LEFTeyex / warmup

A Warmup Scheduler for Pytorch to achieve the warmup learning rate at the beginning of training.

Warmup Scheduler Pytorch

Description

A Warmup Scheduler in Pytorch to make the learning rate change at the beginning of training for warmup.

Install

Notice: need to install pytorch>=1.1.0 manually.
The official website is PyTorch

Then install as follows:

pip install warmup_scheduler_pytorch

Usage

Detail to see GitHub example.py file.

import torch

from torch.optim import SGD  # example
from torch.optim.lr_scheduler import CosineAnnealingLR  # example

from warmup_scheduler_pytorch import WarmUpScheduler

model = Model()
optimizer = SGD(model.parameters(), lr=0.1)
lr_scheduler = CosineAnnealingLR(optimizer, T_max=100, eta_min=0.01)
data_loader = torch.utils.data.DataLoader(...)
warmup_scheduler = WarmUpScheduler(optimizer, lr_scheduler,
                                   len_loader=len(data_loader),
                                   warmup_steps=100,
                                   warmup_start_lr=0.01,
                                   warmup_mode='linear')
epochs = 100
for epoch in range(epochs):
    for batch_data in data_loader:
        output = model(...)
        # loss = loss_fn(output, ...)
        # loss.backward()
        optimizer.step()
        optimizer.zero_grad()
        warmup_scheduler.step()

    # lr_scheduler.step() is no longer needed