), MASE()]).
"""
if loss is None:
loss = NormalDistributionLoss()
if logging_metrics is None:
logging_metrics = nn.ModuleList(
[SMAPE(), MAE(), RMSE(),
MAPE(), MASE()])
if n_plotting_samples is None:
if n_validation_samples is None:
n_plotting_samples = n_validation_samples
else:
n_plotting_samples = 100
self.save_hyperparameters()
# store loss function separately as it is a module
super().__init__(loss=loss, logging_metrics=logging_metrics, **kwargs)
self.embeddings = MultiEmbedding(
embedding_sizes=embedding_sizes,
embedding_paddings=embedding_paddings,
categorical_groups=categorical_groups,
x_categoricals=x_categoricals,
)
assert set(self.encoder_variables) - set(to_list(target)) == set(
self.decoder_variables
), "Encoder and decoder variables have to be the same apart from target variable"
for targeti in to_list(target):
assert (
targeti in time_varying_reals_encoder
), f"target {targeti} has to be real" # todo: remove this restriction
assert (
isinstance(target, str) and isinstance(loss, DistributionLoss)
) or (
isinstance(target,
(list, tuple)) and isinstance(loss, MultiLoss) and
len(loss)
== len(target)
), "number of targets should be equivalent to number of loss metrics"
time_series_rnn = get_cell(cell_type)
self.rnn = time_series_rnn(
input_size=self.input_size,
hidden_size=self.hparams.hidden_size,
num_layers=self.hparams.rnn_layers,
dropout=self.hparams.dropout if self.hparams.rnn_layers > 1 else 0,
batch_first=True,
)
# add linear layers for argument projects
if isinstance(loss, MultiLoss): # multi target
self.distribution_projector = nn.ModuleList([
nn.Linear(self.hparams.hidden_size, len(args))
for args in self.loss.distribution_arguments
])
else:
self.distribution_projector = nn.Linear(
self.hparams.hidden_size,
len(self.loss.distribution_arguments))
def __init__(
self,
activation_class: str = "ReLU",
hidden_size: int = 300,
n_hidden_layers: int = 3,
dropout: float = 0.1,
norm: bool = True,
static_categoricals: List[str] = [],
static_reals: List[str] = [],
time_varying_categoricals_encoder: List[str] = [],
time_varying_categoricals_decoder: List[str] = [],
categorical_groups: Dict[str, List[str]] = {},
time_varying_reals_encoder: List[str] = [],
time_varying_reals_decoder: List[str] = [],
embedding_sizes: Dict[str, Tuple[int, int]] = {},
embedding_paddings: List[str] = [],
embedding_labels: Dict[str, np.ndarray] = {},
x_reals: List[str] = [],
x_categoricals: List[str] = [],
output_size: Union[int, List[int]] = 1,
target: Union[str, List[str]] = None,
loss: MultiHorizonMetric = None,
logging_metrics: nn.ModuleList = None,
**kwargs,
):
"""
Args:
activation_class (str, optional): PyTorch activation class. Defaults to "ReLU".
hidden_size (int, optional): hidden recurrent size - the most important hyperparameter along with
``n_hidden_layers``. Defaults to 10.
n_hidden_layers (int, optional): Number of hidden layers - important hyperparameter. Defaults to 2.
dropout (float, optional): Dropout. Defaults to 0.1.
norm (bool, optional): if to use normalization in the MLP. Defaults to True.
static_categoricals: integer of positions of static categorical variables
static_reals: integer of positions of static continuous variables
time_varying_categoricals_encoder: integer of positions of categorical variables for encoder
time_varying_categoricals_decoder: integer of positions of categorical variables for decoder
time_varying_reals_encoder: integer of positions of continuous variables for encoder
time_varying_reals_decoder: integer of positions of continuous variables for decoder
categorical_groups: dictionary where values
are list of categorical variables that are forming together a new categorical
variable which is the key in the dictionary
x_reals: order of continuous variables in tensor passed to forward function
x_categoricals: order of categorical variables in tensor passed to forward function
embedding_sizes: dictionary mapping (string) indices to tuple of number of categorical classes and
embedding size
embedding_paddings: list of indices for embeddings which transform the zero's embedding to a zero vector
embedding_labels: dictionary mapping (string) indices to list of categorical labels
output_size (Union[int, List[int]], optional): number of outputs (e.g. number of quantiles for
QuantileLoss and one target or list of output sizes).
target (str, optional): Target variable or list of target variables. Defaults to None.
loss (MultiHorizonMetric, optional): loss: loss function taking prediction and targets.
Defaults to QuantileLoss.
logging_metrics (nn.ModuleList, optional): Metrics to log during training.
Defaults to nn.ModuleList([SMAPE(), MAE(), RMSE(), MAPE(), MASE()]).
