Zeroing gradients in PyTorch - Model Pipeline Trace

Model Pipeline - Zeroing gradients

This pipeline shows how zeroing gradients helps prepare a model for a new training step by clearing old gradient values. This prevents mixing old and new gradient information, which keeps training stable and accurate.

Data Flow - 6 Stages

1Data loading

1000 rows x 10 features→Load dataset with 1000 samples and 10 features each→1000 rows x 10 features

[[0.5, 1.2, ..., 0.3], [0.1, 0.4, ..., 0.9], ...]

↓

2Model input

1000 rows x 10 features→Feed input data to model→1000 rows x 1 output

[[0.7], [0.2], ...]

↓

3Loss calculation

1000 rows x 1 output→Calculate loss between predictions and true labels→Scalar loss value

0.45

↓

4Zeroing gradients

Model parameters with old gradients→Clear gradients from previous training step→Model parameters with zero gradients

All gradients set to 0 before backward pass

↓

5Backward pass

Scalar loss→Compute gradients of loss w.r.t. model parameters→Gradients stored in model parameters

Gradients like tensor([0.01, -0.02, ...])

↓

6Optimizer step

Model parameters with gradients→Update model parameters using gradients→Updated model parameters

Weights updated to new values

Training Trace - Epoch by Epoch

Loss
0.7 |*       
0.6 | *      
0.5 |  *     
0.4 |   *    
0.3 |    *   
0.2 |     *  
0.1 |       
    +---------
     1 2 3 4 5 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.65	0.60	Initial loss is high; accuracy is low as model starts learning
2	0.48	0.75	Loss decreases and accuracy improves after zeroing gradients each step
3	0.35	0.82	Continued improvement shows stable training with proper gradient management
4	0.28	0.87	Loss keeps decreasing; zeroing gradients prevents gradient accumulation
5	0.22	0.90	Model converges well with clear gradients each step

Prediction Trace - 6 Layers

Layer 1: Input layer

Layer 2: Forward pass

Layer 3: Loss calculation

Layer 4: Zeroing gradients

Layer 5: Backward pass

Layer 6: Optimizer step

Model Quiz - 3 Questions

Why do we zero gradients before the backward pass?

A To save memory by deleting gradients permanently

B To increase the gradient values for faster learning

C To clear old gradients so they don't add up with new ones

D To reset model weights to initial values

Key Insight

Zeroing gradients is a crucial step in training neural networks. It clears old gradient values before computing new ones, preventing unwanted accumulation. This keeps training stable and helps the model learn correctly over time.

Practice

1. Why do we need to zero gradients in PyTorch before calling loss.backward() during training?

easy

A. To initialize the model parameters randomly

B. To save memory by deleting the model weights

C. To speed up the training by skipping backward pass

D. To clear old gradients so new gradients are calculated correctly

Solution

1. Step 1: Understand gradient accumulation in PyTorch

   PyTorch accumulates gradients by default, so old gradients stay unless cleared.

2. Step 2: Purpose of zeroing gradients before backward pass

   Zeroing gradients ensures only the current batch's gradients are used for updating weights.

3. Final Answer:

   To clear old gradients so new gradients are calculated correctly -> Option D

4. Quick Check:

   Zero gradients = fresh gradient calculation [OK]

Hint: Always zero gradients before backward pass to avoid accumulation [OK]

Common Mistakes:

• Skipping zero_grad() causes wrong gradient sums
• Confusing zeroing gradients with resetting weights
• Thinking zero_grad() speeds training by skipping steps

2. Which of the following is the correct way to zero gradients for an optimizer named optimizer in PyTorch?

easy

A. optimizer.clear_grad()

B. optimizer.zero_gradients()

C. optimizer.zero_grad()

D. optimizer.reset_grad()

Solution

1. Step 1: Recall PyTorch optimizer API

   The correct method to zero gradients is zero_grad() on the optimizer object.

2. Step 2: Check method names for zeroing gradients

   Only zero_grad() is valid; others are incorrect or do not exist.

