REST API inference in PyTorch - Model Metrics & Evaluation

Metrics & Evaluation - REST API inference

Which metric matters for REST API inference and WHY

When using a REST API to get predictions from a machine learning model, the key metrics to watch are latency and accuracy. Latency tells us how fast the API responds, which is important for user experience. Accuracy tells us if the model's predictions are correct. For classification tasks, precision, recall, and F1-score help us understand prediction quality. For regression, mean squared error or mean absolute error matter. We want a balance: fast responses and good prediction quality.

Confusion matrix example for REST API inference

Imagine a REST API that classifies emails as spam or not spam. Here is a confusion matrix from 100 requests:

      | Predicted Spam | Predicted Not Spam |
      |----------------|--------------------|
      | True Positives (TP) = 40          |
      | False Positives (FP) = 10         |
      | False Negatives (FN) = 5          |
      | True Negatives (TN) = 45          |
    

Total requests = 40 + 10 + 5 + 45 = 100

From this, we calculate:

• Precision = TP / (TP + FP) = 40 / (40 + 10) = 0.8
• Recall = TP / (TP + FN) = 40 / (40 + 5) = 0.8889
• F1-score = 2 * (Precision * Recall) / (Precision + Recall) ≈ 0.842
• Accuracy = (TP + TN) / Total = (40 + 45) / 100 = 0.85

Precision vs Recall tradeoff with REST API examples

In REST API inference, sometimes we want to avoid false alarms (false positives), and sometimes we want to catch all true cases (false negatives).

• High Precision Example: A spam filter API where marking a good email as spam is bad. We want high precision to avoid false positives.
• High Recall Example: A medical diagnosis API where missing a disease is dangerous. We want high recall to catch all true cases.

Choosing the right metric depends on the API's purpose and user impact.

What "good" vs "bad" metric values look like for REST API inference

For a REST API serving predictions:

• Good: Accuracy above 90%, precision and recall balanced above 85%, latency under 200ms.
• Bad: Accuracy below 70%, precision or recall below 50%, latency over 1 second causing slow user experience.

Good metrics mean users get fast and reliable predictions. Bad metrics mean slow or wrong results, hurting trust.

Common pitfalls in REST API inference metrics

• Accuracy paradox: High accuracy can be misleading if classes are imbalanced. For example, 95% accuracy if 95% data is one class but model ignores the other.
• Data leakage: If test data leaks into training, API predictions look too good but fail in real use.
• Overfitting: Model performs well on training but poorly on new API requests.
• Ignoring latency: A very accurate model that is too slow makes the API unusable.

Self-check question

Your REST API model has 98% accuracy but only 12% recall on fraud detection. Is it good for production? Why or why not?

Answer: No, it is not good. The low recall means the model misses most fraud cases, which is dangerous. Even with high accuracy, the model fails to catch fraud, so it should be improved before production.

Key Result

For REST API inference, balance prediction quality (precision, recall) with fast response time (latency) to ensure useful and reliable predictions.

Practice

1. What is the main purpose of using a REST API for PyTorch model inference?

easy

A. To train the model faster using multiple GPUs

B. To send data to the model and receive predictions over the web

C. To visualize the model architecture in a browser

D. To save the model weights to disk

Solution

1. Step 1: Understand REST API role

   A REST API allows communication over the web by sending requests and receiving responses.

2. Step 2: Connect REST API with model inference

   In PyTorch, a REST API is used to send input data and get predictions from the model remotely.

3. Final Answer:

   To send data to the model and receive predictions over the web -> Option B

4. Quick Check:

   REST API = Web data exchange for predictions [OK]

Hint: REST API = send input, get prediction over web [OK]

Common Mistakes:

• Confusing inference with training
• Thinking REST API is for saving models
• Assuming REST API visualizes models

2. Which of the following is the correct way to disable gradient calculation during inference in a Flask REST API using PyTorch?

easy

A. with torch.no_grad():

B. model.train()

C. torch.enable_grad()

D. model.backward()

Solution

1. Step 1: Identify how to disable gradients

   PyTorch uses torch.no_grad() context to turn off gradient tracking during inference.

