# Copyright 2018 Amazon.com, Inc. or its affiliates. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License").
# You may not use this file except in compliance with the License.
# A copy of the License is located at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# or in the "license" file accompanying this file. This file is distributed
# on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either
# express or implied. See the License for the specific language governing
# permissions and limitations under the License.
from typing import Optional, Tuple
import numpy as np
import torch
from torch import nn
from gluonts.core.component import validated
from gluonts.model import Input, InputSpec
from gluonts.torch.distributions import StudentTOutput
from gluonts.torch.scaler import StdScaler, MeanScaler, NOPScaler
from gluonts.torch.util import take_last, unsqueeze_expand, weighted_average
from gluonts.torch.model.simple_feedforward import make_linear_layer
[docs]class SinusoidalPositionalEmbedding(nn.Embedding):
"""
This module produces sinusoidal positional embeddings of any length.
"""
def __init__(self, num_positions: int, embedding_dim: int) -> None:
super().__init__(num_positions, embedding_dim)
self.weight = self._init_weight(self.weight) # type: ignore
@staticmethod
def _init_weight(out: torch.Tensor) -> torch.Tensor:
"""
Features are not interleaved.
The cos features are in the 2nd half of the vector. [dim // 2:]
"""
n_pos, dim = out.shape
position_enc = np.array(
[
[pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)]
for pos in range(n_pos)
]
)
# set early to avoid an error in pytorch-1.8+
out.requires_grad = False
sentinel = dim // 2 if dim % 2 == 0 else (dim // 2) + 1
out[:, 0:sentinel] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))
out[:, sentinel:] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
out.detach_()
return out
[docs] @torch.no_grad()
def forward( # type: ignore
self, input_ids_shape: torch.Size, past_key_values_length: int = 0
) -> torch.Tensor:
"""
`input_ids_shape` is expected to be [bsz x seqlen x ...].
"""
_, seq_len = input_ids_shape[:2]
positions = torch.arange(
past_key_values_length,
past_key_values_length + seq_len,
dtype=torch.long,
device=self.weight.device,
)
return super().forward(positions)
[docs]class PatchTSTModel(nn.Module):
"""
Module implementing the PatchTST model for forecasting as described in
https://arxiv.org/abs/2211.14730 extended to be probabilistic.
Parameters
----------
prediction_length
Number of time points to predict.
context_length
Number of time steps prior to prediction time that the model.
num_feat_dynamic_real
Number of dynamic real features in the data (default: 0).
distr_output
Distribution to use to evaluate observations and sample predictions.
Default: ``StudentTOutput()``.
"""
@validated()
def __init__(
self,
prediction_length: int,
context_length: int,
patch_len: int,
stride: int,
padding_patch: str,
d_model: int,
nhead: int,
dim_feedforward: int,
num_feat_dynamic_real: int,
dropout: float,
activation: str,
norm_first: bool,
num_encoder_layers: int,
scaling: Optional[str],
distr_output=StudentTOutput(),
) -> None:
super().__init__()
assert prediction_length > 0
assert context_length > 0
self.prediction_length = prediction_length
self.context_length = context_length
self.patch_len = patch_len
self.stride = stride
self.d_model = d_model
self.padding_patch = padding_patch
self.distr_output = distr_output
self.num_feat_dynamic_real = num_feat_dynamic_real
if scaling == "mean":
self.scaler = MeanScaler(keepdim=True)
elif scaling == "std":
self.scaler = StdScaler(keepdim=True)
else:
self.scaler = NOPScaler(keepdim=True)
self.patch_num = int((context_length - patch_len) / stride + 1)
if padding_patch == "end": # can be modified to general case
self.padding_patch_layer = nn.ReplicationPad1d((0, self.stride))
self.patch_num += 1
# project from `patch_len` + 2 features (`loc` and `scale`) +
# `num_feat_dynamic_real` x `patch_len` to d_model
self.patch_proj = make_linear_layer(
patch_len + 2 + self.num_feat_dynamic_real * patch_len, d_model
)
self.positional_encoding = SinusoidalPositionalEmbedding(
self.patch_num, d_model
)
layer_norm_eps: float = 1e-5
encoder_layer = nn.TransformerEncoderLayer(
d_model=d_model,
nhead=nhead,
dim_feedforward=dim_feedforward,
dropout=dropout,
activation=activation,
layer_norm_eps=layer_norm_eps,
batch_first=True,
norm_first=norm_first,
)
encoder_norm = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.encoder = nn.TransformerEncoder(
encoder_layer, num_encoder_layers, encoder_norm
)
self.flatten = nn.Linear(
d_model * self.patch_num, prediction_length * d_model
)
self.args_proj = self.distr_output.get_args_proj(d_model)
[docs] def forward(
self,
past_target: torch.Tensor,
past_observed_values: torch.Tensor,
past_time_feat: Optional[torch.Tensor] = None,
future_time_feat: Optional[torch.Tensor] = None,
) -> Tuple[Tuple[torch.Tensor, ...], torch.Tensor, torch.Tensor]:
# scale the input
past_target_scaled, loc, scale = self.scaler(
past_target, past_observed_values
)
# do patching
if self.padding_patch == "end":
past_target_scaled = self.padding_patch_layer(past_target_scaled)
past_target_patches = past_target_scaled.unfold(
dimension=1, size=self.patch_len, step=self.stride
)
# do patching for time features as well
if self.num_feat_dynamic_real > 0:
# shift time features by `prediction_length` so that they are
# aligned with the target input.
time_feat = take_last(
torch.cat((past_time_feat, future_time_feat), dim=1),
dim=1,
num=self.context_length,
)
# (bs x T x d) --> (bs x d x T) because the 1D padding is done on
# the last dimension.
time_feat = self.padding_patch_layer(
time_feat.transpose(-2, -1)
).transpose(-2, -1)
time_feat_patches = time_feat.unfold(
dimension=1, size=self.patch_len, step=self.stride
).flatten(-2, -1)
# add loc and scale to past_target_patches as additional features
log_abs_loc = loc.abs().log1p()
log_scale = scale.log()
expanded_static_feat = unsqueeze_expand(
torch.cat([log_abs_loc, log_scale], dim=-1),
dim=1,
size=past_target_patches.shape[1],
)
inputs = torch.cat((past_target_patches, expanded_static_feat), dim=-1)
if self.num_feat_dynamic_real > 0:
inputs = torch.cat((inputs, time_feat_patches), dim=-1)
# project patches
enc_in = self.patch_proj(inputs)
embed_pos = self.positional_encoding(enc_in.size())
# transformer encoder with positional encoding
enc_out = self.encoder(enc_in + embed_pos)
# flatten and project to prediction length * d_model
flatten_out = self.flatten(enc_out.flatten(start_dim=1))
# project to distribution arguments
distr_args = self.args_proj(
flatten_out.reshape(-1, self.prediction_length, self.d_model)
)
return distr_args, loc, scale
[docs] def loss(
self,
past_target: torch.Tensor,
past_observed_values: torch.Tensor,
future_target: torch.Tensor,
future_observed_values: torch.Tensor,
past_time_feat: Optional[torch.Tensor] = None,
future_time_feat: Optional[torch.Tensor] = None,
) -> torch.Tensor:
distr_args, loc, scale = self(
past_target=past_target,
past_observed_values=past_observed_values,
past_time_feat=past_time_feat,
future_time_feat=future_time_feat,
)
loss = self.distr_output.loss(
target=future_target, distr_args=distr_args, loc=loc, scale=scale
)
return weighted_average(loss, weights=future_observed_values, dim=-1)