Warmup strategies in PyTorch

Introduction

Warmup strategies help the model start learning gently by slowly increasing the learning rate. This avoids big jumps that can confuse the model early on.

When training a deep neural network from scratch to avoid unstable updates.

When using a large learning rate that might be too strong at the start.

When fine-tuning a pretrained model to adapt smoothly to new data.

When training on a new dataset that is very different from previous data.

When you notice training loss jumping or not improving at the beginning.

Syntax

PyTorch

from torch.optim.lr_scheduler import LambdaLR

# Define a warmup function
def warmup_lambda(current_step):
    if current_step < warmup_steps:
        return float(current_step) / float(max(1, warmup_steps))
    return 1.0

# Create optimizer
optimizer = torch.optim.Adam(model.parameters(), lr=base_lr)

# Create scheduler with warmup
scheduler = LambdaLR(optimizer, lr_lambda=warmup_lambda)

The warmup function returns a multiplier for the learning rate.

During warmup steps, the multiplier grows from 0 to 1, then stays at 1.

Examples

This linearly increases learning rate from 0 to full over 1000 steps.

PyTorch

def warmup_lambda(step):
    return min(1.0, step / 1000)

Warmup for 500 steps, then keep learning rate constant.

PyTorch

def warmup_lambda(step):
    if step < 500:
        return step / 500
    else:
        return 1.0

Using a lambda function directly to warm up over 2000 steps.

PyTorch

scheduler = LambdaLR(optimizer, lr_lambda=lambda step: min(1.0, step / 2000))

Sample Model

This code trains a simple linear model with a warmup strategy for the learning rate over 5 steps. The learning rate starts at 0 and grows to 0.1 gradually, then stays constant.

PyTorch

import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim.lr_scheduler import LambdaLR

# Simple model
model = nn.Linear(10, 1)

# Parameters
base_lr = 0.1
warmup_steps = 5

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

# Warmup function

def warmup_lambda(step):
    if step < warmup_steps:
        return float(step) / float(max(1, warmup_steps))
    return 1.0

# Scheduler
scheduler = LambdaLR(optimizer, lr_lambda=warmup_lambda)

# Dummy data
inputs = torch.randn(10, 10)
targets = torch.randn(10, 1)

# Loss
criterion = nn.MSELoss()

print('Step | Learning Rate | Loss')
for step in range(10):
    optimizer.zero_grad()
    outputs = model(inputs)
    loss = criterion(outputs, targets)
    loss.backward()
    optimizer.step()
    scheduler.step()
    lr = optimizer.param_groups[0]['lr']
    print(f'{step:4d} | {lr:.4f}        | {loss.item():.4f}')

Important Notes

Warmup helps prevent the model from making large, unstable updates early in training.

You can combine warmup with other learning rate schedules for better results.

Adjust warmup_steps based on your dataset size and model complexity.

Summary

Warmup strategies gradually increase learning rate at the start of training.

This helps models learn smoothly and avoid unstable updates.

In PyTorch, LambdaLR with a custom function is a simple way to add warmup.

Practice

1. What is the main purpose of using a warmup strategy in PyTorch training?

easy

A. To immediately set the learning rate to its maximum value

B. To gradually increase the learning rate at the start of training

C. To decrease the learning rate throughout the entire training

D. To freeze model weights during the first epochs

Solution

1. Step 1: Understand what warmup means

   Warmup means starting with a low learning rate and increasing it slowly.

2. Step 2: Identify the goal of warmup

   This helps the model learn smoothly and avoid sudden big updates that can harm training.

3. Final Answer:

   To gradually increase the learning rate at the start of training -> Option B

4. Quick Check:

   Warmup = gradual learning rate increase [OK]

Hint: Warmup means slowly raising learning rate early [OK]

Common Mistakes:

• Thinking warmup immediately sets max learning rate
• Confusing warmup with learning rate decay
• Assuming warmup freezes model weights

2. Which PyTorch class is commonly used to implement a warmup learning rate schedule with a custom function?

easy

A. torch.optim.lr_scheduler.StepLR

B. torch.optim.lr_scheduler.ReduceLROnPlateau

C. torch.optim.lr_scheduler.LambdaLR

D. torch.optim.lr_scheduler.ExponentialLR

Solution

1. Step 1: Recall PyTorch schedulers for warmup

   LambdaLR allows defining a custom function to adjust learning rate.

