Description

Clothing Recognition is a PyTorch-based image classification project built on the FashionMNIST dataset. The repository contains a custom convolutional neural network implementation, training and evaluation loops, utilities for prediction on sample images, and visualization scripts for performance analysis (precision/recall curves, confusion matrix, F1 vs confidence).

Key Features:

1. Model Architecture:

• Custom CNN implemented as a torch.nn.Module.
• Two convolutional stacks with ReLU activations and average pooling.
• Final fully connected classifier producing logits for 10 classes.
• Custom MinPool2d implementation included for experimentation.

2. Dataset & DataLoader:

• Uses torchvision.datasets.FashionMNIST (28×28 grayscale images).
• Automatic download to data_2 folder if needed.
• Training and test sets wrapped in torch.utils.data.DataLoader with configurable batch size.

3. Training & Evaluation:

• Training loop with SGD optimizer and CrossEntropy loss.
• Accuracy computed per batch and printed per epoch.
• Evaluation utilities returning loss and accuracy summary.

4. Prediction & Visualization:

• Function to run model predictions on arbitrary samples and return class probabilities.
• Plots sample predictions with true vs predicted labels (green for correct, red for wrong).
• Precision–Recall curves per class and macro-averaged curve.
• F1-score vs confidence threshold visualizations and precision/recall vs threshold charts.
• Confusion matrix plotting using sklearn.metrics.

5. