Gradient accumulation in PyTorch

Introduction

Gradient accumulation helps train models with large batch sizes by splitting them into smaller steps, saving memory and keeping training stable.

When your computer's memory is too small to handle a large batch at once.

When you want to simulate training with a big batch but only have resources for smaller batches.

When training large models that need stable updates but limited GPU memory.

When you want to improve training stability by averaging gradients over several mini-batches.

Syntax

PyTorch

optimizer.zero_grad()
for i, data in enumerate(dataloader):
    outputs = model(data.inputs)
    loss = loss_function(outputs, data.labels) / accumulation_steps
    loss.backward()
    if (i + 1) % accumulation_steps == 0:
        optimizer.step()
        optimizer.zero_grad()

Divide the loss by the number of accumulation steps to average gradients correctly.

Call optimizer.step() only after accumulating gradients over several mini-batches.

Examples

Accumulate gradients over 4 mini-batches before updating model weights.

PyTorch

accumulation_steps = 4
optimizer.zero_grad()
for i, (inputs, labels) in enumerate(dataloader):
    outputs = model(inputs)
    loss = criterion(outputs, labels) / accumulation_steps
    loss.backward()
    if (i + 1) % accumulation_steps == 0:
        optimizer.step()
        optimizer.zero_grad()

Here, gradients are accumulated over 2 batches to simulate a larger batch size.

PyTorch

accumulation_steps = 2
optimizer.zero_grad()
for i, batch in enumerate(dataloader):
    inputs, targets = batch
    outputs = model(inputs)
    loss = loss_fn(outputs, targets) / accumulation_steps
    loss.backward()
    if (i + 1) % accumulation_steps == 0:
        optimizer.step()
        optimizer.zero_grad()

Sample Model

This code trains a simple linear model on random data using gradient accumulation over 4 mini-batches before updating weights. It also handles the last few batches if the dataset size is not divisible by accumulation_steps.

PyTorch

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset

# Simple dataset
inputs = torch.randn(20, 5)
labels = torch.randn(20, 1)
dataset = TensorDataset(inputs, labels)
dataloader = DataLoader(dataset, batch_size=2)

# Simple model
model = nn.Linear(5, 1)
criterion = nn.MSELoss()
optimizer = optim.SGD(model.parameters(), lr=0.1)

accumulation_steps = 4
optimizer.zero_grad()

for i, (x, y) in enumerate(dataloader):
    outputs = model(x)
    loss = criterion(outputs, y) / accumulation_steps
    loss.backward()
    if (i + 1) % accumulation_steps == 0:
        optimizer.step()
        optimizer.zero_grad()

# Handle the case where the last batch is not a multiple of accumulation_steps
if (i + 1) % accumulation_steps != 0:
    optimizer.step()
    optimizer.zero_grad()

# Check final loss on the dataset
with torch.no_grad():
    preds = model(inputs)
    final_loss = criterion(preds, labels).item()

print(f"Final loss after gradient accumulation: {final_loss:.4f}")

Important Notes

Always divide the loss by accumulation_steps to keep gradient scale correct.

Remember to call optimizer.zero_grad() after each optimizer.step() to reset gradients.

If dataset size is not divisible by accumulation_steps, handle the last few batches carefully.

Summary

Gradient accumulation splits a large batch into smaller steps to save memory.

It averages gradients over multiple mini-batches before updating model weights.

Useful for training large models or when GPU memory is limited.

Practice

1. What is the main purpose of gradient accumulation in PyTorch training?

easy

A. To reduce the number of model parameters

B. To increase the learning rate automatically

C. To change the model architecture dynamically

D. To split a large batch into smaller steps to save memory

Solution

1. Step 1: Understand gradient accumulation concept

   Gradient accumulation splits a large batch into smaller mini-batches to avoid memory overflow.

2. Step 2: Identify the main benefit

   This technique helps save GPU memory by processing smaller batches sequentially before updating weights.

3. Final Answer:

   To split a large batch into smaller steps to save memory -> Option D

4. Quick Check:

   Gradient accumulation = batch splitting for memory saving [OK]

Hint: Gradient accumulation saves memory by splitting batches [OK]

Common Mistakes:

• Thinking it changes learning rate
• Confusing with model size reduction
• Assuming it alters model layers

2. Which PyTorch code snippet correctly implements gradient accumulation over 4 steps?

easy

A. for i, data in enumerate(dataloader):
outputs = model(data)
loss = criterion(outputs, labels)
optimizer.step()
loss.backward()
optimizer.zero_grad()

B. for i, data in enumerate(dataloader):
outputs = model(data)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
if (i + 1) % 4 == 0:
optimizer.zero_grad()

C. for i, data in enumerate(dataloader):
outputs = model(data)
loss = criterion(outputs, labels)
loss.backward()
if (i + 1) % 4 == 0:
optimizer.step()
optimizer.zero_grad()

D. for i, data in enumerate(dataloader):
outputs = model(data)
loss = criterion(outputs, labels)
optimizer.zero_grad()
loss.backward()
if (i + 1) % 4 == 0:
optimizer.step()

Solution

1. Step 1: Identify correct order of operations

   Gradient accumulation requires calling loss.backward() each step, then optimizer.step() and zero_grad() after accumulating gradients.