2. Step 2: Match scheduler to warmup use

   Warmup needs a custom function to increase learning rate gradually, which LambdaLR supports.

3. Final Answer:

   torch.optim.lr_scheduler.LambdaLR -> Option C

4. Quick Check:

   Custom function scheduler = LambdaLR [OK]

Hint: LambdaLR lets you define custom learning rate changes [OK]

Common Mistakes:

• Choosing StepLR which uses fixed step decay
• Picking ReduceLROnPlateau which reacts to metrics
• Selecting ExponentialLR which decays exponentially

3. Given the following PyTorch code snippet, what will be the learning rate at epoch 3?

import torch
optimizer = torch.optim.SGD([torch.nn.Parameter(torch.randn(2, 2))], lr=0.1)

warmup_epochs = 5
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda epoch: min((epoch + 1) / warmup_epochs, 1))

for epoch in range(5):
    scheduler.step()
    print(f"Epoch {epoch+1} LR: {optimizer.param_groups[0]['lr']}")

medium

A. 0.06

B. 0.03

C. 0.10

D. 0.50

Solution

1. Step 1: Understand the lambda function for LR

   The lambda function returns (epoch+1)/5 until it reaches 1, scaling the base LR 0.1.

2. Step 2: Calculate LR at epoch 3 (0-based index)

   Epoch 3 means epoch=2, so LR factor = (2+1)/5 = 3/5 = 0.6. LR = 0.1 * 0.6 = 0.06.

3. Final Answer:

   0.06 -> Option A

4. Quick Check:

   Epoch 3 LR = 0.1 * 3/5 = 0.06 [OK]

Hint: Multiply base LR by (epoch+1)/warmup_epochs [OK]

Common Mistakes:

• Using epoch number directly without +1
• Confusing epoch index with count
• Assuming LR is constant during warmup

4. Identify the error in this PyTorch warmup scheduler code:

import torch
optimizer = torch.optim.Adam([torch.nn.Parameter(torch.randn(2, 2))], lr=0.01)
warmup_epochs = 3
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda epoch: epoch / warmup_epochs)

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

medium

A. The optimizer should be SGD, not Adam

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

C. The learning rate is not scaled by base LR

D. The lambda function returns 0 at epoch 0 causing zero LR

Solution

1. Step 1: Analyze lambda function behavior at epoch 0

   At epoch 0, lambda returns 0/3 = 0, so LR is zero, which stops learning initially.

2. Step 2: Understand why zero LR is a problem

   Zero LR means no weight updates, which can slow or stop training progress early.

3. Final Answer:

   The lambda function returns 0 at epoch 0 causing zero LR -> Option D

4. Quick Check:

   Epoch 0 LR = 0 causes no learning [OK]

Hint: Check if lambda returns zero at first epoch [OK]

Common Mistakes:

• Ignoring zero LR at start
• Thinking optimizer type causes error
• Confusing scheduler step order

5. You want to implement a warmup strategy that linearly increases the learning rate from 0 to 0.1 over 4 epochs, then keeps it constant. Which lr_lambda function correctly achieves this in PyTorch's LambdaLR?

hard

A. lambda epoch: min((epoch + 1) / 4, 1)

B. lambda epoch: epoch / 4

C. lambda epoch: 1 if epoch >= 4 else 0.1 * epoch

D. lambda epoch: (epoch + 1) * 0.1

Solution

1. Step 1: Understand the warmup goal

   Learning rate should increase linearly from 0 to 1 (scale factor) over 4 epochs, then stay at 1.

2. Step 2: Check each lambda function

   lambda epoch: min((epoch + 1) / 4, 1) uses min((epoch+1)/4, 1), which linearly increases from 0.25 to 1 by epoch 4, then stays at 1.

3. Final Answer:

   lambda epoch: min((epoch + 1) / 4, 1) -> Option A

4. Quick Check:

   Linear increase capped at 1 = lambda epoch: min((epoch + 1) / 4, 1) [OK]

Hint: Use min((epoch+1)/warmup_epochs, 1) for linear warmup [OK]

Common Mistakes:

• Not adding +1 to epoch causing zero start
• Multiplying by 0.1 inside lambda instead of base LR
• Using step function instead of linear increase