Pytorch Change the learning rate based on number of epochs

Question

When I set the learning rate and find the accuracy cannot increase after training few epochs

optimizer = optim.Adam(model.parameters(), lr = 1e-4)

n_epochs = 10
for i in range(n_epochs):

    // some training here

If I want to use a step decay: reduce the learning rate by a factor of 10 every 5 epochs, how can I do so?

Answer 1

You can use learning rate scheduler torch.optim.lr_scheduler.StepLR

import torch.optim.lr_scheduler.StepLR
scheduler = StepLR(optimizer, step_size=5, gamma=0.1)

Decays the learning rate of each parameter group by gamma every step_size epochs see docs here Example from docs

# Assuming optimizer uses lr = 0.05 for all groups
# lr = 0.05     if epoch < 30
# lr = 0.005    if 30 <= epoch < 60
# lr = 0.0005   if 60 <= epoch < 90
# ...
scheduler = StepLR(optimizer, step_size=30, gamma=0.1)
for epoch in range(100):
    train(...)
    validate(...)
    scheduler.step()

Example:

import torch
import torch.optim as optim

optimizer = optim.SGD([torch.rand((2,2), requires_grad=True)], lr=0.1)
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=5, gamma=0.1)

for epoch in range(1, 21):
    scheduler.step()
    print('Epoch-{0} lr: {1}'.format(epoch, optimizer.param_groups[0]['lr']))
    if epoch % 5 == 0:print()

Epoch-1 lr: 0.1
Epoch-2 lr: 0.1
Epoch-3 lr: 0.1
Epoch-4 lr: 0.1
Epoch-5 lr: 0.1

Epoch-6 lr: 0.010000000000000002
Epoch-7 lr: 0.010000000000000002
Epoch-8 lr: 0.010000000000000002
Epoch-9 lr: 0.010000000000000002
Epoch-10 lr: 0.010000000000000002

Epoch-11 lr: 0.0010000000000000002
Epoch-12 lr: 0.0010000000000000002
Epoch-13 lr: 0.0010000000000000002
Epoch-14 lr: 0.0010000000000000002
Epoch-15 lr: 0.0010000000000000002

Epoch-16 lr: 0.00010000000000000003
Epoch-17 lr: 0.00010000000000000003
Epoch-18 lr: 0.00010000000000000003
Epoch-19 lr: 0.00010000000000000003
Epoch-20 lr: 0.00010000000000000003

More on How to adjust Learning Rate - torch.optim.lr_scheduler provides several methods to adjust the learning rate based on the number of epochs.

Answer 2

I mostly prefer this one rather than available from torch module

def adjust_learning_rate_poly(optimizer, initial_lr, iteration, max_iter):
    """Sets the learning rate
    # Adapted from PyTorch Imagenet example:
    # https://github.com/pytorch/examples/blob/master/imagenet/main.py
    """
    lr = initial_lr * ( 1 - (iteration / max_iter)) * ( 1 - (iteration / max_iter))
    if ( lr < 1.0e-7 ):
      lr = 1.0e-7

    for param_group in optimizer.param_groups:
        param_group['lr'] = lr

    return lr

for epoch in range(0, max_epoch):
    lr = adjust_learning_rate_poly(optimizer, lr_rate, epoch, max_epoch)

where lr_rate is my starting learning_rate