The optimizers repository provides a robust collection of advanced optimization algorithms for TensorFlow and Keras, designed to enhance machine learning and deep learning model training. These optimizers extend beyond traditional methods, implementing innovative techniques that help optimize performance across various tasks.

This repository includes:

Optimizers Overview

1. AdaBelief

A modification of the Adam optimizer that adapts learning rates based on gradient variability, making it effective for noisy gradients. Key features include:

• Rectification inspired by RAdam
• Weight decay and gradient clipping

Example Usage:

optimizer = AdaBelief(
    learning_rate=1e-3,
    weight_decay=1e-2,
    rectify=True
)

2. AdamP

AdamP is designed to mitigate weight norm increases in momentum-based optimizers, improving generalization and preventing overfitting through a projection step.

Example Usage:

optimizer = AdamP(
    learning_rate=1e-3,
    weight_decay=1e-2,
    delta=0.1,
    nesterov=True
)

3. LaProp

LaProp dynamically adjusts learning rates in proportion to gradients, supported by techniques such as centered moments and AMSGrad stabilization.

Example Usage:

optimizer = LaProp(
    learning_rate=4e-4,
    centered=True,
    weight_decay=1e-2
)

4. Lars

Layer-wise Adaptive Rate Scaling (LARS) optimizer for large-batch training, adapted for high-dimensional parameter models with features like momentum and trust-region scaling.

Example Usage:

optimizer = Lars(
    learning_rate=1.0,
    momentum=0.9,
    trust_coeff=0.001
)

5. MADGRAD

An advanced optimizer beneficial for training neural networks with sparse and dense gradients. It incorporates features for better optimization on large scales.

Example Usage:

optimizer = MADGRAD(
    learning_rate=1e-2,
    momentum=0.9
)

6. MARS

MARS employs variance reduction techniques for adaptive learning rates, ensuring effective parameter optimization across various scenarios.

Example Usage:

optimizer = Mars(
    learning_rate=3e-3,
    gamma=0.025,
    mars_type="adamw"
)

7. NAdam

Combining Nesterov momentum with adaptive moment estimation for faster convergence and improved optimization dynamics across a range of tasks.

Example Usage:

optimizer = NAdam(
    learning_rate=2e-3,
    schedule_decay=4e-3
)

8. NvNovoGrad

Layer-wise adaptive moments optimize for effective deep learning training, especially suitable for resource-constrained tasks.

Example Usage:

optimizer = NvNovoGrad(
    learning_rate=1e-3,
    grad_averaging=True
)

9. RAdam

Rectified Adam improves stability in early training phases with a variance rectification mechanism for adaptive learning rates.

Example Usage:

optimizer = RAdam(
    learning_rate=1e-3,
    weight_decay=1e-4
)

10. SGDP

Incorporates decoupled weight decay and gradient projection, designed for better convergence in stochastic optimization scenarios.

Example Usage:

optimizer = SGDP(
    learning_rate=1e-3,
    momentum=0.9
)

11. Adan

Adan integrates adaptive gradient estimation with multi-step momentum to accelerate training and improve convergence in deep learning models.

Example Usage:

optimizer = Adan(
    learning_rate=1e-3,
    beta1=0.98
)

12. Lamb

Enhances SGD for large-scale training by adapting learning rates layer-wise for better performance on deep neural networks.

Example Usage:

optimizer = Lamb(
    learning_rate=1e-3,
    trust_clip=True
)

This repository is designed to empower data scientists and machine learning engineers by providing state-of-the-art optimizers that can be seamlessly integrated into TensorFlow and Keras models, ensuring optimized performance and efficient training processes.