Skip to main content

​
Overview

TinyTimeMixer is an efficient time series foundation model that uses MLPMixer architecture for fast and accurate forecasting with minimal computational requirements. Paper: TinyTimeMixers: Fast Pretrained Models for Enhanced Zero/Few-Shot Forecasting

​
Configuration

TinyTimeMixer configuration is typically loaded from a JSON file: 
from samay.utils import load_args

arg_path = "config/tinytimemixer.json"
args = load_args(arg_path)
Example configuration: 
{
    "repo": "ibm/tinytimemixer",
    "context_len": 512,
    "horizon_len": 96
}

​
Loading the Model

from samay.model import TinyTimeMixerModel
from samay.utils import load_args

arg_path = "config/tinytimemixer.json"
args = load_args(arg_path)
model = TinyTimeMixerModel(**args)

​
Loading Dataset

from samay.dataset import TinyTimeMixerDataset

train_dataset = TinyTimeMixerDataset(
    name="ett",
    datetime_col="date",
    path="./data/ETTh1.csv",
    mode="train",
    batch_size=64,
    context_len=512,
    horizon_len=96
)

val_dataset = TinyTimeMixerDataset(
    name="ett",
    datetime_col="date",
    path="./data/ETTh1.csv",
    mode="test",
    batch_size=64,
    context_len=512,
    horizon_len=96
)

​
Zero-Shot Forecasting

metrics = model.evaluate(val_dataset, metric_only=True)
print(metrics)
# {'mse': ..., 'mae': ..., 'mase': ..., 'rmse': ..., ...}

​
Fine-tuning

finetuned_model = model.finetune(train_dataset)

# Evaluate after fine-tuning
metrics = model.evaluate(val_dataset, metric_only=True)

​
Visualization

model.plot(val_dataset)

​
Key Features

  • Lightweight: Significantly smaller than transformer-based models
  • Fast Inference: Quick predictions suitable for real-time applications
  • Efficient Training: Requires less computational resources
  • Good Zero-Shot Performance: Competitive accuracy without fine-tuning

​
Advantages

  1. Speed: Much faster than transformer-based models
  2. Memory Efficient: Lower memory footprint
  3. Easy to Deploy: Suitable for edge devices and production environments
  4. Good Accuracy: Maintains competitive performance despite smaller size

​
When to Use TinyTimeMixer

  • When you need fast inference times
  • When computational resources are limited
  • For real-time forecasting applications
  • When model size is a constraint (e.g., edge deployment)

​
Example Notebook

For a complete working example, see:

Build docs developers (and LLMs) love

Get started for freeTalk to us