"""
if loss is None:
loss = QuantileLoss()
if logging_metrics is None:
logging_metrics = nn.ModuleList(
[SMAPE(), MAE(), RMSE(),
MAPE(), MASE()])
self.save_hyperparameters()
# store loss function separately as it is a module
super().__init__(loss=loss, logging_metrics=logging_metrics, **kwargs)
self.input_embeddings = MultiEmbedding(
embedding_sizes={
name: val
for name, val in embedding_sizes.items()
if name in self.decoder_variables + self.static_variables
},
embedding_paddings=embedding_paddings,
categorical_groups=categorical_groups,
x_categoricals=x_categoricals,
)
# define network
if isinstance(self.hparams.output_size, int):
mlp_output_size = self.hparams.output_size
else:
mlp_output_size = sum(self.hparams.output_size)
cont_size = len(self.decoder_reals_positions)
cat_size = sum(self.input_embeddings.output_size.values())
input_size = cont_size + cat_size
self.mlp = FullyConnectedModule(
dropout=dropout,
norm=self.hparams.norm,
activation_class=getattr(nn, self.hparams.activation_class),
input_size=input_size,
output_size=mlp_output_size,
hidden_size=self.hparams.hidden_size,
n_hidden_layers=self.hparams.n_hidden_layers,
)
and len(dataset.time_varying_unknown_categoricals) == 0
and len(dataset.static_categoricals) == 0
and len(dataset.static_reals) == 0
and len(dataset.time_varying_unknown_reals) == len(
dataset.target_names
) # Expect as as many unknown reals as targets
), "Only covariate should be in 'time_varying_unknown_reals'"
return super().from_dataset(dataset, **new_kwargs)
model = FullyConnectedMultiTargetModel.from_dataset(
multi_target_dataset,
hidden_size=10,
n_hidden_layers=2,
loss=MultiLoss(metrics=[MAE(), SMAPE()], weights=[2.0, 1.0]),
)
model.summarize("full")
model.hparams
# %% [markdown]
# Now, let's pass some data through our model and calculate the loss.
# %%
out = model(x)
out
# %%
y_hat = model.transform_output(
out
) # the model's transform_output method re-scales/de-normalizes the predictions to into the real target space
#%%
"""
Evaluating the trained model
"""
from pytorch_forecasting.metrics import MAE
import torch
# load the best model according to the validation loss (given that
# we use early stopping, this is not necessarily the last epoch)
best_model_path = trainer.checkpoint_callback.best_model_path
best_tft = TemporalFusionTransformer.load_from_checkpoint(best_model_path)
# calculate mean absolute error on validation set
actuals = torch.cat([y for x, y in iter(val_dataloader)])
predictions = best_tft.predict(val_dataloader)
MAE(predictions, actuals)
from pytorch_forecasting.metrics import SMAPE
raw_predictions = best_tft.predict(val_dataloader, mode="raw")
# calculate metric by which to display
predictions, x = best_tft.predict(val_dataloader, return_x=True)
mean_losses = SMAPE(reduction="none")(predictions, actuals).mean(1)
indices = mean_losses.argsort(descending=True) # sort losses
# show only two examples for demonstration purposes
for idx in range(2):
best_tft.plot_prediction(x,
raw_predictions,
idx=indices[idx],
add_loss_to_title=SMAPE())
return_decoder_lengths=True)
finally:
shutil.rmtree(tmp_path, ignore_errors=True)
net.predict(val_dataloader,
fast_dev_run=True,
return_index=True,
return_decoder_lengths=True)
@pytest.mark.parametrize(
"kwargs",
[
{},
dict(
loss=MultiLoss([QuantileLoss(), MAE()]),
data_loader_kwargs=dict(
time_varying_unknown_reals=["volume", "discount"],
target=["volume", "discount"],
),
),
dict(
loss=CrossEntropy(),
data_loader_kwargs=dict(target="agency", ),
),
],
)
def test_integration(data_with_covariates, tmp_path, gpus, kwargs):
_integration(data_with_covariates.assign(target=lambda x: x.volume),
tmp_path, gpus, **kwargs)
train_actuals = actuals[0:-val_size]
test_actuals = actuals[-val_size:-1]
train_pred = predictions[0:-val_size]
test_pred = predictions[-val_size:-1]
val_mse = (test_actuals - test_pred).abs().mean()
print(f"validation MSE: {val_mse}")
# # calcualte metric by which to display
# indices = smape.argsort(descending=True) # sort losses
# for idx in range(10): # plot 10 examples
# best_tft.plot_prediction(x, raw_predictions, idx=indices[idx],
# add_loss_to_title=SMAPE())
# pyplot.show()
interpretation = best_tft.interpret_output(raw_predictions,
reduction="sum")
best_tft.plot_interpretation(interpretation)
pyplot.show()
mae = MAE(reduction="none")(predictions, actuals).mean(1)
mae = mae.data[0]
print(f"Mean average absolute error.: {mae}")
predictions, x = best_tft.predict(val_dataloader, return_x=True)
predictions_vs_actuals = best_tft.calculate_prediction_actual_by_variable(
x, predictions)
best_tft.plot_prediction_actual_by_variable(predictions_vs_actuals)
pyplot.show()
