Skip to content

TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.

Python

initial comming for pytorch patched decoder.

Rajat Sena775855d9fff
diff --git a/src/timesfm_torch/README.md b/src/timesfm_torch/README.md
new file mode 100644
index 0000000..7bd28d7
--- /dev/null
+++ b/src/timesfm_torch/README.md
@@ -0,0 +1,10 @@
+# Official Pytorch implementation of TimesFM
+
+TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google
+Research for time-series forecasting.
+
+* Paper: [A decoder-only foundation model for time-series forecasting](https://arxiv.org/abs/2310.10688), to appear in ICML 2024.
+* [Google Research blog](https://research.google/blog/a-decoder-only-foundation-model-for-time-series-forecasting/)
+* [Hugging Face checkpoint repo](https://huggingface.co/google/timesfm-1.0-200m)
+
+## Please stay tuned for all the same functionalities as the pax version.
diff --git a/src/timesfm_torch/pytorch_patched_decoder.py b/src/timesfm_torch/pytorch_patched_decoder.py
new file mode 100644
index 0000000..9032692
--- /dev/null
+++ b/src/timesfm_torch/pytorch_patched_decoder.py
@@ -0,0 +1,811 @@
+# Copyright 2024 Google LLC
+#
+# 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 version of patched decoder."""
+
+import dataclasses
+import math
+from typing import List, Tuple
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+
+def _create_quantiles() -> list[float]:
+ return [0.1, 0.25, 0.5, 0.75, 0.9]
+
+
+class TimesFMConfig:
+ """Config for initializing timesfm patched_decoder class."""
+
+ # The number of blocks in the model.
+ num_layers: int = 20
+ # The number of attention heads used in the attention layers of the model.
+ num_heads: int = 16
+ # The number of key-value heads for implementing attention.
+ num_kv_heads: int = 16
+ # The hidden size of the model.
+ hidden_size: int = 1280
+ # The dimension of the MLP representations.
+ intermediate_size: int = 1280
+ # The number of head dimensions.
+ head_dim: int = 80
+ # The epsilon used by the rms normalization layers.
+ rms_norm_eps: float = 1e-6
+ # Patch length
+ patch_len: int = 32
+ # Horizon length
+ horizon_len: int = 128
+ # quantiles
+ quantiles: List[float] = dataclasses.field(default_factory=_create_quantiles)
+ # Padding value
+ pad_val: float = 1123581321.0
+ # Tolerance
+ tolerance: float = 1e-6
+ # The dtype of the weights.
+ dtype: str = "bfloat32"
+ # use positional embedding
+ use_positional_embedding: bool = True
+
+
+def _masked_mean_std(
+ inputs: torch.Tensor,
+ padding: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+ """Calculates mean and standard deviation of `inputs` across axis 1.
+
+ It excludes values where `padding` is 1. This should reproduce
+ http://google3/learning/multipod/pax/foundation_ts/patched_decoder.py?cl=664903156&l=320
+
+ Args:
+ inputs: A PyTorch tensor of shape [b, n, p].
+ padding: A PyTorch tensor of shape [b, n, p] with values 0 or 1.
+
+ Returns:
+ A tuple containing the mean and standard deviation.
+ We return the statistics of the first patch with more than three non-padded
+ values.
+ """
+ # TODO(senrajat): will be tested as a part of preprocess input directly.
+ # Selecting the first patch with more than 3 unpadded values.
+ pad_sum = torch.sum(1 - padding, dim=2)
+
+ def _get_patch_index(arr: torch.Tensor):
+ indices = torch.argmax((arr >= 3).to(torch.int32), dim=1)
+ row_sum = (arr >= 3).to(torch.int32).sum(dim=1)
+ return torch.where(row_sum == 0, arr.shape[1] - 1, indices)
+
+ patch_indices = _get_patch_index(pad_sum)
+ bidxs = torch.arange(inputs.shape[0])
+
+ arr = inputs[bidxs, patch_indices, :]
+ pad = padding[bidxs, patch_indices, :]
+
+ # Create a mask where padding is 0
+ mask = 1 - pad
+
+ # Calculate the number of valid elements
+ num_valid_elements = torch.sum(mask, dim=1)
+ num_valid_elements = torch.where(
+ num_valid_elements == 0,
+ torch.tensor(1,
+ dtype=num_valid_elements.dtype,
+ device=num_valid_elements.device),
+ num_valid_elements,
+ )
+
+ # Calculate the masked sum and squared sum
+ masked_sum = torch.sum(arr * mask, dim=1)
+ masked_squared_sum = torch.sum((arr * mask)**2, dim=1)
+
+ # Calculate the masked mean and standard deviation
+ masked_mean = masked_sum / num_valid_elements
+ masked_var = masked_squared_sum / num_valid_elements - masked_mean**2
+ masked_var = torch.where(
+ masked_var < 0.0,
+ torch.tensor(0.0, dtype=masked_var.dtype, device=masked_var.device),
+ masked_var,
+ )
+ masked_std = torch.sqrt(masked_var)
+
+ return masked_mean, masked_std
+
+
+def _shift_padded_seq(mask: torch.Tensor, seq: torch.Tensor) -> torch.Tensor:
+ """Shifts rows of seq based on the first 0 in each row of the mask.
+
+ This should reproduce:
+ https://google3/learning/multipod/pax/foundation_ts/patched_decoder.py;l=125.
+
+ Args:
+ mask: mask tensor of shape [B, N]
+ seq: seq tensor of shape [B, N, P]
+
+ Returns:
+ Returns the shifted sequence.
+ """
+ batch_size, num_seq, feature_dim = seq.shape
+
+ new_mask: torch.BoolTensor = mask == 0
+
+ # Use argmax to find the first True value in each row
+ indices = new_mask.to(torch.int32).argmax(dim=1)
+
+ # Handle rows with all zeros
+ indices[~new_mask.any(dim=1)] = -1
+
+ # Create index ranges for each sequence in the batch
+ idx_range = (torch.arange(num_seq).to(
+ seq.device).unsqueeze(0).unsqueeze(-1).expand(batch_size, -1,
+ feature_dim))
+
+ # Calculate shifted indices for each element in each sequence
+ shifted_idx = (idx_range - indices[:, None, None]) % num_seq
+
+ # Gather values from seq using shifted indices
+ shifted_seq = seq.gather(1, shifted_idx)
+
+ return shifted_seq
+
+
+def get_large_negative_number(dtype: torch.dtype) -> torch.Tensor:
+ """Returns a large negative value for the given dtype."""
+ if dtype.is_floating_point:
+ dtype_max = torch.finfo(dtype).max
+ else:
+ dtype_max = torch.iinfo(dtype).max
+ return torch.tensor(-0.7 * dtype_max, dtype=dtype)
+
+
+def apply_mask_to_logits(logits: torch.Tensor,
+ mask: torch.Tensor) -> torch.Tensor:
+ """Applies a floating-point mask to a set of logits.
+
+ Args:
+ logits: A torch.Tensor of logit values.
+ mask: A torch.Tensor (float32) of mask values with the encoding described
+ in the function documentation.
+
+ Returns:
+ Masked logits.
+ """
+
+ min_value = get_large_negative_number(logits.dtype)
+
+ return torch.where((mask >= min_value * 0.5), logits, min_value)
+
+
+def convert_paddings_to_mask(
+ paddings: torch.Tensor, dtype: torch.dtype = torch.float32) -> torch.Tensor:
+ """Converts binary paddings to a logit mask ready to add to attention matrix.
+
+ It should mimic:
+ https://google3/third_party/py/praxis/layers/attentions.py;l=191
+
+ Args:
+ paddings: binary torch.Tensor of shape [B, T], with 1 denoting padding
+ token.
