Learning rate schedulers in PyTorch

Introduction

Learning rate schedulers help the model learn better by changing the speed of learning during training. This can make training faster and improve results.

When training a neural network and you want to start with a higher learning speed and slow down later.

When the model stops improving and you want to reduce the learning rate to fine-tune it.

When training for many epochs and you want to adjust the learning rate automatically.

When you want to avoid the model jumping around too much by lowering the learning rate gradually.

Syntax

PyTorch

scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.1)

for epoch in range(num_epochs):
    train(...)  # your training code
    scheduler.step()

optimizer is the optimizer you use for training, like Adam or SGD.

step_size is how many epochs before the learning rate changes.

Examples

Reduces learning rate by half every 5 epochs.

PyTorch

scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=5, gamma=0.5)

Reduces learning rate by 10% every epoch.

PyTorch

scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.9)

Changes learning rate following a cosine curve over 10 epochs.

PyTorch

scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=10)

Sample Model

This code trains a simple linear model on dummy data. The learning rate starts at 0.1 and drops by 10 times every 3 epochs. The print shows loss and current learning rate each epoch.

PyTorch

import torch
import torch.nn as nn
import torch.optim as optim

# Simple model
model = nn.Linear(2, 1)

# Optimizer
optimizer = optim.SGD(model.parameters(), lr=0.1)

# Scheduler: reduce LR by 0.1 every 3 epochs
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=3, gamma=0.1)

# Dummy data
inputs = torch.tensor([[1.0, 2.0], [3.0, 4.0]])
targets = torch.tensor([[1.0], [2.0]])

loss_fn = nn.MSELoss()

for epoch in range(6):
    optimizer.zero_grad()
    outputs = model(inputs)
    loss = loss_fn(outputs, targets)
    loss.backward()
    optimizer.step()
    scheduler.step()
    print(f"Epoch {epoch+1}, Loss: {loss.item():.4f}, LR: {scheduler.get_last_lr()[0]:.5f}")

Important Notes

Call scheduler.step() after each epoch to update the learning rate.

You can check the current learning rate with scheduler.get_last_lr().

Different schedulers change the learning rate in different ways; choose one that fits your training needs.

Summary

Learning rate schedulers adjust the learning speed during training to help the model learn better.

They are used to reduce the learning rate gradually or at specific times.

PyTorch provides many schedulers like StepLR, ExponentialLR, and CosineAnnealingLR.

Practice

1. What is the main purpose of using a learning rate scheduler in PyTorch training?

easy

A. To change the model architecture dynamically

B. To increase the batch size automatically

C. To shuffle the training data at each epoch

D. To adjust the learning rate during training for better model performance

Solution

1. Step 1: Understand the role of learning rate

   The learning rate controls how fast the model updates its knowledge during training.

2. Step 2: Identify what a scheduler does

   A learning rate scheduler changes the learning rate over time to improve training stability and performance.

3. Final Answer:

   To adjust the learning rate during training for better model performance -> Option D

4. Quick Check:

   Learning rate scheduler adjusts learning rate [OK]

Hint: Schedulers change learning rate, not batch size or model structure [OK]

Common Mistakes:

• Confusing scheduler with batch size adjustment
• Thinking scheduler changes model layers
• Assuming scheduler shuffles data

2. Which of the following is the correct way to create a StepLR scheduler in PyTorch for optimizer opt with step size 10 and gamma 0.1?

easy

A. scheduler = torch.optim.lr_scheduler.StepLR(opt, step_size=10, gamma=0.1)

B. scheduler = torch.optim.StepLR(opt, step=10, decay=0.1)

C. scheduler = torch.optim.lr_scheduler.StepLR(opt, steps=10, gamma=0.1)

D. scheduler = torch.optim.lr_scheduler.StepLR(opt, step_size=10, decay=0.1)

Solution

1. Step 1: Recall PyTorch StepLR syntax

   The correct class is torch.optim.lr_scheduler.StepLR with parameters step_size and gamma.

2. Step 2: Match parameters correctly

   step_size=10 and gamma=0.1 are the correct parameter names and values.

