# coding=utf-8
# Copyright 2023 Amazon and The HuggingFace Inc. team. 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.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License 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.
"""PyTorch Informer model."""
from typing import List, Optional, Tuple, Union
import numpy as np
import torch
from torch import nn
from transformers.activations import ACT2FN
from transformers.modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
from transformers.modeling_outputs import (
BaseModelOutput,
BaseModelOutputWithPastAndCrossAttentions,
SampleTSPredictionOutput,
Seq2SeqTSModelOutput,
Seq2SeqTSPredictionOutput,
)
from transformers.modeling_utils import PreTrainedModel
from transformers.time_series_utils import NegativeBinomialOutput, NormalOutput, StudentTOutput
from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
from .configuration_informer import InformerConfig
logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "InformerConfig"
# Copied from transformers.models.time_series_transformer.modeling_time_series_transformer.TimeSeriesFeatureEmbedder with TimeSeries->Informer
class InformerFeatureEmbedder(nn.Module):
"""
Embed a sequence of categorical features.
Args:
cardinalities (`list[int]`):
List of cardinalities of the categorical features.
embedding_dims (`list[int]`):
List of embedding dimensions of the categorical features.
"""
def __init__(self, cardinalities: List[int], embedding_dims: List[int]) -> None:
super().__init__()
self.num_features = len(cardinalities)
self.embedders = nn.ModuleList([nn.Embedding(c, d) for c, d in zip(cardinalities, embedding_dims)])
def forward(self, features: torch.Tensor) -> torch.Tensor:
if self.num_features > 1:
# we slice the last dimension, giving an array of length
# self.num_features with shape (N,T) or (N)
cat_feature_slices = torch.chunk(features, self.num_features, dim=-1)
else:
cat_feature_slices = [features]
return torch.cat(
[
embed(cat_feature_slice.squeeze(-1))
for embed, cat_feature_slice in zip(self.embedders, cat_feature_slices)
],
dim=-1,
)
# Copied from transformers.models.time_series_transformer.modeling_time_series_transformer.TimeSeriesStdScaler with TimeSeriesTransformer->Informer,TimeSeries->Informer
class InformerStdScaler(nn.Module):
"""
Standardize features by calculating the mean and scaling along the first dimension, and then normalizes it by
subtracting from the mean and dividing by the standard deviation.
"""
def __init__(self, config: InformerConfig):
super().__init__()
self.dim = config.scaling_dim if hasattr(config, "scaling_dim") else 1
self.keepdim = config.keepdim if hasattr(config, "keepdim") else True
self.minimum_scale = config.minimum_scale if hasattr(config, "minimum_scale") else 1e-5
def forward(
self, data: torch.Tensor, observed_indicator: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Parameters:
data (`torch.Tensor` of shape `(batch_size, sequence_length, num_input_channels)`):
input for Batch norm calculation
observed_indicator (`torch.BoolTensor` of shape `(batch_size, sequence_length, num_input_channels)`):
Calculating the scale on the observed indicator.
Returns:
tuple of `torch.Tensor` of shapes
(`(batch_size, sequence_length, num_input_channels)`,`(batch_size, 1, num_input_channels)`,
`(batch_size, 1, num_input_channels)`)
"""
denominator = observed_indicator.sum(self.dim, keepdim=self.keepdim)
denominator = denominator.clamp_min(1.0)
loc = (data * observed_indicator).sum(self.dim, keepdim=self.keepdim) / denominator
variance = (((data - loc) * observed_indicator) ** 2).sum(self.dim, keepdim=self.keepdim) / denominator
scale = torch.sqrt(variance + self.minimum_scale)
return (data - loc) / scale, loc, scale
# Copied from transformers.models.time_series_transformer.modeling_time_series_transformer.TimeSeriesMeanScaler with TimeSeriesTransformer->Informer,TimeSeries->Informer
class InformerMeanScaler(nn.Module):
"""
Computes a scaling factor as the weighted average absolute value along the first dimension, and scales the data
accordingly.
"""
def __init__(self, config: InformerConfig):
super().__init__()
self.dim = config.scaling_dim if hasattr(config, "scaling_dim") else 1
self.keepdim = config.keepdim if hasattr(config, "keepdim") else True
self.minimum_scale = config.minimum_scale if hasattr(config, "minimum_scale") else 1e-10
self.default_scale = config.default_scale if hasattr(config, "default_scale") else None
def forward(
self, data: torch.Tensor, observed_indicator: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Parameters:
data (`torch.Tensor` of shape `(batch_size, sequence_length, num_input_channels)`):
input for Batch norm calculation
observed_indicator (`torch.BoolTensor` of shape `(batch_size, sequence_length, num_input_channels)`):
Calculating the scale on the observed indicator.