+ dtype: data type of the input.
+
+ Returns:
+ A torch.Tensor of shape [B, 1, 1, T] ready to add to attention logits.
+ """
+ attention_mask = paddings[:, None, None, :] # Equivalent to jnp.newaxis
+ attention_mask *= get_large_negative_number(dtype)
+ return attention_mask
+
+
+def causal_mask(input_t: torch.Tensor) -> torch.Tensor:
+ """Computes and returns causal mask.
+
+ It should mimic:
+ https://source.corp.google.com/piper///depot/google3/third_party/py/praxis/layers/attentions.py;l=77
+
+ Args:
+ input_t: A torch.Tensor of shape [B, T, D].
+
+ Returns:
+ An attention_mask torch.Tensor of shape [1, 1, T, T]. Attention mask has
+ already been converted to large negative values.
+ """
+ assert input_t.dtype.is_floating_point, input_t.dtype
+ large_negative_number = get_large_negative_number(input_t.dtype)
+ t = input_t.shape[1]
+ col_idx = torch.arange(t).unsqueeze(0).repeat(t, 1)
+ row_idx = torch.arange(t).unsqueeze(1).repeat(1, t)
+ mask = (row_idx < col_idx).to(input_t.dtype) * large_negative_number
+ return (mask.unsqueeze(0).unsqueeze(0).to(input_t.device)
+ ) # Equivalent to jnp.newaxis
+
+
+def merge_masks(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
+ """Merges 2 masks.
+
+ logscale mask is expected but 0/1 mask is also fine.
+ It should mimic:
+ https://google3/third_party/py/praxis/layers/attentions.py;l=127
+
+ Args:
+ a: torch.Tensor of shape [1|B, 1, 1|T, S].
+ b: torch.Tensor of shape [1|B, 1, 1|T, S].
+
+ Returns:
+ torch.Tensor of shape [1|B, 1, 1|T, S].
+ """
+
+ def expand_t(key_mask):
+ query_mask = key_mask.transpose(-1, -2) # Equivalent of jnp.transpose
+ return torch.minimum(query_mask, key_mask)
+
+ if a.shape[2] != b.shape[2]:
+ if a.shape[2] == 1:
+ a = expand_t(a)
+ else:
+ assert b.shape[2] == 1
+ b = expand_t(b)
+
+ assert a.shape[1:] == b.shape[1:], f"a.shape={a.shape}, b.shape={b.shape}."
+ return torch.minimum(a, b) # Element-wise minimum, similar to jnp.minimum
+
+
+class ResidualBlock(nn.Module):
+ """TimesFM residual block."""
+
+ def __init__(
+ self,
+ input_dims,
+ hidden_dims,
+ output_dims,
+ ):
+ super(ResidualBlock, self).__init__()
+ self.input_dims = input_dims
+ self.hidden_dims = hidden_dims
+ self.output_dims = output_dims
+
+ # Hidden Layer
+ self.hidden_layer = nn.Sequential(
+ nn.Linear(input_dims, hidden_dims),
+ nn.SiLU(),
+ )
+
+ # Output Layer
+ self.output_layer = nn.Linear(hidden_dims, output_dims)
+ # Residual Layer
+ self.residual_layer = nn.Linear(input_dims, output_dims)
+
+ def forward(self, x):
+ hidden = self.hidden_layer(x)
+ output = self.output_layer(hidden)
+ residual = self.residual_layer(x)
+ return output + residual
+
+
+class RMSNorm(torch.nn.Module):
+ """Pax rms norm in pytorch."""
