Model packaging (.mar files) in PyTorch - Model Metrics & Evaluation

Metrics & Evaluation - Model packaging (.mar files)

Which metric matters for Model Packaging (.mar files) and WHY

When packaging a model into a .mar file, the main goal is to keep the model's performance intact after deployment. Therefore, the key metrics to check are the model's prediction accuracy, loss, and inference speed before and after packaging. This ensures the model behaves the same and runs efficiently in production.

Confusion Matrix or Equivalent Visualization

For classification models packaged in .mar files, the confusion matrix before and after packaging should be identical. For example:

      | Predicted Positive | Predicted Negative |
      |--------------------|--------------------|
      | True Positive (TP)  | False Negative (FN) |
      | False Positive (FP) | True Negative (TN)  |

      TP + FP + FN + TN = total samples
    

If the confusion matrix changes after packaging, it means the model predictions changed, indicating a packaging issue.

Precision vs Recall Tradeoff with Concrete Examples

Packaging should not affect the tradeoff between precision and recall. For example, if a spam filter model packaged as a .mar file had 90% precision and 85% recall before packaging, it should keep similar values after packaging.

If precision drops, the model may mark good emails as spam (false positives). If recall drops, it may miss spam emails (false negatives). Packaging must preserve this balance.

What "Good" vs "Bad" Metric Values Look Like for Model Packaging

Good: Metrics before and after packaging are nearly identical (e.g., accuracy difference < 1%). Inference speed is stable or improved. No errors during loading or prediction.

Bad: Large drops in accuracy, precision, recall, or F1 score after packaging. Increased latency or failures when loading the .mar file. This means the packaging corrupted the model or environment.

Common Metrics Pitfalls in Model Packaging

• Accuracy Paradox: High accuracy but poor recall or precision after packaging can hide problems.
• Data Leakage: Testing metrics on training data can falsely show no change after packaging.
• Overfitting Indicators: If metrics improve unrealistically after packaging, it may be a sign of testing on the wrong data.
• Environment Differences: Differences in hardware or software versions can cause metric changes unrelated to packaging.

Self Check

Your model packaged as a .mar file shows 98% accuracy but only 12% recall on fraud detection. Is it good for production? Why or why not?

Answer: No, it is not good. Although accuracy is high, the recall is very low, meaning the model misses most fraud cases. In fraud detection, recall is critical to catch as many frauds as possible. Packaging should preserve or improve recall, so this indicates a problem.

Key Result

Model packaging must preserve key metrics like accuracy, precision, recall, and inference speed to ensure reliable deployment.

Practice

1. What is the main purpose of packaging a PyTorch model into a .mar file?

easy

A. To visualize the model architecture

B. To convert the model into a different training format

C. To compress the model for storage only

D. To bundle the model and code for easy deployment with TorchServe

Solution

1. Step 1: Understand what a .mar file contains

   A .mar file bundles the trained model and the handler code needed to serve it.

2. Step 2: Identify the use case of .mar files

   These files are used by TorchServe to deploy models as APIs easily.

3. Final Answer:

   To bundle the model and code for easy deployment with TorchServe -> Option D

4. Quick Check:

   Model packaging = deployment bundle [OK]

Hint: Think deployment: .mar bundles model + code [OK]

Common Mistakes:

• Confusing .mar with just compression
• Thinking .mar changes training format
• Assuming .mar is for visualization

2. Which command correctly creates a .mar file using torch-model-archiver?

easy

A. torch-model-archiver --model-name mymodel --model-file model.pth --handler handler.py --export-path model_store

B. torch-model-archiver --model-name mymodel --model-file model.pth --handler handler.txt --export-path model_store

C. torch-model-archiver --model-name mymodel --model-file model.txt --handler handler.py --export-path model_store

D. torch-model-archiver --model-name mymodel --model-file model.py --handler handler.py --export-path model_store

Solution

1. Step 1: Identify correct model file type

   The model file should be a PyTorch model file like .pth, not a Python script or text file.

2. Step 2: Confirm handler file type

   The handler must be a Python script .py, not a text file.

3. Final Answer:

   torch-model-archiver --model-name mymodel --model-file model.pth --handler handler.py --export-path model_store -> Option A

4. Quick Check:

   Model file = .pth, handler = .py [OK]

Hint: Model file is .pth, handler is .py script [OK]

Common Mistakes:

• Using .py file as model file
• Using .txt as handler file
• Mixing file extensions incorrectly

3. Given this command:

torch-model-archiver --model-name resnet --model-file resnet.pth --handler image_classifier.py --export-path model_store

What will be the output?

medium

A. A Python script named resnet.py created

B. An error because resnet.pth is not a valid model file

C. A file named resnet.mar created in model_store directory

D. A folder named resnet.mar created

Solution

1. Step 1: Understand torch-model-archiver output

   The command packages the model and handler into a single .mar file.

2. Step 2: Check output location and file name

   The .mar file is named after the model name and saved in the export path.

3. Final Answer:

   A file named resnet.mar created in model_store directory -> Option C

4. Quick Check:

   .mar file = model + handler bundle [OK]

Hint: Output is a .mar file named after model [OK]

Common Mistakes:

• Expecting a Python script output
• Thinking .mar is a folder
• Assuming error due to .pth file

4. You ran this command:

torch-model-archiver --model-name mymodel --model-file model.pth --handler handler.py --export-path model_store

But got an error: Handler file not found. What is the likely cause?

medium

A. The handler.py file is missing or path is incorrect

B. The model.pth file is corrupted

C. The export path model_store does not exist

D. The model name contains invalid characters

Solution

1. Step 1: Analyze the error message

   The error says handler file not found, so the problem is with the handler.py file location or name.

2. Step 2: Check other options

   Model file corruption or export path issues would cause different errors, not handler missing.

3. Final Answer:

   The handler.py file is missing or path is incorrect -> Option A

4. Quick Check:

   Error message = handler file missing [OK]

Hint: Check handler file path if 'Handler file not found' error occurs [OK]

Common Mistakes:

• Blaming model file for handler error
• Ignoring file path correctness
• Assuming export path causes handler error

5. You want to package two different PyTorch models with custom handlers into .mar files for TorchServe. Which approach correctly packages both models for deployment?

hard

A. Package one model as .mar and copy the second model file into the model_store folder without packaging

B. Run torch-model-archiver twice, each time with the respective model file, handler, and unique model name, exporting to the same model_store folder

C. Combine both models into one .pth file and use a single handler to package once

D. Use torch-model-archiver once with both model files and handlers listed separated by commas

Solution

1. Step 1: Understand packaging multiple models

   Each model requires its own .mar file with its specific handler and unique model name.

2. Step 2: Confirm correct packaging method

   Running torch-model-archiver twice with different parameters and exporting to the same folder is the correct way.

3. Final Answer:

   Run torch-model-archiver twice, each time with the respective model file, handler, and unique model name, exporting to the same model_store folder -> Option B

4. Quick Check:

   One .mar per model + handler [OK]

Hint: Package each model separately with unique names [OK]

Common Mistakes:

• Trying to combine models in one .mar file
• Copying model files without packaging
• Passing multiple files in one command