Returns:
tuple of `torch.Tensor` of shapes
(`(batch_size, sequence_length, num_input_channels)`,`(batch_size, 1, num_input_channels)`,
`(batch_size, 1, num_input_channels)`)
"""
ts_sum = (data * observed_indicator).abs().sum(self.dim, keepdim=True)
num_observed = observed_indicator.sum(self.dim, keepdim=True)
scale = ts_sum / torch.clamp(num_observed, min=1)
# If `default_scale` is provided, we use it, otherwise we use the scale
# of the batch.
if self.default_scale is None:
batch_sum = ts_sum.sum(dim=0)
batch_observations = torch.clamp(num_observed.sum(0), min=1)
default_scale = torch.squeeze(batch_sum / batch_observations)
else:
default_scale = self.default_scale * torch.ones_like(scale)
# apply default scale where there are no observations
scale = torch.where(num_observed > 0, scale, default_scale)
# ensure the scale is at least `self.minimum_scale`
scale = torch.clamp(scale, min=self.minimum_scale)
scaled_data = data / scale
if not self.keepdim:
scale = scale.squeeze(dim=self.dim)
return scaled_data, torch.zeros_like(scale), scale
# Copied from transformers.models.time_series_transformer.modeling_time_series_transformer.TimeSeriesNOPScaler with TimeSeriesTransformer->Informer,TimeSeries->Informer
class InformerNOPScaler(nn.Module):
"""
Assigns a scaling factor equal to 1 along the first dimension, and therefore applies no scaling to the input data.
"""
def __init__(self, config: InformerConfig):
super().__init__()
self.dim = config.scaling_dim if hasattr(config, "scaling_dim") else 1
self.keepdim = config.keepdim if hasattr(config, "keepdim") else True
def forward(
self, data: torch.Tensor, observed_indicator: torch.Tensor = None
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Parameters:
data (`torch.Tensor` of shape `(batch_size, sequence_length, num_input_channels)`):
input for Batch norm calculation
Returns:
tuple of `torch.Tensor` of shapes
(`(batch_size, sequence_length, num_input_channels)`,`(batch_size, 1, num_input_channels)`,
`(batch_size, 1, num_input_channels)`)
"""
scale = torch.ones_like(data, requires_grad=False).mean(dim=self.dim, keepdim=self.keepdim)
loc = torch.zeros_like(data, requires_grad=False).mean(dim=self.dim, keepdim=self.keepdim)
return data, loc, scale
# Copied from transformers.models.time_series_transformer.modeling_time_series_transformer.weighted_average
def weighted_average(input_tensor: torch.Tensor, weights: Optional[torch.Tensor] = None, dim=None) -> torch.Tensor:
"""
Computes the weighted average of a given tensor across a given `dim`, masking values associated with weight zero,
meaning instead of `nan * 0 = nan` you will get `0 * 0 = 0`.
Args:
input_tensor (`torch.FloatTensor`):
Input tensor, of which the average must be computed.
weights (`torch.FloatTensor`, *optional*):
Weights tensor, of the same shape as `input_tensor`.
dim (`int`, *optional*):
The dim along which to average `input_tensor`.
Returns:
`torch.FloatTensor`: The tensor with values averaged along the specified `dim`.
"""
if weights is not None:
weighted_tensor = torch.where(weights != 0, input_tensor * weights, torch.zeros_like(input_tensor))
sum_weights = torch.clamp(weights.sum(dim=dim) if dim else weights.sum(), min=1.0)
return (weighted_tensor.sum(dim=dim) if dim else weighted_tensor.sum()) / sum_weights
else:
return input_tensor.mean(dim=dim)
# Copied from transformers.models.time_series_transformer.modeling_time_series_transformer.nll
def nll(input: torch.distributions.Distribution, target: torch.Tensor) -> torch.Tensor:
"""
Computes the negative log likelihood loss from input distribution with respect to target.
"""
return -input.log_prob(target)
# Copied from transformers.models.marian.modeling_marian.MarianSinusoidalPositionalEmbedding with Marian->Informer
class InformerSinusoidalPositionalEmbedding(nn.Embedding):
"""This module produces sinusoidal positional embeddings of any length."""
def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None) -> None:
super().__init__(num_positions, embedding_dim)
self.weight = self._init_weight(self.weight)
@staticmethod
def _init_weight(out: nn.Parameter) -> nn.Parameter:
"""
Identical to the XLM create_sinusoidal_embeddings except 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)]
)
out.requires_grad = False # set early to avoid an error in pytorch-1.8+
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
@torch.no_grad()
def forward(self, input_ids_shape: torch.Size, past_key_values_length: int = 0) -> torch.Tensor:
"""`input_ids_shape` is expected to be [bsz x seqlen]."""
bsz, 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)
# Copied from transformers.models.time_series_transformer.modeling_time_series_transformer.TimeSeriesValueEmbedding with TimeSeries->Info
class InformerValueEmbedding(nn.Module):
def __init__(self, feature_size, d_model):
super().__init__()
self.value_projection = nn.Linear(in_features=feature_size, out_features=d_model, bias=False)
def forward(self, x):
return self.value_projection(x)
# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->Informer
class InformerAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
def __init__(
self,
embed_dim: int,
num_heads: int,
dropout: float = 0.0,
is_decoder: bool = False,
bias: bool = True,
is_causal: bool = False,
config: Optional[InformerConfig] = None,
):
super().__init__()
self.embed_dim = embed_dim
self.num_heads = num_heads
self.dropout = dropout
self.head_dim = embed_dim // num_heads
self.config = config
if (self.head_dim * num_heads) != self.embed_dim:
raise ValueError(
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
f" and `num_heads`: {num_heads})."
)
self.scaling = self.head_dim**-0.5
self.is_decoder = is_decoder
self.is_causal = is_causal
self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
def forward(
self,
hidden_states: torch.Tensor,
key_value_states: Optional[torch.Tensor] = None,
past_key_value: Optional[Tuple[torch.Tensor]] = None,
attention_mask: Optional[torch.Tensor] = None,
layer_head_mask: Optional[torch.Tensor] = None,
output_attentions: bool = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
"""Input shape: Batch x Time x Channel"""
# if key_value_states are provided this layer is used as a cross-attention layer
# for the decoder
is_cross_attention = key_value_states is not None
bsz, tgt_len, _ = hidden_states.size()
# get query proj
query_states = self.q_proj(hidden_states) * self.scaling
# get key, value proj
# `past_key_value[0].shape[2] == key_value_states.shape[1]`
# is checking that the `sequence_length` of the `past_key_value` is the same as
# the provided `key_value_states` to support prefix tuning
if (
is_cross_attention
and past_key_value is not None
and past_key_value[0].shape[2] == key_value_states.shape[1]
):
# reuse k,v, cross_attentions
key_states = past_key_value[0]
value_states = past_key_value[1]
elif is_cross_attention:
# cross_attentions
key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
elif past_key_value is not None:
# reuse k, v, self_attention
key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
key_states = torch.cat([past_key_value[0], key_states], dim=2)
value_states = torch.cat([past_key_value[1], value_states], dim=2)
else:
# self_attention
key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
if self.is_decoder:
# if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
# Further calls to cross_attention layer can then reuse all cross-attention
# key/value_states (first "if" case)
# if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
# all previous decoder key/value_states. Further calls to uni-directional self-attention
# can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
# if encoder bi-directional self-attention `past_key_value` is always `None`
past_key_value = (key_states, value_states)
proj_shape = (bsz * self.num_heads, -1, self.head_dim)
query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
key_states = key_states.reshape(*proj_shape)
value_states = value_states.reshape(*proj_shape)
src_len = key_states.size(1)
attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
raise ValueError(
f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
f" {attn_weights.size()}"
)
if attention_mask is not None:
if attention_mask.size() != (bsz, 1, tgt_len, src_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
)
attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
if layer_head_mask is not None:
if layer_head_mask.size() != (self.num_heads,):
raise ValueError(
f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
f" {layer_head_mask.size()}"
)
attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
if output_attentions:
# this operation is a bit awkward, but it's required to
# make sure that attn_weights keeps its gradient.
# In order to do so, attn_weights have to be reshaped
# twice and have to be reused in the following
attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
else:
attn_weights_reshaped = None
attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
attn_output = torch.bmm(attn_probs, value_states)
if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
attn_output = attn_output.transpose(1, 2)
# Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
# partitioned across GPUs when using tensor-parallelism.
attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
attn_output = self.out_proj(attn_output)
return attn_output, attn_weights_reshaped, past_key_value
class InformerProbSparseAttention(nn.Module):
"""Probabilistic Attention mechanism to select the "active"
queries rather than the "lazy" queries and provides a sparse Transformer thus mitigating the quadratic compute and
memory requirements of vanilla attention"""
def __init__(
self,
embed_dim: int,
num_heads: int,
dropout: float = 0.0,
is_decoder: bool = False,
sampling_factor: int = 5,
bias: bool = True,
):
super().__init__()
self.factor = sampling_factor
self.embed_dim = embed_dim
self.num_heads = num_heads
self.dropout = dropout
self.head_dim = embed_dim // num_heads
if (self.head_dim * num_heads) != self.embed_dim:
raise ValueError(
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
f" and `num_heads`: {num_heads})."