2. Step 2: Check modulo condition for accumulation

   Only after 4 mini-batches (i+1 % 4 == 0) should optimizer.step() and zero_grad() be called to update weights and reset gradients.

3. Final Answer:

   Code that calls loss.backward() every step and optimizer.step() plus zero_grad() every 4 steps -> Option C

4. Quick Check:

   Backward every step, step and zero_grad every 4 steps [OK]

Hint: Call backward every step, step and zero_grad every N steps [OK]

Common Mistakes:

• Calling optimizer.step() before loss.backward()
• Resetting gradients too early
• Not using modulo condition correctly

3. Given this PyTorch code with gradient accumulation steps=2, what is the printed loss after 4 batches?

optimizer.zero_grad()
for i, (inputs, labels) in enumerate(dataloader):
    outputs = model(inputs)
    loss = criterion(outputs, labels) / 2
    loss.backward()
    if (i + 1) % 2 == 0:
        optimizer.step()
        optimizer.zero_grad()
    print(f"Batch {i+1} loss: {loss.item()}")

medium

A. Loss values cause runtime error due to division

B. Loss values divided by 2 printed for each batch

C. Loss values accumulated and printed only every 2 batches

D. Loss values printed without division

Solution

1. Step 1: Understand loss scaling in code

   Loss is divided by 2 before backward to average gradients over 2 accumulation steps.

2. Step 2: Check print statement inside loop

   Loss.item() prints the scaled loss for each batch, so printed values are divided by 2 every batch.

3. Final Answer:

   Loss values divided by 2 printed for each batch -> Option B

4. Quick Check:

   Loss scaled by 2 and printed each batch [OK]

Hint: Loss is scaled before backward, printed as scaled every batch [OK]

Common Mistakes:

• Thinking loss prints only every 2 batches
• Ignoring loss division by 2
• Assuming runtime error from division

4. Identify the error in this gradient accumulation code snippet:

optimizer.zero_grad()
for i, (inputs, labels) in enumerate(dataloader):
    outputs = model(inputs)
    loss = criterion(outputs, labels)
    loss.backward()
    if i % 4 == 0:
        optimizer.step()
        optimizer.zero_grad()

medium

A. optimizer.step() and zero_grad() are called too early at i=0

B. loss.backward() is missing gradient scaling

C. optimizer.zero_grad() should be called before loss.backward()

D. The loop should use (i + 1) % 4 == 0 instead of i % 4 == 0

Solution

1. Step 1: Analyze the modulo condition

   Condition i % 4 == 0 triggers optimizer.step() and zero_grad() at i=0, which is before any gradients accumulate.

2. Step 2: Understand correct timing for optimizer step

   Optimizer step should happen after 4 batches, so condition should be (i + 1) % 4 == 0 to avoid early update.

3. Final Answer:

   optimizer.step() and zero_grad() are called too early at i=0 -> Option A

4. Quick Check:

   Modulo condition triggers step too early [OK]

Hint: Use (i + 1) % N == 0 to delay step until after N batches [OK]

Common Mistakes:

• Calling optimizer.step() at first batch
• Confusing zero_grad() timing
• Ignoring off-by-one in modulo condition

5. You want to train a large model on a GPU with limited memory using gradient accumulation with batch size 64 split into 8 steps. Which approach correctly adjusts the learning rate and loss scaling to keep training stable?

hard

A. Divide the loss by 8 before backward and keep the original learning rate

B. Multiply the learning rate by 8 and do not scale the loss

C. Divide the learning rate by 8 and multiply the loss by 8

D. Keep loss and learning rate unchanged, just accumulate gradients

Solution

1. Step 1: Understand gradient accumulation effect on loss

   Splitting batch into 8 steps means gradients accumulate over 8 mini-batches, so loss must be divided by 8 to average gradients correctly.

2. Step 2: Adjust learning rate accordingly

   Learning rate stays the same because gradient accumulation simulates a larger batch size without changing optimization dynamics.

3. Final Answer:

   Divide the loss by 8 before backward and keep the original learning rate -> Option A

4. Quick Check:

   Loss scaled down, learning rate unchanged for stable accumulation [OK]

Hint: Scale loss down by steps, keep learning rate same [OK]

Common Mistakes:

• Scaling learning rate instead of loss
• Not scaling loss causing gradient explosion
• Assuming no changes needed for accumulation