+
+ def __init__(
+ self,
+ dim: int,
+ eps: float = 1e-6,
+ add_unit_offset: bool = False,
+ ):
+ super().__init__()
+ self.eps = eps
+ self.add_unit_offset = add_unit_offset
+ self.weight = nn.Parameter(torch.zeros(dim))
+
+ def _norm(self, x):
+ return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+ def forward(self, x):
+ output = self._norm(x.float())
+ if self.add_unit_offset:
+ output = output * (1 + self.weight.float())
+ else:
+ output = output * self.weight.float()
+ return output.type_as(x)
+
+
+class TransformerMLP(nn.Module):
+ """Pax transformer MLP in pytorch."""
+
+ def __init__(
+ self,
+ hidden_size: int,
+ intermediate_size: int,
+ ):
+ super().__init__()
+ self.gate_proj = nn.Linear(hidden_size, intermediate_size)
+ self.down_proj = nn.Linear(intermediate_size, hidden_size)
+ self.layer_norm = nn.LayerNorm(normalized_shape=hidden_size, eps=1e-6)
+
+ def forward(self, x, paddings=None):
+ gate_inp = self.layer_norm(x)
+ gate = self.gate_proj(gate_inp)
+ gate = F.relu(gate)
+ outputs = self.down_proj(gate)
+ if paddings is not None:
+ outputs = outputs * (1.0 - paddings[:, :, None])
+ return outputs + x
+
+
+class TimesFMAttention(nn.Module):
+ """Implements the attention used in TimesFM."""
+
+ def __init__(
+ self,
+ hidden_size: int,
+ num_heads: int,
+ num_kv_heads: int,
+ head_dim: int,
+ ):
+ super().__init__()
+
+ self.num_heads = num_heads
+ self.num_kv_heads = num_kv_heads
+
+ assert self.num_heads % self.num_kv_heads == 0
+ self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+ self.hidden_size = hidden_size
+ self.head_dim = head_dim
+
+ self.q_size = self.num_heads * self.head_dim
+ self.kv_size = self.num_kv_heads * self.head_dim
+ self.scaling = nn.Parameter(
+ torch.empty((self.head_dim,), dtype=torch.float32),)
+
+ self.qkv_proj = nn.Linear(
+ self.hidden_size,
+ (self.num_heads + 2 * self.num_kv_heads) * self.head_dim,
+ )
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size)
+
+ def _per_dim_scaling(self, query: torch.Tensor) -> torch.Tensor:
+ # [batch_size, n_local_heads, input_len, head_dim]
+ r_softplus_0 = 1.442695041
+ softplus_func = torch.nn.Softplus()
+ scale = r_softplus_0 / math.sqrt(self.head_dim)
+ scale = scale * softplus_func(self.scaling)
+ return query * scale[None, None, None, :]
+
+ def forward(
+ self,
+ hidden_states: torch.Tensor,
+ mask: torch.Tensor,
+ kv_write_indices: torch.Tensor | None = None,
+ kv_cache: Tuple[torch.Tensor, torch.Tensor] | None = None,
+ ) -> torch.Tensor:
+ hidden_states_shape = hidden_states.shape
+ assert len(hidden_states_shape) == 3
+
+ batch_size, input_len, _ = hidden_states_shape
+
+ qkv = self.qkv_proj(hidden_states)
+ xq, xk, xv = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+
+ xq = xq.view(batch_size, -1, self.num_heads, self.head_dim)
+ xk = xk.view(batch_size, -1, self.num_kv_heads, self.head_dim)
+ xv = xv.view(batch_size, -1, self.num_kv_heads, self.head_dim)
+ xq = self._per_dim_scaling(xq)
+
+ # Write new kv cache.