3. Final Answer:

   scheduler = torch.optim.lr_scheduler.StepLR(opt, step_size=10, gamma=0.1) -> Option A

4. Quick Check:

   StepLR uses step_size and gamma [OK]

Hint: Use exact parameter names: step_size and gamma [OK]

Common Mistakes:

• Using wrong parameter names like step or decay
• Calling StepLR from wrong module
• Mixing up parameter order

3. Given the code below, what will be the learning rate after 3 calls to scheduler.step()?

import torch
opt = torch.optim.SGD([torch.nn.Parameter(torch.randn(2, 2, requires_grad=True))], lr=0.1)
scheduler = torch.optim.lr_scheduler.StepLR(opt, step_size=2, gamma=0.5)

for _ in range(3):
    scheduler.step()
    current_lr = opt.param_groups[0]['lr']

medium

A. 0.05

B. 0.1

C. 0.025

D. 0.0125

Solution

1. Step 1: Understand StepLR behavior

   StepLR reduces learning rate by gamma every step_size epochs. Here, step_size=2, gamma=0.5.

2. Step 2: Calculate learning rate after 3 steps

   After 1 step: lr=0.1 (no change, step 1 < 2)
   After 2 steps: lr=0.1 * 0.5 = 0.05 (step 2 reached)
   After 3 steps: lr remains 0.05 (step 3 < 4)

3. Final Answer:

   0.05 -> Option A

4. Quick Check:

   StepLR halves lr every 2 steps [OK]

Hint: Learning rate changes only at multiples of step_size [OK]

Common Mistakes:

• Reducing learning rate every step instead of every step_size
• Multiplying gamma incorrectly
• Ignoring initial learning rate

4. Identify the error in the following PyTorch learning rate scheduler code:

import torch
opt = torch.optim.Adam([torch.nn.Parameter(torch.randn(3, 3, requires_grad=True))], lr=0.01)
scheduler = torch.optim.lr_scheduler.ExponentialLR(opt, gamma=0.9)

for epoch in range(5):
    scheduler.step()
    print(f"Epoch {epoch}: lr = {opt.param_groups[0]['lr']}")

medium

A. Learning rate should be set inside the loop

B. scheduler.step() should be called after optimizer.step()

C. ExponentialLR does not exist in PyTorch

D. gamma value must be greater than 1

Solution

1. Step 1: Recall correct scheduler usage

   In PyTorch, scheduler.step() should be called after optimizer.step() to update learning rate correctly.

2. Step 2: Check code order

   The code calls scheduler.step() before any optimizer.step(), which is incorrect and may cause unexpected lr updates.

3. Final Answer:

   scheduler.step() should be called after optimizer.step() -> Option B

4. Quick Check:

   Call scheduler.step() after optimizer.step() [OK]

Hint: Always call scheduler.step() after optimizer.step() [OK]

Common Mistakes:

• Calling scheduler.step() before optimizer.step()
• Using invalid gamma values
• Misunderstanding scheduler existence

5. You want to train a model where the learning rate starts at 0.1, then reduces by half every 5 epochs, but after 20 epochs, it should decay exponentially by 0.9 every epoch. Which PyTorch scheduler setup achieves this behavior?

hard

A. Use CosineAnnealingLR with T_max=20 and then StepLR with step_size=5, gamma=0.5

B. Use ExponentialLR with gamma=0.9 from start and manually adjust learning rate at epoch 20

C. Use StepLR with step_size=5, gamma=0.5 for first 20 epochs, then switch to ExponentialLR with gamma=0.9

D. Use StepLR with step_size=20, gamma=0.5 and ignore exponential decay

Solution

1. Step 1: Understand the two-phase learning rate schedule

   First phase: reduce lr by half every 5 epochs for 20 epochs.
   Second phase: after 20 epochs, apply exponential decay by 0.9 every epoch.

2. Step 2: Match PyTorch schedulers to phases

   StepLR with step_size=5, gamma=0.5 fits first phase.
   ExponentialLR with gamma=0.9 fits second phase.
   Switching schedulers after 20 epochs achieves desired behavior.

3. Final Answer:

   Use StepLR with step_size=5, gamma=0.5 for first 20 epochs, then switch to ExponentialLR with gamma=0.9 -> Option C

4. Quick Check:

   Combine StepLR then ExponentialLR for phased decay [OK]

Hint: Combine schedulers for multi-phase learning rate changes [OK]

Common Mistakes:

• Trying to use one scheduler for both phases
• Ignoring the switch at epoch 20
• Using wrong scheduler types for phases