Model serialization formats (pickle, ONNX, TorchScript) in MLOps - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding model serialization formats

Which of the following statements correctly describes the main difference between pickle and ONNX model serialization formats?

A Pickle serializes Python objects including models, but ONNX is a platform-independent format for representing models to run on different frameworks.

B Pickle is a binary format for GPU acceleration, while ONNX only works with CPU models.

C Pickle converts models to JSON format, whereas ONNX saves models as plain text files.

D Pickle is used only for TensorFlow models, and ONNX is exclusive to PyTorch.

💡 Hint

Think about portability and framework independence.

Pickle saves Python objects as they are, which makes them Python-specific and less portable. ONNX is designed to be a universal format that can be used across different frameworks and platforms.

💻 Command Output

intermediate

TorchScript model saving output

What is the expected output when running the following PyTorch code snippet?

import torch
import torchvision.models as models
model = models.resnet18(pretrained=False)
scripted_model = torch.jit.script(model)
torch.jit.save(scripted_model, 'resnet18_scripted.pt')
print('Model saved successfully')

A Model saved successfully

B TypeError: 'ResNet' object is not callable

C FileNotFoundError: [Errno 2] No such file or directory: 'resnet18_scripted.pt'

D RuntimeError: torch.jit.script failed due to unsupported operation

💡 Hint

Consider what the print statement outputs after saving.

The code scripts the model and saves it to a file. If no errors occur, the print statement outputs the success message.

❓ Troubleshoot

advanced

Troubleshooting ONNX model loading error

You exported a PyTorch model to ONNX format but get this error when loading it in another framework: ValueError: Unsupported ONNX opset version. What is the most likely cause?

A The PyTorch model was not scripted before exporting to ONNX.

B The model file is corrupted due to incomplete download.

C The ONNX runtime is not installed on the system.

D The ONNX model was exported with a newer opset version than the target framework supports.

💡 Hint

Opset versions define supported operations in ONNX.

If the ONNX model uses a newer opset version than the runtime supports, it cannot load the model properly, causing this error.

🔀 Workflow

advanced

Choosing the right serialization format for deployment

You have a PyTorch model that you want to deploy on a mobile device with limited resources and no Python environment. Which serialization format is best suited for this deployment?

A Saving the raw PyTorch model weights as a .pth file.

B TorchScript, because it compiles the model into a standalone format runnable without Python.

C ONNX, because it requires a Python interpreter to run.

D Pickle, because it preserves all Python objects and dependencies.

💡 Hint

Consider the environment constraints and runtime requirements.

TorchScript compiles models into a format that can run independently of Python, ideal for mobile deployment. Pickle requires Python, ONNX needs compatible runtime, and raw weights need Python code to load.

✅ Best Practice

expert

Best practice for model versioning with serialization formats

When managing multiple versions of machine learning models serialized with different formats (pickle, ONNX, TorchScript), what is the best practice to ensure reproducibility and smooth deployment?

A Convert all models to ONNX format regardless of original framework to unify deployment.

B Only keep the latest model version in pickle format to reduce storage space.

C Store each model version with metadata including format, framework version, and opset version in a version control system or model registry.

D Avoid storing metadata and rely on file timestamps to track versions.

💡 Hint

Think about traceability and compatibility.

Storing detailed metadata with each model version helps track compatibility, reproduce results, and manage deployments effectively. Ignoring metadata or relying on timestamps can cause confusion and errors.