+ # [batch_size, input_len, n_local_kv_heads, head_dim]
+ if kv_cache is not None and kv_write_indices is not None:
+ k_cache, v_cache = kv_cache
+ k_cache.index_copy_(1, kv_write_indices, xk)
+ v_cache.index_copy_(1, kv_write_indices, xv)
+
+ key = k_cache
+ value = v_cache
+ else:
+ key = xk
+ value = xv
+ if self.num_kv_heads != self.num_heads:
+ # [batch_size, max_seq_len, n_local_heads, head_dim]
+ key = torch.repeat_interleave(key, self.num_queries_per_kv, dim=2)
+ value = torch.repeat_interleave(value, self.num_queries_per_kv, dim=2)
+
+ # [batch_size, n_local_heads, input_len, head_dim]
+ q = xq.transpose(1, 2)
+ # [batch_size, n_local_heads, max_seq_len, head_dim]
+ k = key.transpose(1, 2)
+ v = value.transpose(1, 2)
+
+ # [batch_size, n_local_heads, input_len, max_seq_len]
+ scores = torch.matmul(q, k.transpose(2, 3))
+ scores = scores + mask
+ scores = F.softmax(scores.float(), dim=-1).type_as(q)
+
+ # [batch_size, n_local_heads, input_len, head_dim]
+ output = torch.matmul(scores, v)
+ # return scores, output.transpose(1, 2).contiguous()
+
+ # [batch_size, input_len, hidden_dim]
+ output = output.transpose(1, 2).contiguous().view(batch_size, input_len, -1)
+ output = self.o_proj(output)
+ return scores, output
+
+
+class TimesFMDecoderLayer(nn.Module):
+ """Transformer layer."""
+
+ def __init__(
+ self,
+ hidden_size: int,
+ intermediate_size: int,
+ num_heads: int,
+ num_kv_heads: int,
+ head_dim: int,
+ rms_norm_eps: float = 1e-6,
+ ):
+ super().__init__()
+ self.self_attn = TimesFMAttention(
+ hidden_size=hidden_size,
+ num_heads=num_heads,
+ num_kv_heads=num_kv_heads,
+ head_dim=head_dim,
+ )
+ self.mlp = TransformerMLP(
+ hidden_size=hidden_size,
+ intermediate_size=intermediate_size,
+ )
+ self.input_layernorm = RMSNorm(hidden_size, eps=rms_norm_eps)
+
+ def forward(
+ self,
+ hidden_states: torch.Tensor,
+ mask: torch.Tensor,
+ paddings: torch.Tensor,
+ kv_write_indices: torch.Tensor | None = None,
+ kv_cache: Tuple[torch.Tensor, torch.Tensor] | None = None,
+ ) -> torch.Tensor:
+ # Self Attention
+ residual = hidden_states
+ hidden_states = self.input_layernorm(hidden_states)
+ scores, hidden_states = self.self_attn(
+ hidden_states=hidden_states,
+ mask=mask,
+ kv_write_indices=kv_write_indices,
+ kv_cache=kv_cache,
+ )
+ hidden_states = residual + hidden_states
+
+ # MLP
+ hidden_states = self.mlp(hidden_states, paddings=paddings)
+
+ return scores, hidden_states
+
+
+class StackedDecoder(nn.Module):
+ """Stacked transformer layer."""
+
+ def __init__(
+ self,
+ hidden_size: int,
+ intermediate_size: int,
+ num_heads: int,
+ num_kv_heads: int,
+ head_dim: int,
+ num_layers: int,
+ rms_norm_eps: float = 1e-6,
+ ):
+ super().__init__()
+
+ self.layers = nn.ModuleList()
+ for _ in range(num_layers):
+ self.layers.append(
+ TimesFMDecoderLayer(
+ hidden_size=hidden_size,
+ intermediate_size=intermediate_size,
+ num_heads=num_heads,
+ num_kv_heads=num_kv_heads,
+ head_dim=head_dim,
+ rms_norm_eps=rms_norm_eps,
+ ))
+
+ def forward(
+ self,
+ hidden_states: torch.Tensor,
+ paddings: torch.Tensor,
+ kv_write_indices: torch.Tensor | None = None,
+ kv_caches: List[Tuple[torch.Tensor, torch.Tensor]] | None = None,
+ ) -> torch.Tensor:
+ padding_mask = convert_paddings_to_mask(paddings, hidden_states.dtype)
+ atten_mask = causal_mask(hidden_states)
+ mask = merge_masks(padding_mask, atten_mask)
+ for i in range(len(self.layers)):
+ layer = self.layers[i]
+ kv_cache = kv_caches[i] if kv_caches is not None else None
+ _, hidden_states = layer(
+ hidden_states=hidden_states,
+ mask=mask,
+ paddings=paddings,
+ kv_write_indices=kv_write_indices,
+ kv_cache=kv_cache,
+ )
+ return hidden_states
+
+
+class PositionalEmbedding(torch.nn.Module):
+ """Generates position embedding for a given 1-d sequence.
