custom-onecyclelr 0.2.0

A Custom PyTorch implementation of the OneCycleLR learning rate scheduler. (With some modifications)

OneCycle Learning Rate Scheduler

A custom implementation of the OneCycle learning rate scheduler for PyTorch.

Features

• Customized version of the OneCycleLR algorithm with four distinct phases: warmup, idling, annealing, and decay.
• Flexibility in defining various hyperparameters such as:
  • Warmup iterations and type (linear or exponential)
  • Idling period duration
  • Annealing phase duration and minimum learning rate
  • Decay phase duration and minimum learning rate
• Compatibility with any PyTorch optimizer

Installation

pip install custom-onecyclelr

Usage

Here's an example of how to integrate the scheduler into your training loop:

import torch
from custom_onecyclelr import scheduler

# Initialize model and optimizer
model = YourModel()
optimizer = torch.optim.SGD(model.parameters(), lr=1e-3)

# Create the OneCycleLR scheduler with desired parameters
scheduler_instance = scheduler.OneCycleLr(
    optimizer,
    warmup_iters=6,  # Number of iterations for the warmup phase
    lr_idling_iters=8,  # Number of iterations where learning rate remains at max
    annealing_iters=56,  # Cosine annealing phase duration
    decay_iters=100,  # Linear decay phase duration
    max_lr=0.01,
    annealing_lr_min=0.001,
    decay_lr_min=0.0001,
    warmup_start_lr=0.0001,
    warmup_type="exp"  # "linear" or "exp"
)

# You can also use the config_helper function to simplify the configuration (works best for higher total iterations)
# from custom_onecyclelr import config
# scheduler_instance = scheduler.OneCycleLr(
#     optimizer,
#     config.config_helper(200, 0.008)
# )

# Training loop
for epoch in range(total_epochs):
    for inputs, targets in dataloader:
        optimizer.zero_grad()
        outputs = model(inputs)
        loss = criterion(outputs, targets)
        loss.backward()
        optimizer.step()

        scheduler_instance.step()

Visualization

You can visualize how the learning rate changes over iterations by running:

python examples/vis.py

This will generate a plot showing the different phases of the learning rate schedule.

License

This project is licensed under MIT License - see LICENSE for details.