)
self.scaling = self.head_dim**-0.5
self.is_decoder = is_decoder
self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
def forward(
self,
hidden_states: torch.Tensor,
key_value_states: Optional[torch.Tensor] = None,
past_key_value: Optional[Tuple[torch.Tensor]] = None,
attention_mask: Optional[torch.Tensor] = None,
layer_head_mask: Optional[torch.Tensor] = None,
output_attentions: bool = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
"""Input shape: Batch x Time x Channel"""
# if key_value_states are provided this layer is used as a cross-attention layer
# for the decoder
is_cross_attention = key_value_states is not None
bsz, tgt_len, _ = hidden_states.size()
# get query proj
query_states = self.q_proj(hidden_states) * self.scaling
# get key, value proj
# `past_key_value[0].shape[2] == key_value_states.shape[1]`
# is checking that the `sequence_length` of the `past_key_value` is the same as
# the provided `key_value_states` to support prefix tuning
if (
is_cross_attention
and past_key_value is not None
and past_key_value[0].shape[2] == key_value_states.shape[1]
):
# reuse k,v, cross_attentions
key_states = past_key_value[0]
value_states = past_key_value[1]
elif is_cross_attention:
# cross_attentions
key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
elif past_key_value is not None:
# reuse k, v, self_attention
key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
key_states = torch.cat([past_key_value[0], key_states], dim=2)
value_states = torch.cat([past_key_value[1], value_states], dim=2)
else:
# self_attention
key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
if self.is_decoder:
# if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
# Further calls to cross_attention layer can then reuse all cross-attention
# key/value_states (first "if" case)
# if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
# all previous decoder key/value_states. Further calls to uni-directional self-attention
# can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
# if encoder bi-directional self-attention `past_key_value` is always `None`
past_key_value = (key_states, value_states)
proj_shape = (bsz * self.num_heads, -1, self.head_dim)
query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
key_states = key_states.reshape(*proj_shape)
value_states = value_states.reshape(*proj_shape)
key_states_time_length = key_states.size(1) # L_K
log_key_states_time_length = np.ceil(np.log1p(key_states_time_length)).astype("int").item() # log_L_K
query_states_time_length = query_states.size(1) # L_Q
log_query_states_time_length = np.ceil(np.log1p(query_states_time_length)).astype("int").item() # log_L_Q
u_part = min(self.factor * query_states_time_length * log_key_states_time_length, key_states_time_length)
u = min(self.factor * log_query_states_time_length, query_states_time_length)
if key_states_time_length > 0:
index_sample = torch.randint(0, key_states_time_length, (u_part,))
k_sample = key_states[:, index_sample, :]
else:
k_sample = key_states
queries_keys_sample = torch.bmm(query_states, k_sample.transpose(1, 2)) # Q_K_sampled
# find the Top_k query with sparsity measurement
if u > 0:
sparsity_measurement = queries_keys_sample.max(dim=-1)[0] - torch.div(
queries_keys_sample.sum(dim=-1), key_states_time_length
) # M
top_u_sparsity_measurement = sparsity_measurement.topk(u, sorted=False)[1] # M_top
# calculate q_reduce: query_states[:, top_u_sparsity_measurement]
dim_for_slice = torch.arange(query_states.size(0)).unsqueeze(-1)
q_reduce = query_states[dim_for_slice, top_u_sparsity_measurement]
else:
q_reduce = query_states
top_u_sparsity_measurement = None
# Use q_reduce to calculate attention weights
attn_weights = torch.bmm(q_reduce, key_states.transpose(1, 2))
src_len = key_states.size(1)
if attn_weights.size() != (bsz * self.num_heads, u, src_len):
raise ValueError(
f"Attention weights should be of size {(bsz * self.num_heads, u, src_len)}, but is"
f" {attn_weights.size()}"
)
if attention_mask is not None:
if attention_mask.size() != (bsz, 1, tgt_len, src_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
)
prob_mask = attention_mask.expand(bsz, self.num_heads, tgt_len, src_len).reshape(
bsz * self.num_heads, tgt_len, src_len
)
if top_u_sparsity_measurement is not None:
dim_for_slice = torch.arange(prob_mask.size(0)).unsqueeze(-1)
prob_mask = prob_mask[dim_for_slice, top_u_sparsity_measurement, :]
attn_weights = attn_weights.view(bsz, self.num_heads, u, src_len) + prob_mask.view(
bsz, self.num_heads, u, src_len
)
attn_weights = attn_weights.view(bsz * self.num_heads, u, src_len)
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
if layer_head_mask is not None:
if layer_head_mask.size() != (self.num_heads,):
raise ValueError(
f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
f" {layer_head_mask.size()}"
)
attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, u, src_len)
attn_weights = attn_weights.view(bsz * self.num_heads, u, src_len)
if output_attentions:
# this operation is a bit awkward, but it's required to
# make sure that attn_weights keeps its gradient.
# In order to do so, attn_weights have to be reshaped
# twice and have to be reused in the following
attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, u, src_len)
attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, u, src_len)
else:
attn_weights_reshaped = None
attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
attn_output = torch.bmm(attn_probs, value_states)
# calculate context for updating the attn_output, based on:
# https://github.com/zhouhaoyi/Informer2020/blob/ac59c7447135473fb2aafeafe94395f884d5c7a5/models/attn.py#L74
if self.is_decoder:
# cast to float32 before operation to avoid overflow
context = value_states.cumsum(dim=-2, dtype=torch.float32).to(value_states.dtype)
else:
v_mean_dim_time = value_states.mean(dim=-2)
context = (
v_mean_dim_time.unsqueeze(dim=1)
.expand(bsz * self.num_heads, query_states_time_length, v_mean_dim_time.size(-1))
.clone()
)
if top_u_sparsity_measurement is not None:
# update context: copy the attention output to the context at top_u_sparsity_measurement index
dim_for_slice = torch.arange(context.size(0)).unsqueeze(-1)
context[dim_for_slice, top_u_sparsity_measurement, :] = attn_output
attn_output = context
if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
attn_output = attn_output.transpose(1, 2)
# Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
# partitioned across GPUs when using tensor-parallelism.
attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
attn_output = self.out_proj(attn_output)
return attn_output, attn_weights_reshaped, past_key_value
# source: https://github.com/zhouhaoyi/Informer2020/blob/main/models/encoder.py
class InformerConvLayer(nn.Module):
def __init__(self, c_in):
super().__init__()
self.downConv = nn.Conv1d(
in_channels=c_in,
out_channels=c_in,
kernel_size=3,
padding=1,
padding_mode="circular",
)
self.norm = nn.BatchNorm1d(c_in)
self.activation = nn.ELU()
self.maxPool = nn.MaxPool1d(kernel_size=3, stride=2, padding=1)
def forward(self, x):
x = self.downConv(x.permute(0, 2, 1))
x = self.norm(x)
x = self.activation(x)
x = self.maxPool(x)
x = x.transpose(1, 2)
return x
class InformerEncoderLayer(nn.Module):
def __init__(self, config: InformerConfig):
super().__init__()
self.embed_dim = config.d_model
if config.attention_type == "prob":
self.self_attn = InformerProbSparseAttention(
embed_dim=self.embed_dim,
num_heads=config.encoder_attention_heads,
dropout=config.attention_dropout,
sampling_factor=config.sampling_factor,
)
else:
self.self_attn = InformerAttention(
embed_dim=self.embed_dim,
num_heads=config.encoder_attention_heads,
dropout=config.attention_dropout,
)
self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
self.dropout = config.dropout
self.activation_fn = ACT2FN[config.activation_function]
self.activation_dropout = config.activation_dropout
self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
self.final_layer_norm = nn.LayerNorm(self.embed_dim)
def forward(
self,
hidden_states: torch.FloatTensor,
attention_mask: torch.FloatTensor,
layer_head_mask: torch.FloatTensor,
output_attentions: Optional[bool] = False,
) -> Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]:
"""
Args:
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
attention_mask (`torch.FloatTensor`): attention mask of size
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
`(encoder_attention_heads,)`.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
"""
residual = hidden_states
hidden_states, attn_weights, _ = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
layer_head_mask=layer_head_mask,
output_attentions=output_attentions,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.self_attn_layer_norm(hidden_states)
residual = hidden_states
hidden_states = self.activation_fn(self.fc1(hidden_states))
hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
hidden_states = self.fc2(hidden_states)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.final_layer_norm(hidden_states)
if hidden_states.dtype == torch.float16 and (
torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any()
):
clamp_value = torch.finfo(hidden_states.dtype).max - 1000
hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
outputs = (hidden_states,)
if output_attentions:
outputs += (attn_weights,)
return outputs
class InformerDecoderLayer(nn.Module):
def __init__(self, config: InformerConfig):
super().__init__()
self.embed_dim = config.d_model
if config.attention_type == "prob":
self.self_attn = InformerProbSparseAttention(
embed_dim=self.embed_dim,
num_heads=config.decoder_attention_heads,
dropout=config.attention_dropout,
sampling_factor=config.sampling_factor,
is_decoder=True,
)
else:
self.self_attn = InformerAttention(
embed_dim=self.embed_dim,
num_heads=config.decoder_attention_heads,
dropout=config.attention_dropout,
is_decoder=True,
)
self.dropout = config.dropout
self.activation_fn = ACT2FN[config.activation_function]
self.activation_dropout = config.activation_dropout
self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
self.encoder_attn = InformerAttention(
self.embed_dim,
config.decoder_attention_heads,
dropout=config.attention_dropout,
is_decoder=True,
)
self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
self.final_layer_norm = nn.LayerNorm(self.embed_dim)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
encoder_hidden_states: Optional[torch.Tensor] = None,
encoder_attention_mask: Optional[torch.Tensor] = None,
layer_head_mask: Optional[torch.Tensor] = None,
cross_attn_layer_head_mask: Optional[torch.Tensor] = None,
past_key_value: Optional[Tuple[torch.Tensor]] = None,
output_attentions: Optional[bool] = False,
use_cache: Optional[bool] = True,
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
"""
Args:
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
attention_mask (`torch.FloatTensor`): attention mask of size
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
encoder_hidden_states (`torch.FloatTensor`):
cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
`(encoder_attention_heads,)`.
cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of
size `(decoder_attention_heads,)`.
