ONNX export in PyTorch - Model Metrics & Evaluation

Metrics & Evaluation - ONNX export

Which metric matters for ONNX export and WHY

When exporting a PyTorch model to ONNX format, the key metric is model output consistency. This means the ONNX model should produce the same predictions as the original PyTorch model for the same inputs. This ensures the exported model works correctly in other environments.

Metrics like accuracy, precision, or recall are not directly relevant to the export process itself but are important to verify before export. The main focus during export is that the model's outputs match closely between PyTorch and ONNX.

Confusion matrix or equivalent visualization

For ONNX export, a confusion matrix is not applicable. Instead, we compare outputs from PyTorch and ONNX models using numerical checks.

PyTorch output: [0.8, 0.1, 0.1]
ONNX output:    [0.799, 0.101, 0.100]
Difference:    [0.001, 0.001, 0.000]
Max difference < tolerance (e.g., 1e-3) means export is good.
    

Tradeoff: Export fidelity vs model complexity

Exporting complex models to ONNX can sometimes cause small differences in outputs due to unsupported operations or precision changes.

There is a tradeoff between:

• High fidelity: Outputs match very closely, but export may require simplifying or changing some model parts.
• Full complexity: Export the exact model, but outputs may differ slightly or export may fail.

Choosing the right balance depends on your use case. For critical applications, prioritize fidelity and test outputs carefully.

What "good" vs "bad" export looks like

Good export:

• ONNX model runs without errors.
• Outputs match PyTorch outputs within a small tolerance (e.g., max difference < 1e-3).
• Model size and structure are reasonable.

Bad export:

• Export fails or crashes.
• ONNX outputs differ greatly from PyTorch outputs (large numerical differences).
• Unsupported operations cause missing or incorrect behavior.

Common pitfalls when exporting to ONNX

• Dynamic axes not set: Causes fixed input sizes, limiting model flexibility.
• Unsupported ops: Some PyTorch operations may not be supported in ONNX, causing export errors or wrong outputs.
• Model in training mode: Exporting while model is in training mode can cause different outputs (e.g., dropout active).
• Not testing outputs: Failing to compare PyTorch and ONNX outputs can hide export issues.
• Ignoring input preprocessing: Differences in input data format or normalization can cause output mismatches.

Self-check question

Your PyTorch model and ONNX model outputs differ by a max of 0.05 on some inputs. Is this export good?

Answer: No, a max difference of 0.05 is quite large and may cause wrong predictions or behavior. You should investigate the cause, check for unsupported ops, and try to reduce the difference below a small tolerance like 1e-3.

Key Result

ONNX export quality is measured by how closely the ONNX model outputs match the original PyTorch model outputs.

Practice

1. What is the main purpose of exporting a PyTorch model to ONNX format?

easy

A. To save the model in a universal format for sharing and deployment

B. To train the model faster on GPUs

C. To convert the model into a TensorFlow model automatically

D. To visualize the model architecture in PyTorch

Solution

1. Step 1: Understand ONNX export purpose

   ONNX is designed to save models in a format that can be used across different frameworks and platforms.

2. Step 2: Compare options

   Only To save the model in a universal format for sharing and deployment correctly describes the universal sharing and deployment purpose of ONNX export.

3. Final Answer:

   To save the model in a universal format for sharing and deployment -> Option A

4. Quick Check:

   ONNX export = universal format [OK]

Hint: ONNX = share and deploy models universally [OK]

Common Mistakes:

• Confusing ONNX export with training speedup
• Thinking ONNX converts models to TensorFlow automatically
• Assuming ONNX export is for visualization only

2. Which of the following is the correct way to prepare a PyTorch model for ONNX export?

easy

A. Call model.train() before export

B. Export without setting any input

C. Use a dummy input tensor matching the model input shape

D. Use model.eval() after export

Solution

1. Step 1: Identify preparation steps for ONNX export

   Model should be in evaluation mode and a dummy input tensor matching input shape is needed for tracing.

