Model optimization (quantization, pruning) in PyTorch - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding Quantization Impact on Model Size

Which of the following statements best describes the main effect of applying post-training quantization to a PyTorch model?

A It increases the model size by adding extra layers to improve accuracy after training.

B It reduces the model size by converting weights from 32-bit floats to lower-bit integers, often 8-bit, with minimal accuracy loss.

C It removes neurons from the model to reduce size but always causes significant accuracy loss.

D It converts the model to use 64-bit floating point weights for higher precision.

💡 Hint

Think about how quantization changes the data type of weights and its effect on storage.

Quantization reduces model size by changing weights from 32-bit floats to smaller integer types like 8-bit integers, which saves memory and can speed up inference with little accuracy loss.

❓ Predict Output

intermediate

Output of Pruning Code on Model Parameters

Consider the following PyTorch code that applies pruning to a linear layer. What will be the number of zero weights in model.fc.weight after pruning?

PyTorch

import torch
import torch.nn as nn
import torch.nn.utils.prune as prune

class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc = nn.Linear(10, 5)

model = SimpleModel()
prune.l1_unstructured(model.fc, name='weight', amount=0.4)

zero_weights = torch.sum(model.fc.weight == 0).item()
print(zero_weights)

A 20

B 40

C 12

D 0

💡 Hint

Calculate total weights and 40% of them.

The linear layer has 10 inputs and 5 outputs, so 50 weights total. 40% of 50 is 20, so 20 weights are zeroed by pruning.

❓ Hyperparameter

advanced

Choosing Pruning Amount for Accuracy Preservation

You want to prune a PyTorch model to reduce size but keep accuracy loss under 2%. Which pruning amount is most likely to meet this goal?

A Pruning 5% of weights using L1 unstructured pruning

B Pruning 10% of weights using structured pruning on entire channels

C Pruning 50% of weights randomly without considering importance

D Pruning 80% of weights using unstructured pruning

💡 Hint

Smaller pruning amounts usually preserve accuracy better.

Pruning only 5% of weights with L1 unstructured pruning removes the least important weights, minimizing accuracy loss under 2%. Larger amounts or random pruning risk higher loss.

❓ Metrics

advanced

Evaluating Quantized Model Accuracy

After applying dynamic quantization to a PyTorch LSTM model, you observe the following accuracies on the test set:

• Original model accuracy: 92%
• Quantized model accuracy: 89%

What is the best interpretation of this result?

A The quantized model should have the same accuracy; the difference means the test data changed.

B Dynamic quantization always improves accuracy, so this result is unexpected and indicates a bug.

C Quantization caused the model to overfit, increasing accuracy on training but decreasing on test.

D Dynamic quantization caused a small accuracy drop but improved inference speed and reduced model size.

💡 Hint

Quantization trades off some accuracy for efficiency.

Dynamic quantization reduces model size and speeds up inference but can cause a small accuracy drop, which is normal and expected.

🔧 Debug

expert

Identifying Error in Pruning Application Code

What error will the following PyTorch code raise when trying to prune a model's convolutional layer?

PyTorch

import torch
import torch.nn as nn
import torch.nn.utils.prune as prune

class ConvModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv = nn.Conv2d(3, 16, 3, bias=False)

model = ConvModel()
prune.l1_unstructured(model.conv, name='bias', amount=0.3)

A No error; pruning applied successfully

B AttributeError: 'Conv2d' object has no attribute 'bias_mask'

C ValueError: Cannot prune parameter 'bias' because it does not exist or is not a tensor

D RuntimeError: Pruning amount must be between 0 and 1

💡 Hint

Check if the layer has a bias parameter by default.

The Conv2d layer by default has a bias parameter, but if bias=False is set, it does not. If bias is None, pruning 'bias' raises ValueError because the parameter does not exist.