3. Final Answer:

   optimizer.zero_grad() -> Option C

4. Quick Check:

   Use zero_grad() method to clear gradients [OK]

Hint: Remember method name ends with '_grad', not '_gradients' [OK]

Common Mistakes:

• Using zero_gradients() which does not exist
• Confusing with clear_grad() or reset_grad() methods
• Calling zero_grad on model instead of optimizer

3. What will be the output of the following PyTorch code snippet?

import torch
x = torch.tensor([2.0], requires_grad=True)
optimizer = torch.optim.SGD([x], lr=0.1)
optimizer.zero_grad()
y = x * 3
loss = y ** 2
loss.backward()
print(x.grad.item())

medium

A. 36.0

B. 12.0

C. 18.0

D. 6.0

Solution

1. Step 1: Calculate loss and gradients

   y = 3 * x = 6.0; loss = y^2 = 36.0; gradient of loss wrt x is d(loss)/dx = 2*y * dy/dx = 2*6*3 = 36.

2. Step 2: Check printed gradient value

   x.grad will hold 36.0 after backward, but since x.grad is printed directly, it shows 36.0.

3. Final Answer:

   36.0 -> Option A

4. Quick Check:

   Gradient = 2*y*3 = 36.0 [OK]

Hint: Use chain rule: d(loss)/dx = 2*y * dy/dx [OK]

Common Mistakes:

• Calculating gradient as 12.0 by missing chain rule
• Confusing loss value with gradient value
• Forgetting to call zero_grad() before backward

4. Consider this training loop snippet:

for data, target in dataloader:
    output = model(data)
    loss = loss_fn(output, target)
    loss.backward()
    optimizer.step()
    optimizer.zero_grad()

What is the problem here?

medium

A. No problem, this is the correct order

B. Gradients are zeroed after optimizer.step(), causing stale gradients in next iteration

C. loss.backward() should be called after optimizer.step()

D. optimizer.zero_grad() should be called before loss.backward() to clear old gradients

Solution

1. Step 1: Understand correct order of zeroing gradients

   Gradients must be zeroed before calling loss.backward() to avoid accumulation.

2. Step 2: Identify error in given code

   Here, zero_grad() is called after optimizer.step(), so old gradients accumulate during backward pass.

3. Final Answer:

   Gradients are zeroed after optimizer.step(), causing stale gradients in next iteration -> Option B

4. Quick Check:

   Zero gradients before backward pass [OK]

Hint: Call zero_grad() before backward() to reset gradients [OK]

Common Mistakes:

• Calling zero_grad() after optimizer.step() instead of before backward()
• Swapping order of backward() and step()
• Assuming zero_grad() is optional

5. You want to accumulate gradients over 4 mini-batches before updating model weights to simulate a larger batch size. Which code snippet correctly implements this using zeroing gradients in PyTorch?

hard

A. optimizer.zero_grad()
for i, (data, target) in enumerate(dataloader):
output = model(data)
loss = loss_fn(output, target) / 4
loss.backward()
if (i + 1) % 4 == 0:
optimizer.step()
optimizer.zero_grad()

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

C. for i, (data, target) in enumerate(dataloader):
output = model(data)
loss = loss_fn(output, target)
loss.backward()
optimizer.step()
optimizer.zero_grad()

D. optimizer.zero_grad()
for data, target in dataloader:
output = model(data)
loss = loss_fn(output, target)
loss.backward()
optimizer.step()
optimizer.zero_grad()

Solution

1. Step 1: Understand gradient accumulation

   To accumulate gradients over multiple batches, zero_grad() is called once before loop, then backward() accumulates gradients.

2. Step 2: Check correct update and zeroing frequency

   After 4 batches, optimizer.step() updates weights, then zero_grad() clears gradients for next accumulation cycle.

3. Final Answer:

   Code snippet C correctly accumulates gradients and updates every 4 batches -> Option A

4. Quick Check:

   Zero once, accumulate 4 backward calls, then step and zero [OK]

Hint: Zero gradients once, step and zero every N batches to accumulate [OK]

Common Mistakes:

• Zeroing gradients inside loop every batch prevents accumulation
• Not dividing loss by accumulation steps causing large gradients
• Calling optimizer.step() every batch defeats accumulation purpose