past_key_value (`Tuple(torch.FloatTensor)`): cached past key and value projection states
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
"""
residual = hidden_states
# Self Attention
# decoder uni-directional self-attention cached key/values tuple is at positions 1,2
self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
# add present self-attn cache to positions 1,2 of present_key_value tuple
hidden_states, self_attn_weights, present_key_value = self.self_attn(
hidden_states=hidden_states,
past_key_value=self_attn_past_key_value,
attention_mask=attention_mask,
layer_head_mask=layer_head_mask,
output_attentions=output_attentions,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.self_attn_layer_norm(hidden_states)
# Cross-Attention Block
cross_attn_present_key_value = None
cross_attn_weights = None
if encoder_hidden_states is not None:
residual = hidden_states
# cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
hidden_states=hidden_states,
key_value_states=encoder_hidden_states,
attention_mask=encoder_attention_mask,
layer_head_mask=cross_attn_layer_head_mask,
past_key_value=cross_attn_past_key_value,
output_attentions=output_attentions,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.encoder_attn_layer_norm(hidden_states)
# add cross-attn to positions 3,4 of present_key_value tuple
present_key_value = present_key_value + cross_attn_present_key_value
# Fully Connected
residual = hidden_states
hidden_states = self.activation_fn(self.fc1(hidden_states))
hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
hidden_states = self.fc2(hidden_states)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.final_layer_norm(hidden_states)
outputs = (hidden_states,)
if output_attentions:
outputs += (self_attn_weights, cross_attn_weights)
if use_cache:
outputs += (present_key_value,)
return outputs
INFORMER_START_DOCSTRING = r"""
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.
Parameters:
config ([`TimeSeriesTransformerConfig`]):
Model configuration class with all the parameters of the model. Initializing with a config file does not
load the weights associated with the model, only the configuration. Check out the
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
INFORMER_INPUTS_DOCSTRING = r"""
Args:
past_values (`torch.FloatTensor` of shape `(batch_size, sequence_length)` or `(batch_size, sequence_length, input_size)`):
Past values of the time series, that serve as context in order to predict the future. The sequence size of
this tensor must be larger than the `context_length` of the model, since the model will use the larger size
to construct lag features, i.e. additional values from the past which are added in order to serve as "extra
context".
The `sequence_length` here is equal to `config.context_length` + `max(config.lags_sequence)`, which if no
`lags_sequence` is configured, is equal to `config.context_length` + 7 (as by default, the largest
look-back index in `config.lags_sequence` is 7). The property `_past_length` returns the actual length of
the past.
The `past_values` is what the Transformer encoder gets as input (with optional additional features, such as
`static_categorical_features`, `static_real_features`, `past_time_features` and lags).
Optionally, missing values need to be replaced with zeros and indicated via the `past_observed_mask`.
For multivariate time series, the `input_size` > 1 dimension is required and corresponds to the number of
variates in the time series per time step.
past_time_features (`torch.FloatTensor` of shape `(batch_size, sequence_length, num_features)`):
Required time features, which the model internally will add to `past_values`. These could be things like
"month of year", "day of the month", etc. encoded as vectors (for instance as Fourier features). These
could also be so-called "age" features, which basically help the model know "at which point in life" a
time-series is. Age features have small values for distant past time steps and increase monotonically the
more we approach the current time step. Holiday features are also a good example of time features.
These features serve as the "positional encodings" of the inputs. So contrary to a model like BERT, where
the position encodings are learned from scratch internally as parameters of the model, the Time Series
Transformer requires to provide additional time features. The Time Series Transformer only learns
additional embeddings for `static_categorical_features`.
Additional dynamic real covariates can be concatenated to this tensor, with the caveat that these features
must but known at prediction time.
The `num_features` here is equal to `config.`num_time_features` + `config.num_dynamic_real_features`.
past_observed_mask (`torch.BoolTensor` of shape `(batch_size, sequence_length)` or `(batch_size, sequence_length, input_size)`, *optional*):
Boolean mask to indicate which `past_values` were observed and which were missing. Mask values selected in
`[0, 1]`:
- 1 for values that are **observed**,
- 0 for values that are **missing** (i.e. NaNs that were replaced by zeros).
static_categorical_features (`torch.LongTensor` of shape `(batch_size, number of static categorical features)`, *optional*):
Optional static categorical features for which the model will learn an embedding, which it will add to the
values of the time series.
Static categorical features are features which have the same value for all time steps (static over time).
A typical example of a static categorical feature is a time series ID.
static_real_features (`torch.FloatTensor` of shape `(batch_size, number of static real features)`, *optional*):
Optional static real features which the model will add to the values of the time series.
Static real features are features which have the same value for all time steps (static over time).
A typical example of a static real feature is promotion information.
future_values (`torch.FloatTensor` of shape `(batch_size, prediction_length)` or `(batch_size, prediction_length, input_size)`, *optional*):
Future values of the time series, that serve as labels for the model. The `future_values` is what the
Transformer needs during training to learn to output, given the `past_values`.
The sequence length here is equal to `prediction_length`.
See the demo notebook and code snippets for details.
Optionally, during training any missing values need to be replaced with zeros and indicated via the
`future_observed_mask`.
For multivariate time series, the `input_size` > 1 dimension is required and corresponds to the number of
variates in the time series per time step.
future_time_features (`torch.FloatTensor` of shape `(batch_size, prediction_length, num_features)`):
Required time features for the prediction window, which the model internally will add to `future_values`.
These could be things like "month of year", "day of the month", etc. encoded as vectors (for instance as
Fourier features). These could also be so-called "age" features, which basically help the model know "at
which point in life" a time-series is. Age features have small values for distant past time steps and
increase monotonically the more we approach the current time step. Holiday features are also a good example
of time features.
These features serve as the "positional encodings" of the inputs. So contrary to a model like BERT, where
the position encodings are learned from scratch internally as parameters of the model, the Time Series
Transformer requires to provide additional time features. The Time Series Transformer only learns
additional embeddings for `static_categorical_features`.
Additional dynamic real covariates can be concatenated to this tensor, with the caveat that these features
must but known at prediction time.
The `num_features` here is equal to `config.`num_time_features` + `config.num_dynamic_real_features`.
future_observed_mask (`torch.BoolTensor` of shape `(batch_size, sequence_length)` or `(batch_size, sequence_length, input_size)`, *optional*):
Boolean mask to indicate which `future_values` were observed and which were missing. Mask values selected
in `[0, 1]`:
- 1 for values that are **observed**,
- 0 for values that are **missing** (i.e. NaNs that were replaced by zeros).
This mask is used to filter out missing values for the final loss calculation.
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on certain token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
Mask to avoid performing attention on certain token indices. By default, a causal mask will be used, to
make sure the model can only look at previous inputs in order to predict the future.
head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
decoder_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
Tuple consists of `last_hidden_state`, `hidden_states` (*optional*) and `attentions` (*optional*)
`last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)` (*optional*) is a sequence of
hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
model's internal embedding lookup matrix.
use_cache (`bool`, *optional*):
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
`past_key_values`).
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
class InformerEncoder(InformerPreTrainedModel):
"""
Informer encoder consisting of *config.encoder_layers* self attention layers with distillation layers. Each
attention layer is an [`InformerEncoderLayer`].
Args:
config: InformerConfig
"""
def __init__(self, config: InformerConfig):
super().__init__(config)
self.dropout = config.dropout
self.layerdrop = config.encoder_layerdrop
self.gradient_checkpointing = False
if config.prediction_length is None:
raise ValueError("The `prediction_length` config needs to be specified.")
self.value_embedding = InformerValueEmbedding(feature_size=config.feature_size, d_model=config.d_model)
self.embed_positions = InformerSinusoidalPositionalEmbedding(
config.context_length + config.prediction_length, config.d_model
)
self.layers = nn.ModuleList([InformerEncoderLayer(config) for _ in range(config.encoder_layers)])
self.layernorm_embedding = nn.LayerNorm(config.d_model)
if config.distil:
self.conv_layers = nn.ModuleList(
[InformerConvLayer(config.d_model) for _ in range(config.encoder_layers - 1)]
)
self.conv_layers.append(None)
else:
self.conv_layers = [None] * config.encoder_layers
# Initialize weights and apply final processing
self.post_init()
def forward(
self,
attention_mask: Optional[torch.Tensor] = None,
head_mask: Optional[torch.Tensor] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutput]:
r"""
Args:
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
This is useful if you want more control over how to convert `input_ids` indices into associated vectors
than the model's internal embedding lookup matrix.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
for more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
hidden_states = self.value_embedding(inputs_embeds)
embed_pos = self.embed_positions(inputs_embeds.size())
hidden_states = self.layernorm_embedding(hidden_states + embed_pos)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
# expand attention_mask
if attention_mask is not None:
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
attention_mask = _prepare_4d_attention_mask(attention_mask, inputs_embeds.dtype)
encoder_states = () if output_hidden_states else None
all_attentions = () if output_attentions else None
# check if head_mask has a correct number of layers specified if desired
if head_mask is not None:
if head_mask.size()[0] != (len(self.layers)):
raise ValueError(
f"The head_mask should be specified for {len(self.layers)} layers, but it is for"
f" {head_mask.size()[0]}."
