TinyTimeMixerDataset

​

Overview

​

Class signature

class TinyTimeMixerDataset(BaseDataset):
    def __init__(
        self,
        name: str = None,
        datetime_col: str = "ds",
        path: str = None,
        boundaries: list = [0, 0, 0],
        batch_size: int = 128,
        mode: str = None,
        stride: int = 10,
        context_len: int = 512,
        horizon_len: int = 64,
    )

​

Parameters

​

Methods

​

__len__()

def __len__(self) -> int

​

__getitem__(index)

def __getitem__(self, index: int)

• index (int): Index of the data chunk.

• input_seq (np.ndarray): Input sequence of shape (num_channels, context_len)
• forecast_seq (np.ndarray): Forecast sequence of shape (num_channels, horizon_len)

​

get_data_loader()

def get_data_loader()

​

Example usage

from samay.dataset import TinyTimeMixerDataset

dataset = TinyTimeMixerDataset(
    path="data/timeseries.csv",
    datetime_col="date",
    context_len=512,
    horizon_len=96,
    batch_size=128,
    mode="train"
)

loader = dataset.get_data_loader()
for input_seq, forecast_seq in loader:
    # Training logic here
    # input_seq: (batch_size, n_channels, context_len)
    # forecast_seq: (batch_size, n_channels, horizon_len)
    pass

​

Features

• Efficient chunking for multivariate time series (max 64 channels per chunk)
• Automatic padding for short sequences
• High batch size support (default 128) for fast training
• Simple and clean data interface
• Automatic horizon length adjustment (max 30% of data length)

​

Data format

​

Input shape

​

Output shape

​

Notes

• The dataset automatically adjusts horizon length to be at most 30% of the data length
• Supports special boundary values: [-1, -1, -1] uses all data for training
• Data is chunked into groups of up to 64 channels for efficient processing
• No normalization is applied by default (raw data is used)