Skip to main content

​
Class Signature

class TimesFM_2p5_Dataset(BaseDataset):
    def __init__(
        name: str = None,
        datetime_col: str = None,
        path: str = None,
        batch_size: int = 16,
        mode: str = "train",
        boundaries: list = [0, 0, 0],
        task_name: str = "evaluation",
        stride: int = 10,
        context_len: int = 512,
        horizon_len: int = 96,
        normalize: bool = True,
        **kwargs
    )
Dataset wrapper for preparing time-series data for the TimesFM 2.5 model.

​
Parameters

​
name
str
default:"None"
Dataset name (e.g., “ett”, “etth1”).
​
datetime_col
str
default:"None"
Name of the datetime column in the CSV file.
​
path
str
default:"None"
Path to the CSV file containing time-series data.
​
batch_size
int
default:"16"
Batch size for dataloaders.
​
mode
str
default:"train"
Dataset mode: “train” or “test”.
​
boundaries
list[int]
default:"[0, 0, 0]"
Train/val/test split boundaries. If [0,0,0], uses 50%/20%/30% split.
​
task_name
str
default:"evaluation"
Task name: “evaluation” or “finetuning”.
​
stride
int
default:"10"
Stride for windowing when creating sequences.
​
context_len
int
default:"512"
Length of historical context to use for forecasting.
​
horizon_len
int
default:"96"
Forecast horizon length. Automatically capped at 30% of dataset length.
​
normalize
bool
default:"True"
Whether to normalize the data.

​
Attributes

​
n_channels
int
Number of channels (columns) in the time series, computed as df.shape[1] - 1.

​
Methods

​
get_data_loader()

Returns a PyTorch DataLoader for the dataset. 
def get_data_loader()
​
return
DataLoader
PyTorch DataLoader configured with the specified batch size.

​
len()

Returns the number of samples in the dataset. 
def __len__()
​
return
int
Total number of samples available.

​
Usage Example

from samay.dataset import TimesFM_2p5_Dataset

# Create dataset for evaluation
eval_dataset = TimesFM_2p5_Dataset(
    name="ett",
    path="data/ETTh1.csv",
    datetime_col="date",
    mode="test",
    task_name="evaluation",
    context_len=512,
    horizon_len=96,
    batch_size=32,
    normalize=True
)

# Create dataset for training
train_dataset = TimesFM_2p5_Dataset(
    name="ett",
    path="data/ETTh1.csv",
    datetime_col="date",
    mode="train",
    task_name="finetuning",
    context_len=512,
    horizon_len=96,
    batch_size=16,
    normalize=True,
    stride=10
)

# Get dataloader
dataloader = eval_dataset.get_data_loader()

print(f"Dataset size: {len(eval_dataset)}")
print(f"Number of channels: {eval_dataset.n_channels}")
print(f"Context length: {eval_dataset.context_len}")
print(f"Horizon length: {eval_dataset.horizon_len}")

​
Notes

  • Horizon length is automatically capped at 30% of the dataset length
  • Default boundaries use 50%/20%/30% split for train/val/test if not specified
  • Data normalization is applied by default for better model performance
  • Datetime column is dropped after reading if specified
  • Supports both evaluation and fine-tuning tasks
  • Windowing with stride creates overlapping sequences for training

Build docs developers (and LLMs) love

Get started for freeTalk to us