)
for idx, (encoder_layer, conv_layer) in enumerate(zip(self.layers, self.conv_layers)):
if output_hidden_states:
encoder_states = encoder_states + (hidden_states,)
# add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
to_drop = False
if self.training:
dropout_probability = torch.rand([])
if dropout_probability < self.layerdrop: # skip the layer
to_drop = True
if to_drop:
layer_outputs = (None, None)
else:
if self.gradient_checkpointing and self.training:
layer_outputs = self._gradient_checkpointing_func(
encoder_layer.__call__,
hidden_states,
attention_mask,
(head_mask[idx] if head_mask is not None else None),
output_attentions,
)
if conv_layer is not None:
output = self._gradient_checkpointing_func(conv_layer, layer_outputs[0])
layer_outputs = (output,) + layer_outputs[1:]
else:
layer_outputs = encoder_layer(
hidden_states,
attention_mask,
layer_head_mask=(head_mask[idx] if head_mask is not None else None),
output_attentions=output_attentions,
)
if conv_layer is not None:
output = conv_layer(layer_outputs[0])
layer_outputs = (output,) + layer_outputs[1:]
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions = all_attentions + (layer_outputs[1],)
if output_hidden_states:
encoder_states = encoder_states + (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
return BaseModelOutput(
last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
)
# Copied from transformers.models.time_series_transformer.modeling_time_series_transformer.TimeSeriesTransformerDecoder with TimeSeriesTransformer->Informer,TimeSeriesTransformerConfig->InformerConfig,time-series-transformer->informer,Transformer->Informer,TimeSeries->Informer
class InformerDecoder(InformerPreTrainedModel):
"""
Informer decoder consisting of *config.decoder_layers* layers. Each layer is a
[`InformerDecoderLayer`]
Args:
config: InformerConfig
"""
def __init__(self, config: InformerConfig):
super().__init__(config)
self.dropout = config.dropout
self.layerdrop = config.decoder_layerdrop
if config.prediction_length is None:
raise ValueError("The `prediction_length` config needs to be specified.")
self.value_embedding = InformerValueEmbedding(feature_size=config.feature_size, d_model=config.d_model)
self.embed_positions = InformerSinusoidalPositionalEmbedding(
config.context_length + config.prediction_length, config.d_model
)
self.layers = nn.ModuleList([InformerDecoderLayer(config) for _ in range(config.decoder_layers)])
self.layernorm_embedding = nn.LayerNorm(config.d_model)
self.gradient_checkpointing = False
# Initialize weights and apply final processing
self.post_init()
def forward(
self,
attention_mask: Optional[torch.Tensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
encoder_attention_mask: Optional[torch.LongTensor] = None,
head_mask: Optional[torch.Tensor] = None,
cross_attn_head_mask: Optional[torch.Tensor] = None,
past_key_values: Optional[List[torch.FloatTensor]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
r"""
Args:
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
of the decoder.
encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*):
Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
Mask to nullify selected heads of the cross-attention modules in the decoder to avoid performing
cross-attention on hidden heads. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
This is useful if you want more control over how to convert `input_ids` indices into associated vectors
than the model's internal embedding lookup matrix.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
for more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
use_cache = use_cache if use_cache is not None else self.config.use_cache
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
input_shape = inputs_embeds.size()[:-1]
# past_key_values_length
past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
attention_mask = _prepare_4d_causal_attention_mask(
attention_mask, input_shape, inputs_embeds, past_key_values_length
)
# expand encoder attention mask
if encoder_hidden_states is not None and encoder_attention_mask is not None:
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
encoder_attention_mask = _prepare_4d_attention_mask(
encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
)
hidden_states = self.value_embedding(inputs_embeds)
embed_pos = self.embed_positions(inputs_embeds.size(), past_key_values_length=self.config.context_length)
hidden_states = self.layernorm_embedding(hidden_states + embed_pos)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
if self.gradient_checkpointing and self.training:
if use_cache:
logger.warning_once(
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
)
use_cache = False
# decoder layers
all_hidden_states = () if output_hidden_states else None
all_self_attns = () if output_attentions else None
all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
next_decoder_cache = () if use_cache else None
# check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]):
if attn_mask is not None:
if attn_mask.size()[0] != (len(self.layers)):
raise ValueError(
f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
f" {head_mask.size()[0]}."
)
for idx, decoder_layer in enumerate(self.layers):
# add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
if output_hidden_states:
all_hidden_states += (hidden_states,)
if self.training:
dropout_probability = torch.rand([])
if dropout_probability < self.layerdrop:
continue
past_key_value = past_key_values[idx] if past_key_values is not None else None
if self.gradient_checkpointing and self.training:
layer_outputs = self._gradient_checkpointing_func(
decoder_layer.__call__,
hidden_states,
attention_mask,
encoder_hidden_states,
encoder_attention_mask,
head_mask[idx] if head_mask is not None else None,
cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
None,
output_attentions,
use_cache,
)
else:
layer_outputs = decoder_layer(
hidden_states,
attention_mask=attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
layer_head_mask=(head_mask[idx] if head_mask is not None else None),
cross_attn_layer_head_mask=(
cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None
),
past_key_value=past_key_value,
output_attentions=output_attentions,
use_cache=use_cache,
)
hidden_states = layer_outputs[0]
if use_cache:
next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)
if output_attentions:
all_self_attns += (layer_outputs[1],)
if encoder_hidden_states is not None:
all_cross_attentions += (layer_outputs[2],)
# add hidden states from the last decoder layer
if output_hidden_states:
all_hidden_states += (hidden_states,)
next_cache = next_decoder_cache if use_cache else None
if not return_dict:
return tuple(
v
for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
if v is not None
)
return BaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=hidden_states,
past_key_values=next_cache,
hidden_states=all_hidden_states,
attentions=all_self_attns,
cross_attentions=all_cross_attentions,
)
__all__ = ["InformerForPrediction", "InformerModel", "InformerPreTrainedModel"]
