TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.
Python
Quantiles PR nit fix
misha-chertushkin67e0eaaa3d2c
diff --git a/src/finetuning/finetuning_torch.py b/src/finetuning/finetuning_torch.pyindex 48f7dce..5c2d8b3 100644--- a/src/finetuning/finetuning_torch.py+++ b/src/finetuning/finetuning_torch.py@@ -13,7 +13,7 @@ import torch.distributed as distimport torch.nn as nnfrom torch.nn.parallel import DistributedDataParallel as DDPfrom torch.utils.data import DataLoader, Dataset-from timesfm.pytorch_patched_decoder import _create_quantiles+from timesfm.pytorch_patched_decoder import create_quantilesimport wandb@@ -267,7 +267,7 @@ class TimesFMFinetuner:loss = self.loss_fn(last_patch_pred, x_future.squeeze(-1))if self.config.use_quantile_loss:- quantiles = self.config.quantiles or _create_quantiles()+ quantiles = self.config.quantiles or create_quantiles()for i, quantile in enumerate(quantiles):last_patch_quantile = predictions[:, -1, :, i + 1]loss += torch.mean(diff --git a/src/timesfm/pytorch_patched_decoder.py b/src/timesfm/pytorch_patched_decoder.pyindex 67f6be4..15bf428 100644--- a/src/timesfm/pytorch_patched_decoder.py+++ b/src/timesfm/pytorch_patched_decoder.py@@ -21,7 +21,7 @@ from torch import nnimport torch.nn.functional as F-def _create_quantiles() -> list[float]:+def create_quantiles() -> list[float]:return [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]@@ -48,7 +48,7 @@ class TimesFMConfig:# Horizon lengthhorizon_len: int = 128# quantiles- quantiles: List[float] = dataclasses.field(default_factory=_create_quantiles)+ quantiles: List[float] = dataclasses.field(default_factory=create_quantiles)# Padding valuepad_val: float = 1123581321.0# Tolerance