Skip to content

Latest commit

 

History

History
125 lines (78 loc) · 6.23 KB

CHANGELOG.md

File metadata and controls

125 lines (78 loc) · 6.23 KB
Raw

Changelog

0.1.20

🚀 Feature Enhancements

Enhanced Model Fine-tuning with finetune_loss and finetune_steps 🛠️

The latest update brings a significant enhancement to the fine-tuning capabilities of our forecasting models. With the introduction of the finetune_loss, users now have the ability to not only specify the number of steps for fine-tunning (with finetune_steps) but also to define the target loss for fine-tunning, offering more granular control over model optimization.

  • finetune_loss Options:

    • default: Adopts the model's preset loss function, optimized during initial training.
    • mae (Mean Absolute Error): Focuses on the mean of the absolute differences between predicted and actual values.
    • mse (Mean Squared Error): Emphasizes the mean of the squares of the differences between predicted and actual values.
    • rmse (Root Mean Squared Error): Provides the square root of MSE, offering error terms in the same units as the predictions.
    • mape (Mean Absolute Percentage Error): Measures the mean absolute percent difference between predicted and actual values.
    • smape (Symmetric Mean Absolute Percentage Error): Offers a symmetric version of MAPE, ensuring equal treatment of over and underestimations.

  • finetune_steps: Determines the number of steps to execute during the fine-tuning process. It is crucial to set finetune_steps to a value greater than 0 to activate the fine-tuning mechanism with the chosen finetune_loss function. This allows for a more tailored optimization, aligning the model closely with specific forecasting requirements and improving its predictive performance.

# Configure model fine-tuning with custom loss function and steps
fcst_df = timegpt.forecast(df, model='timegpt-1-long-horizon', finetune_loss='mape', finetune_steps=50)

# Apply fine-tuning to cross-validation for enhanced model validation
cv_df = timegpt.cross_validation(df, model='timegpt-1', finetune_loss='smape', finetune_steps=50)

This update opens up new possibilities for refining forecasting models, ensuring they are finely tuned to the specific characteristics and challenges of the forecasting task at hand.

See full changelog here.

0.1.19

🚀 Feature Enhancements

Advanced Data Partitioning with num_partitions 🔄

We're excited to introduce the num_partitions argument for our forecast, cross_validation, and detect_anomalies methods, offering more control over data processing and parallelization:

  • Optimized Resource Utilization in Distributed Environments: For Spark, Ray, or Dask dataframes, num_partitions enables the system to either leverage all available parallel resources or to specify the number of parallel processes. This ensures efficient resource allocation and utilization across distributed computing environments.
# Utilize num_partitions in distributed environments
fcst_df = timegpt.forecast(df, model='timegpt-1-long-horizon', num_partitions=10)
  • Efficient Handling of Large Pandas Dataframes: When working with Pandas dataframes, num_partitions groups series into specified partitions, allowing for sequential API calls. This is particularly useful for large dataframes that are impractical to send over the internet in one go, enhancing performance and efficiency.
# Efficiently process large Pandas dataframes
cv_df = timegpt.cross_validation(df, model='timegpt-1', num_partitions=5)

This new feature provides a flexible approach to handling data across different environments, ensuring optimal performance and resource management.

See full changelog here.

0.1.18

🚀 Feature Enhancements

Forecast Using Diverse Models 🌐

Release of new forecasting methods. Among the updates, we've unveiled the timegpt-1-long-horizon model, crafted specifically for long-term forecasts that span multiple seasonalities. To use it, simply specify the model in your methods like so:

from nixtlats import TimeGPT

# Initialize the TimeGPT model
timegpt = TimeGPT()

# Generate forecasts using the long-horizon model
fcst_df = timegpt.forecast(..., model='timegpt-1-long-horizon')

# Perform cross-validation with the long-horizon model
cv_df = timegpt.cross_validation(..., model='timegpt-1-long-horizon')

# Detect anomalies with the long-horizon model
anomalies_df = timegpt.detect_anomalies(..., model='timegpt-1-long-horizon')

Choose between timegpt-1 for the first version of TimeGPT or timegpt-1-long-horizon for long horizon tasks..

Cross-Validation Methodology 📊

You can dive deeper into your forecasting pipelines with the new cross_validation feature. This method enables you to validate forecasts across different windows efficiently:

# Set up cross-validation with a custom horizon, number of windows, and step size
cv_df = timegpt.cross_validation(df, h=35, n_windows=5, step_size=5)

This will generate 5 distinct forecast sets, each with a horizon of 35, stepping through your data every 5 timestamps.

🔁 Retry Behavior for Robust API Calls

The new retry mechanism allows the making of more robust API calls (preventing them from crashing with large-scale tasks).

  • max_retries: Number of max retries for an API call.
  • retry_interval: Pause between retries.
  • max_wait_time: Total duration of retries.
timegpt = TimeGPT(max_retries=10, retry_interval=5, max_wait_time=360)

🔑 Token Inference Made Simple

The TimeGPT class now automatically infers your TIMEGPT_TOKEN using os.environ.get('TIMEGPT_TOKEN'), streamlining your setup:

# No more manual token handling - TimeGPT has got you covered
timegpt = TimeGPT()

For more information visit our FAQS section.

📘 Introducing the FAQ Section

Questions? We've got answers! Our new FAQ section tackles the most common inquiries, from integrating exogenous variables to configuring authorization tokens and understanding long-horizon forecasts.

See full changelog here.