2. Step 2: Evaluate options

   Use a dummy input tensor matching the model input shape correctly states the use of a dummy input tensor. Call model.train() before export is wrong because model.train() is for training mode. Export without setting any input misses input, and Use model.eval() after export is incorrect order.

3. Final Answer:

   Use a dummy input tensor matching the model input shape -> Option C

4. Quick Check:

   Dummy input needed = Use a dummy input tensor matching the model input shape [OK]

Hint: Always use dummy input tensor before export [OK]

Common Mistakes:

• Forgetting to set model.eval() before export
• Not providing dummy input tensor
• Calling model.train() instead of eval()

3. Given the code below, what will be the output of print(onnx_model.graph.input[0].name) after export?

import torch
import torch.nn as nn

class SimpleModel(nn.Module):
    def forward(self, x):
        return x * 2

model = SimpleModel()
model.eval()
dummy_input = torch.randn(1, 3)
torch.onnx.export(model, dummy_input, "model.onnx", input_names=["input_tensor"])
import onnx
onnx_model = onnx.load("model.onnx")
print(onnx_model.graph.input[0].name)

medium

A. x

B. input_tensor

C. input0

D. data

Solution

1. Step 1: Check input_names parameter in export

   The export call sets input_names=["input_tensor"], so the input name in ONNX graph should be "input_tensor".

2. Step 2: Confirm printed input name

   Loading the ONNX model and printing the first input name will output "input_tensor" as set.

3. Final Answer:

   input_tensor -> Option B

4. Quick Check:

   input_names param = input_tensor [OK]

Hint: input_names param sets ONNX input name [OK]

Common Mistakes:

• Assuming default input name like 'input0'
• Confusing PyTorch variable name with ONNX input name
• Not setting input_names and expecting custom name

4. You try to export a PyTorch model to ONNX but get an error: RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.FloatTensor) should be the same. What is the best fix?

medium

A. Export without specifying input_names

B. Use model.train() instead of model.eval()

C. Remove dummy input tensor

D. Move both the model and the dummy input to CPU before export

Solution

1. Step 1: Understand the error cause

   The error means model weights are on CPU but input tensor is on GPU (cuda), causing type mismatch.

2. Step 2: Fix by aligning device

   Moving both the model and the dummy input to CPU ensures both are on the same device (CPU), fixing the mismatch.

3. Final Answer:

   Move both the model and the dummy input to CPU before export -> Option D

4. Quick Check:

   Device mismatch fix = move model and input to CPU [OK]

Hint: Ensure model and input are on same device before export [OK]

Common Mistakes:

• Ignoring device mismatch and exporting anyway
• Switching to train mode instead of fixing device
• Removing dummy input tensor causing other errors

5. You want to export a PyTorch model to ONNX with dynamic batch size support. Which argument should you add to torch.onnx.export to enable this?

hard

A. dynamic_axes={'input': {0: 'batch_size'}, 'output': {0: 'batch_size'}}

B. enable_dynamic_batch=True

C. set_dynamic=True

D. dynamic_batch=True

Solution

1. Step 1: Identify how to specify dynamic axes in ONNX export

   PyTorch ONNX export uses the dynamic_axes argument to mark dimensions as dynamic, e.g., batch size dimension 0.

2. Step 2: Check options for dynamic batch size

   Only dynamic_axes={'input': {0: 'batch_size'}, 'output': {0: 'batch_size'}} correctly uses dynamic_axes with dictionary specifying batch dimension 0 for input and output.

3. Final Answer:

   dynamic_axes={'input': {0: 'batch_size'}, 'output': {0: 'batch_size'}} -> Option A

4. Quick Check:

   Dynamic batch size = dynamic_axes param [OK]

Hint: Use dynamic_axes dict to set dynamic batch size [OK]

Common Mistakes:

• Using nonexistent parameters like enable_dynamic_batch
• Forgetting to specify dynamic axes for output
• Assuming batch size is dynamic by default