+
+ Attributes:
+ min_timescale: Start of the geometric index. Determines the periodicity of
+ the added signal.
+ max_timescale: End of the geometric index. Determines the frequency of the
+ added signal.
+ embedding_dims: Dimension of the embedding to be generated.
+ """
+
+ def __init__(
+ self,
+ embedding_dims: int,
+ min_timescale: int = 1,
+ max_timescale: int = 10_000,
+ ) -> None:
+ super().__init__()
+ self.min_timescale = min_timescale
+ self.max_timescale = max_timescale
+ self.embedding_dims = embedding_dims
+
+ def forward(self, seq_length=None, position=None):
+ """Generates a Tensor of sinusoids with different frequencies.
+
+ Args:
+ seq_length: an optional Python int defining the output sequence length.
+ if the `position` argument is specified.
+ position: [B, seq_length], optional position for each token in the
+ sequence, only required when the sequence is packed.
+
+ Returns:
+ [B, seqlen, D] if `position` is specified, else [1, seqlen, D]
+ """
+ if position is None:
+ assert seq_length is not None
+ # [1, seqlen]
+ position = torch.arange(seq_length, dtype=torch.float32).unsqueeze(0)
+ else:
+ assert position.ndim == 2, position.shape
+
+ num_timescales = self.embedding_dims // 2
+ log_timescale_increment = math.log(
+ float(self.max_timescale) / float(self.min_timescale)) / max(
+ num_timescales - 1, 1)
+ inv_timescales = self.min_timescale * torch.exp(
+ torch.arange(num_timescales, dtype=torch.float32) *
+ -log_timescale_increment)
+ scaled_time = position.unsqueeze(2) * inv_timescales.unsqueeze(0).unsqueeze(
+ 0)
+ signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=2)
+ # Padding to ensure correct embedding dimension
+ signal = F.pad(signal, (0, 0, 0, self.embedding_dims % 2))
+ return signal
+
+
+class PatchedTimeSeriesDecoder(nn.Module):
+ """Patched time-series decoder."""
+
+ def __init__(self, config: TimesFMConfig):
+ super().__init__()
+ self.config = config
+ self.input_ff_layer = ResidualBlock(
+ input_dims=2 * config.patch_len,
+ output_dims=config.hidden_size,
+ hidden_dims=config.intermediate_size,
+ )
+ self.freq_emb = nn.Embedding(num_embeddings=3,
+ embedding_dim=config.hidden_size)
+ self.horizon_ff_layer = ResidualBlock(
+ input_dims=config.hidden_size,
+ output_dims=config.horizon_len * (1 + len(config.quantiles)),
+ hidden_dims=config.intermediate_size,
+ )
+ self.stacked_transformer = StackedDecoder(
+ hidden_size=self.config.hidden_size,
+ intermediate_size=self.config.intermediate_size,
+ num_heads=self.config.num_heads,
+ num_kv_heads=self.config.num_kv_heads,
+ head_dim=self.config.head_dim,
+ num_layers=self.config.num_layers,
+ rms_norm_eps=self.config.rms_norm_eps,
+ )
+ if self.config.use_positional_embedding:
+ self.position_emb = PositionalEmbedding(self.config.hidden_size)
+
+ def _forward_transform(
+ self, inputs: torch.Tensor, patched_pads: torch.Tensor
+ ) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
+ """Input is of shape [B, N, P]."""
