brach-assignment-bp-2026 0.1.0

A simple PyTorch wrapper for image classification using a pretrained ResNet18 model

Vision Model Package: ResNet18 Wrapper

This package provides a simple, production-ready interface for image classification using a pretrained ResNet18 model. It is designed to be easily installable and used for quick inference tasks.

---

Installation

The package can be installed via pip:

pip install Dmytro-Shapovalov-brach-assignment-2026

---

Quick Start

from my_package import ModelWrapper

# Initialize the model (weights are loaded automatically)
inferer = ModelWrapper()

# Run prediction
img_path = "test_image.jpg"
predict = inferer.predict(img_path)

print(f"Predicted Class Index: {predict}")

---

API reference

• ModelWrapper(model_path=None)

  The main class for managing the model and inference.

  model_path: Optional path to a custom .pth file. If None, it automatically loads the weights.pth bundled with the package.

• ModelWrapper.predict(image_path)

  Performs the full inference pipeline on a single image.

  Input: str path to the image file.

  Returns: int representing the predicted class index.

  Process: Handles image loading (PIL), resizing (224x224), normalization (ImageNet stats), and tensor conversion (Torch)

---

Evaluation Proposal

To evaluate this model on a specific image classification task, I would use a standard validation pipeline to evaluate both accuracy and robustness.

• Dataset: A separate test set relevant to the task.

• Metrics:

  Accuracy: measures the percentage of correct predictions.

  F1-Score: evaluates performance across potentially imbalanced classes.

  Inference Latency: measures the average time taken for a single prediction.

• Pipeline:

  Data Loading: using torchvision.datasets and DataLoader feed images in batches.

  Inference Mode: setting the model to .eval() and using torch.no_grad() to disable gradient calculation and save memory.

  Comparison: comparing predicted indices against ground truth labels to generate a confusion matrix.

  Reporting: aggregating results to identify specific classes where the model might be underperforming.

• Evaluation Pseudocode:

  model.eval()
  results = []
  with torch.no_grad():
      for images, labels in test_loader:
          outputs = model(images)
          preds = torch.argmax(outputs, dim=1)
          results.extend((preds == labels).tolist())
  
  accuracy = sum(results) / len(results)
  print(f"Test Accuracy: {accuracy:.2%}")