kikogazda/Efficient_NetV2_Edition

🚗 EfficientNetV2 Car Classifier: Fine-Grained Vehicle Recognition

EfficientNetV2 Car Classifier delivers robust, fine-grained recognition for 196 car makes and models, powered by EfficientNetV2, state-of-the-art augmentations, rigorous metric tracking, and full visual explainability with Grad-CAM.
Developed by kikogazda, 2025.

---

📁 Project Structure

Efficient_NetV2_Edition/
├── efficientnetv2_best_model.pth      # Best model weights
├── Last_model.ipynb                   # Full training & evaluation pipeline
├── class_mapping.json                 # Class index to name mapping
├── *.csv                              # Logs, splits, labels, and metrics
├── *.png                              # Visualizations and Grad-CAM outputs
├── README.md                          # Model card (this file)
└── ...                                # Additional scripts, reports, and assets

---

🚦 Table of Contents

• Overview
• Dataset & Preprocessing
• Model Architecture
• Training Pipeline
• Explainability (Grad-CAM)
• Visualizations
• Metrics & Results
• Hugging Face & Demo
• Download Resources
• Usage & Inference

---

Overview

EfficientNetV2 Car Classifier tackles the real-world challenge of distinguishing between 196 car makes and models, even when differences are nearly imperceptible.
Highlights:

• Modern EfficientNetV2 backbone with transfer learning
• Aggressive, real-world augmentation pipeline
• Class balancing for rare makes/models
• Extensive, scriptable metric tracking and reporting
• End-to-end explainability with Grad-CAM
• Fully reproducible, robust, and deployment-ready

---

Dataset & Preprocessing

• Dataset: Stanford Cars 196
  • 196 classes, 16,185 images (official train/test split)
  • Detailed make/model/year for each image
• Preprocessing:
  • Annotation CSV export and class mapping JSON
  • Stratified train/val/test split (maintains class distribution)
  • Outlier cleaning and normalization
  • Augmentations: random resized crop, flip, rotate, color jitter, blur
  • ImageNet mean/std normalization

---

Model Architecture

• Backbone: EfficientNetV2 (pretrained)
  • All but the last blocks frozen initially
  • Custom classifier head for 196 classes (Linear → ReLU → Dropout → Linear)
• Optimization:
  • Adam optimizer
  • Cross-Entropy loss (with label smoothing)
  • Learning rate scheduling (ReduceLROnPlateau)
  • Early stopping (macro F1 on validation)
  • WeightedRandomSampler for class balance

Flow:
Input → [Augmentations] → EfficientNetV2 Backbone → Custom Head → Softmax (196 classes)

---

Training Pipeline

• Epochs: Up to 25 (early stopping enabled)
• Batch Size: 32 (weighted sampling)
• Validation: Macro/micro metrics, confusion matrix, Top-3/Top-5 accuracy
• Logging: All metrics and losses to CSV, plus high-res visual plots:
  • Accuracy/F1 per epoch
  • Precision/Recall (macro, weighted)
  • Loss curve
  • Top-3/Top-5 accuracy
• Artifacts: All reports, CSVs, and visuals in repo for transparency

---

Explainability (Grad-CAM)

Grad-CAM overlays highlight image regions most responsible for model predictions—letting you "see" what the network is using for its decisions.

• Why? Trust, transparency, debugging.
• How? For every prediction, a heatmap overlay shows most influential pixels.

Heatmaps visualize key decision regions for each sample.

---

Visualizations

Here are key visualizations from the training and evaluation process, including loss curves, accuracy plots, and Grad-CAM++ overlays that illustrate what the model focuses on.

🎯 Accuracy & F1 Score per Epoch

Visualizing training and validation accuracy alongside macro F1 score.

---

📉 Training vs Validation Loss

Clear comparison of model learning over time.

---

📈 Precision & Recall Trends

Macro and weighted precision/recall for detailed class-wise performance.

---

📊 Top-3 and Top-5 Accuracy Over Epochs

Measuring how often the correct class is within the top predictions.

---

🏆 Top-20 Most Accurate Classes

Sorted bar plot of classes the model predicts with the highest accuracy.

---

🧩 Confusion Matrix

High-resolution heatmap showing misclassifications and accuracy by class.

---

📈 Metrics & Results

Metric	Value
train_loss	0.97
train_acc	0.997
val_loss	1.40
val_acc	0.87
val_precision_macro	0.89
val_precision_weighted	0.89
val_recall_macro	0.87
val_recall_weighted	0.87
val_f1_macro	0.87
val_f1_weighted	0.88
val_top3	0.95
val_top5	0.97

---

Hugging Face Demo

Live Gradio Demo:
Click here to launch the demo

---

Download Resources

• Stanford Cars 196 Dataset:
  Direct download from Stanford
• Trained Model Weights:
  Download from Hugging Face (efficientnetv2_best_model.pth)
• Class mapping/metadata:
  Included as class_mapping.json in this repo

---

Usage & Inference

1. Install dependencies

pip install -r requirements.txt
pip install torch torchvision pytorch-grad-cam gradio

import torch
from torchvision import transforms
from PIL import Image
import json
from efficientnet_pytorch import EfficientNet

# Load model
model = EfficientNet.from_pretrained('efficientnet-b2', num_classes=196)
model.load_state_dict(torch.load("efficientnetv2_best_model.pth", map_location="cpu"))
model.eval()

# Preprocess
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
img = Image.open("your_image.jpg").convert("RGB")
input_tensor = transform(img).unsqueeze(0)

# Predict
with torch.no_grad():
    output = model(input_tensor)
    pred = output.argmax(1).item()

# Class name
with open("class_mapping.json") as f:
    class_map = json.load(f)
print("Predicted class:", class_map[str(pred)])