+ mu, sigma = _masked_mean_std(inputs, patched_pads)
+ sigma = torch.where(
+ sigma < self.config.tolerance,
+ torch.tensor(1.0, dtype=sigma.dtype, device=sigma.device),
+ sigma,
+ )
+
+ # Normalize each patch
+ outputs = (inputs - mu[:, None, None]) / sigma[:, None, None]
+ outputs = torch.where(
+ torch.abs(inputs - self.config.pad_val) < self.config.tolerance,
+ torch.tensor(self.config.pad_val,
+ dtype=outputs.dtype,
+ device=outputs.device),
+ outputs,
+ )
+ return outputs, (mu, sigma)
+
+ def _reverse_transform(
+ self, outputs: torch.Tensor, stats: tuple[torch.Tensor,
+ torch.Tensor]) -> torch.Tensor:
+ """Output is of shape [B, N, P, Q]."""
+ mu, sigma = stats
+ return outputs * sigma[:, None, None, None] + mu[:, None, None, None]
+
+ def _preprocess_input(
+ self,
+ input_ts: torch.Tensor,
+ input_padding: torch.Tensor,
+ ) -> tuple[
+ torch.Tensor,
+ torch.Tensor,
+ tuple[torch.Tensor, torch.Tensor] | None,
+ torch.Tensor,
+ ]:
+ """Preprocess input for stacked transformer."""
+
+ # Reshape into patches (using view for efficiency)
+ bsize = input_ts.shape[0]
+ patched_inputs = input_ts.view(bsize, -1, self.config.patch_len)
+ patched_pads = input_padding.view(bsize, -1, self.config.patch_len)
+
+ patched_inputs = torch.where(
+ torch.abs(patched_pads - 1.0) < self.config.tolerance,
+ torch.tensor(0.0,
+ dtype=patched_inputs.dtype,
+ device=patched_inputs.device),
+ patched_inputs,
+ )
+ patched_pads = torch.where(
+ torch.abs(patched_inputs - self.config.pad_val) < self.config.tolerance,
+ torch.tensor(1.0, dtype=patched_pads.dtype, device=patched_pads.device),
+ patched_pads,
+ )
+ patched_inputs, stats = self._forward_transform(patched_inputs,
+ patched_pads)
+
+ # B x N x D
+ patched_inputs = patched_inputs * (1.0 - patched_pads)
+ concat_inputs = torch.cat([patched_inputs, patched_pads], dim=-1)
+ model_input = self.input_ff_layer(concat_inputs)
+
+ # A patch should not be padded even if there is at least one zero.
+ patched_padding = torch.min(patched_pads,
+ dim=-1)[0] # Get the values from the min result
+ if self.config.use_positional_embedding:
+ pos_emb = self.position_emb(model_input.shape[1]).to(model_input.device)
+ pos_emb = torch.concat([pos_emb] * model_input.shape[0], dim=0)
+ pos_emb = _shift_padded_seq(patched_padding, pos_emb)
+ model_input += pos_emb
+
+ return model_input, patched_padding, stats, patched_inputs
+
+ def _postprocess_output(
+ self,
+ model_output: torch.Tensor,
+ num_outputs: int,
+ stats: tuple[torch.Tensor, torch.Tensor],
+ ) -> torch.Tensor:
+ """Postprocess output of stacked transformer."""
