Skip to main content

​
Class Signature

class TimesFM_2p5_Model(Basemodel):
    def __init__(self, config=None, repo=None, **kwargs)
The TimesFM_2p5_Model class implements TimesFM 2.5, the next generation of Google’s Time-Series Foundation Model with enhanced forecasting capabilities and improved accuracy.

​
Initialization Parameters

​
config
dict
default:"None"
Model configuration dictionary using ForecastConfig parameters. Required for model compilation.
​
repo
str
default:"None"
Hugging Face model repository ID. If provided, loads the pre-trained TimesFM 2.5 model. If not provided, initializes a new model instance.
​
**kwargs
dict
Additional keyword arguments passed to the base class.

​
Properties

​
quantiles
list[float]
Quantile levels configured in the model, accessed from model.model.config.quantiles.

​
Methods

​
evaluate()

def evaluate(
    dataset: TimesFM_2p5_Dataset,
    metric_only: bool = False,
    **kwargs
)
Evaluate the TimesFM 2.5 model on a dataset.
​
dataset
TimesFM_2p5_Dataset
required
Dataset containing the input data and relevant functions like dataloaders.
​
metric_only
bool
default:"False"
If True, return only metrics. If False, return metrics plus arrays.
​
return
Dict[str, float] | Tuple
When metric_only=True:Dictionary containing evaluation metrics:
  • mse: Mean Squared Error
  • mae: Mean Absolute Error
  • mase: Mean Absolute Scaled Error
  • mape: Mean Absolute Percentage Error
  • rmse: Root Mean Squared Error
  • nrmse: Normalized RMSE
  • smape: Symmetric Mean Absolute Percentage Error
  • msis: Mean Scaled Interval Score
  • nd: Normalized Deviation
  • mwsq: Mean Weighted Scaled Quantile Loss
  • crps: Continuous Ranked Probability Score
When metric_only=False:Tuple of (metrics, trues, preds, histories):
  • metrics: Dictionary containing evaluation metrics (as above)
  • trues: True values of shape (batch_size, n_channels, horizon_len)
  • preds: Predicted values of shape (batch_size, n_channels, horizon_len)
  • histories: Historical context values of shape (batch_size, n_channels, context_len)

​
Usage Example

from samay.model import TimesFM_2p5_Model
from samay.dataset import TimesFM_2p5_Dataset

# Load pre-trained model with config
config = {
    "context_len": 512,
    "horizon_len": 96,
    "input_patch_len": 32,
    "output_patch_len": 128,
    "quantiles": [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
}

model = TimesFM_2p5_Model(
    config=config,
    repo="google/timesfm-2.5-200m-pytorch"
)

# Prepare dataset
dataset = TimesFM_2p5_Dataset(
    name="ett",
    path="data/ETTh1.csv",
    datetime_col="date",
    context_len=512,
    horizon_len=96,
    mode="test"
)

# Evaluate with metrics only
metrics = model.evaluate(dataset, metric_only=True)
print(f"MSE: {metrics['mse']:.4f}")
print(f"MAE: {metrics['mae']:.4f}")
print(f"CRPS: {metrics['crps']:.4f}")

# Get full evaluation results
metrics, trues, preds, histories = model.evaluate(dataset, metric_only=False)

print(f"True values shape: {trues.shape}")
print(f"Predictions shape: {preds.shape}")
print(f"Histories shape: {histories.shape}")

​
Notes

  • TimesFM 2.5 requires a config dictionary with ForecastConfig parameters for model compilation
  • The model must be compiled with model.compile(config) after initialization
  • Supports probabilistic forecasting through quantile predictions
  • Improved over TimesFM 1.0 with better accuracy and faster inference
  • The model is evaluated in eval mode and moved to the appropriate device automatically
  • Context length and horizon length are determined by the dataset configuration

Build docs developers (and LLMs) love

Get started for freeTalk to us