TorchScript export in PyTorch - Model Metrics & Evaluation

Metrics & Evaluation - TorchScript export

Which metric matters for TorchScript export and WHY

TorchScript export is about saving a PyTorch model so it can run fast and independently from Python. The key metric to check after export is model output consistency. This means the exported model should give the same predictions as the original PyTorch model. We check this by comparing outputs on the same input data.

Additionally, inference speed is important because TorchScript aims to make models faster and easier to deploy.

Confusion matrix or equivalent visualization

Since TorchScript export is about model saving and running, we don't use a confusion matrix here. Instead, we compare outputs before and after export.

Original model output: [0.8, 0.1, 0.1]
Exported model output: [0.8, 0.1, 0.1]

Difference (L2 norm): 0.0001 (very small)
    

This shows the exported model predicts almost exactly the same as the original.

Tradeoff: Output consistency vs export flexibility

When exporting with TorchScript, you can choose tracing or scripting. Tracing is fast but may miss some dynamic code paths, causing output differences. Scripting handles dynamic code better but can be more complex.

If you want perfect output match, scripting is safer. If you want faster export and your model is simple, tracing might be enough.

What "good" vs "bad" output consistency looks like

Good: The exported model outputs differ from the original by less than 1e-5 on test inputs. Predictions and probabilities match closely.

Bad: Large differences in outputs or different predicted classes. This means the export failed to capture model logic correctly.

Common pitfalls when exporting with TorchScript

• Using tracing on models with dynamic control flow causes wrong outputs.
• Not testing exported model outputs on varied inputs.
• Ignoring differences in floating point precision or device (CPU vs GPU).
• Overlooking that some PyTorch operations are not supported in TorchScript.

Self-check question

Your original PyTorch model and TorchScript exported model give 98% accuracy on test data, but the exported model's outputs differ significantly on some inputs. Is this export good for production?

Answer: No. Even if accuracy looks high, large output differences mean the exported model may behave unpredictably on new data. You should fix export issues to ensure consistent predictions.

Key Result

Output consistency between original and exported model is key to trust TorchScript export.

Practice

1. What is the main purpose of exporting a PyTorch model using TorchScript?

easy

A. To increase the training speed of the model

B. To save the model so it can run independently without Python

C. To convert the model into a TensorFlow format

D. To visualize the model architecture

Solution

1. Step 1: Understand TorchScript export purpose

   TorchScript export saves PyTorch models in a format that can run without Python, making deployment easier.

2. Step 2: Compare options with purpose

   Only To save the model so it can run independently without Python correctly states the main purpose: saving for standalone use without Python.

3. Final Answer:

   To save the model so it can run independently without Python -> Option B

4. Quick Check:

   TorchScript export = standalone model saving [OK]

Hint: TorchScript export = run model without Python [OK]

Common Mistakes:

• Thinking it speeds up training
• Confusing with TensorFlow conversion
• Assuming it is for visualization

2. Which of the following is the correct way to export a PyTorch model using scripting in TorchScript?

easy

A. torch.jit.trace(model, example_input)

B. torch.load('model.pt')

C. torch.save(model.state_dict(), 'model.pt')

D. torch.jit.script(model)

Solution

1. Step 1: Identify scripting syntax

   Using scripting to export a model requires torch.jit.script(model).

2. Step 2: Differentiate from tracing and saving

   torch.jit.trace(model, example_input) is tracing, torch.save(model.state_dict(), 'model.pt') saves weights only, torch.load('model.pt') loads a model, so only torch.jit.script(model) is correct for scripting export.

3. Final Answer:

   torch.jit.script(model) -> Option D

4. Quick Check:

   Scripting export uses torch.jit.script [OK]

Hint: Scripting export uses torch.jit.script(model) [OK]

Common Mistakes:

• Confusing scripting with tracing
• Using torch.save instead of torch.jit.script
• Trying to load instead of export

3. Given the code below, what will be the output of print(traced_model(torch.tensor([2.0])))?

import torch
class SimpleModel(torch.nn.Module):
    def forward(self, x):
        return x * 3

model = SimpleModel()
example_input = torch.tensor([1.0])
traced_model = torch.jit.trace(model, example_input)
print(traced_model(torch.tensor([2.0])))

medium

A. tensor([2.0])

B. tensor([3.0])

C. tensor([6.0])

D. RuntimeError

Solution

1. Step 1: Understand model behavior

   The model multiplies input by 3, so input 2.0 becomes 6.0.

2. Step 2: Check traced model output

   Tracing records the multiply by 3 operation, so traced_model(2.0) outputs tensor([6.0]).

3. Final Answer:

   tensor([6.0]) -> Option C

4. Quick Check:

   Input 2.0 * 3 = 6.0 [OK]

Hint: Model multiplies input by 3, so output is input*3 [OK]

Common Mistakes:

• Confusing input with output
• Expecting tracing to fail
• Thinking output is unchanged input

4. What is the error in the following code snippet when exporting a model with TorchScript scripting?

import torch
class MyModel(torch.nn.Module):
    def forward(self, x):
        if x.sum() > 0:
            return x * 2
        else:
            return x - 2

model = MyModel()
scripted_model = torch.jit.trace(model, torch.tensor([1.0]))

medium

A. Using torch.jit.trace instead of torch.jit.script for model with conditions

B. Missing example input tensor

C. Model class missing __init__ method

D. Incorrect tensor datatype

Solution

1. Step 1: Identify model features

   The model has a condition (if statement) in forward, which tracing cannot capture correctly.

2. Step 2: Understand TorchScript export methods

   Tracing works only for simple models without control flow; scripting is needed for conditions.

3. Final Answer:

   Using torch.jit.trace instead of torch.jit.script for model with conditions -> Option A

4. Quick Check:

   Model with conditions requires scripting, not tracing [OK]

Hint: Use scripting for models with if/else, not tracing [OK]

Common Mistakes:

• Using trace on models with control flow
• Assuming missing input tensor causes error
• Thinking __init__ is mandatory here

5. You want to export a PyTorch model that uses a loop and conditional statements inside its forward method. Which approach should you use to export it with TorchScript, and why?

hard

A. Use torch.jit.script because it supports control flow like loops and conditions

B. Use torch.jit.trace because it records operations for any model

C. Use torch.save to save the model weights only

D. Use torch.jit.trace with multiple example inputs to cover all paths

Solution

1. Step 1: Analyze model features

   The model has loops and conditions, which require TorchScript to understand control flow.

2. Step 2: Choose correct export method

   torch.jit.script compiles the model including control flow, while tracing cannot handle dynamic paths.

3. Final Answer:

   Use torch.jit.script because it supports control flow like loops and conditions -> Option A

4. Quick Check:

   Loops and conditions need scripting export [OK]

Hint: Loops and conditions require torch.jit.script export [OK]

Common Mistakes:

• Using tracing for models with dynamic control flow
• Saving weights only instead of full model
• Trying to cover all paths with tracing