+
+ # B x N x (H.Q)
+ output_ts = self.horizon_ff_layer(model_output)
+
+ # Reshape using view
+ b, n, _ = output_ts.shape
+ output_ts = output_ts.view(b, n, self.config.horizon_len, num_outputs)
+
+ return self._reverse_transform(output_ts, stats)
+
+ def forward(
+ self,
+ input_ts: torch.Tensor,
+ input_padding: torch.LongTensor,
+ freq: torch.Tensor,
+ ) -> torch.Tensor:
+ num_outputs = len(self.config.quantiles) + 1
+ model_input, patched_padding, stats, _ = self._preprocess_input(
+ input_ts=input_ts,
+ input_padding=input_padding,
+ )
+ f_emb = self.freq_emb(freq) # B x 1 x D
+ model_input += f_emb
+ model_output = self.stacked_transformer(model_input, patched_padding)
+
+ output_ts = self._postprocess_output(model_output, num_outputs, stats)
+ return output_ts
+
+ def decode(
+ self,
+ input_ts: torch.Tensor,
+ paddings: torch.Tensor,
+ freq: torch.LongTensor,
+ horizon_len: int,
+ output_patch_len: int | None = None,
+ max_len: int = 512,
+ return_forecast_on_context: bool = False,
+ ) -> tuple[torch.Tensor, torch.Tensor]:
+ """Auto-regressive decoding without caching.
+
+ Args:
+ input_ts: input time-series and paddings. Time-series shape B x C.
+ paddings: padding shape B x (C + H) where H is the prediction length.
+ freq: frequency shape B x 1
+ horizon_len: prediction length.
+ output_patch_len: output length to be fetched from one step of
+ auto-regressive decoding.
+ max_len: maximum training context length.
+ return_forecast_on_context: whether to return the model forecast on the
+ context except the first input patch.
+
+ Returns:
+ Tuple of two forecasting results:
+ - Point (mean) output predictions as a tensor with shape B x H'.
+ - Full predictions (mean and quantiles) as a tensor with shape
+ B x H' x (1 + # quantiles).
+ In particular, if return_forecast_on_context is True, H' is H plus
+ the forecastable context length, i.e. context_len - (first) patch_len.
+ """
+ final_out = input_ts
+ context_len = final_out.shape[1]
+ full_outputs = []
+ if paddings.shape[1] != final_out.shape[1] + horizon_len:
+ raise ValueError(
+ "Length of paddings must match length of input + horizon_len:"
+ f" {paddings.shape[1]} != {final_out.shape[1]} + {horizon_len}")
+ if output_patch_len is None:
+ output_patch_len = self.config.horizon_len
+ num_decode_patches = (horizon_len + output_patch_len -
+ 1) // output_patch_len
+ for step_index in range(num_decode_patches):
+ current_padding = paddings[:, 0:final_out.shape[1]]
+ input_ts = final_out[:, -max_len:]
+ input_padding = current_padding[:, -max_len:]
+ fprop_outputs = self(input_ts, input_padding, freq)
+ if return_forecast_on_context and step_index == 0:
+ # For the first decodings step, collect the model forecast on the
+ # context except the unavailable first input batch forecast.
+ new_full_ts = fprop_outputs[:, :-1, :self.config.patch_len, :]
+ new_full_ts = fprop_outputs.view(new_full_ts.size(0), -1,
+ new_full_ts.size(3))
+
+ full_outputs.append(new_full_ts)
+
+ # (full batch, last patch, output_patch_len, index of mean forecast = 0)
+ new_ts = fprop_outputs[:, -1, :output_patch_len, 0]
+ new_full_ts = fprop_outputs[:, -1, :output_patch_len, :]
+ # (full batch, last patch, output_patch_len, all output indices)
+ full_outputs.append(new_full_ts)
+ final_out = torch.concatenate([final_out, new_ts], axis=-1)
+
+ if return_forecast_on_context:
+ # `full_outputs` indexing starts at after the first input patch.
+ full_outputs = torch.concatenate(
+ full_outputs,
+ axis=1)[:, :(context_len - self.config.patch_len + horizon_len), :]
+ else:
+ # `full_outputs` indexing starts at the forecast horizon.
+ full_outputs = torch.concatenate(full_outputs, axis=1)[:,
+ 0:horizon_len, :]
+
+ return (full_outputs[:, :, 0], full_outputs)