2. Step 2: Check other options

   model.train() enables training mode, torch.enable_grad() enables gradients, and model.backward() is invalid.

3. Final Answer:

   with torch.no_grad(): -> Option A

4. Quick Check:

   Disable gradients = torch.no_grad() [OK]

Hint: Use torch.no_grad() to save memory during inference [OK]

Common Mistakes:

• Using model.train() instead of model.eval()
• Not disabling gradients causing slow inference
• Calling backward() during inference

3. Given this Flask route for inference, what will be the output if the input JSON is {'input': [1.0, 2.0]} and the model returns tensor([3.0])?

from flask import Flask, request, jsonify
import torch
app = Flask(__name__)
model = torch.nn.Linear(2, 1)
model.eval()

@app.route('/predict', methods=['POST'])
def predict():
    data = request.get_json()
    x = torch.tensor(data['input']).unsqueeze(0).float()
    with torch.no_grad():
        output = model(x)
    return jsonify({'prediction': output.item()})

medium

A. {"prediction": "tensor([3.0])"}

B. {"prediction": [3.0]}

C. Error: model not loaded

D. {"prediction": 3.0}

Solution

1. Step 1: Understand input and output processing

   The input JSON is converted to a tensor with shape (1, 2). The model outputs a tensor with shape (1, 1).

2. Step 2: Check output conversion

   The output tensor's single value is extracted using output.item(), which returns a float, then wrapped in JSON.

3. Final Answer:

   {"prediction": 3.0} -> Option D

4. Quick Check:

   output.item() returns float for JSON [OK]

Hint: Use output.item() to get scalar from tensor for JSON [OK]

Common Mistakes:

• Returning tensor object directly in JSON
• Forgetting to convert input to tensor
• Not using torch.no_grad() causing slow inference

4. You wrote this Flask REST API code for inference but get an error: 'RuntimeError: expected scalar type Float but found Double'. What is the likely fix?

@app.route('/predict', methods=['POST'])
def predict():
    data = request.get_json()
    x = torch.tensor(data['input'])
    with torch.no_grad():
        output = model(x)
    return jsonify({'prediction': output.item()})

medium

A. Convert input tensor to float using .float() before passing to model

B. Call model.train() before inference

C. Remove torch.no_grad() context

D. Change model output to int

Solution

1. Step 1: Understand the error cause

   The model expects float tensors but input tensor is double (default from Python floats).

2. Step 2: Fix input tensor type

   Convert input tensor to float type using .float() before passing to the model.

3. Final Answer:

   Convert input tensor to float using .float() before passing to model -> Option A

4. Quick Check:

   Tensor dtype mismatch fixed by .float() [OK]

Hint: Always convert input to float tensor before model call [OK]

Common Mistakes:

• Not converting input data type
• Calling model.train() during inference
• Removing torch.no_grad() causing slow inference

5. You want to deploy a PyTorch model as a REST API that accepts images, preprocesses them, and returns class predictions. Which sequence of steps is best practice inside your Flask route?

hard

A. Load model globally, call model.eval(), preprocess input image to numpy array, call model backward(), return JSON

B. Load model inside route on every request, call model.train(), preprocess input, get prediction, return JSON

C. Load model once globally, set model.eval(), preprocess input image to tensor, use torch.no_grad(), get prediction, return JSON

D. Preprocess input image after prediction, load model globally, call torch.enable_grad(), return raw tensor

Solution

1. Step 1: Model loading and mode

   Load the model once globally to avoid overhead and set it to evaluation mode with model.eval().

2. Step 2: Input preprocessing and inference

   Preprocess the input image to a tensor, use torch.no_grad() to disable gradients, get prediction, and return JSON response.

3. Final Answer:

   Load model once globally, set model.eval(), preprocess input image to tensor, use torch.no_grad(), get prediction, return JSON -> Option C

4. Quick Check:

   Efficient inference = global load + eval + no_grad + preprocess [OK]

Hint: Load model once, eval mode, no_grad, preprocess input tensor [OK]

Common Mistakes:

• Loading model on every request causing slow response
• Using model.train() during inference
• Not preprocessing input to tensor
• Calling backward() during inference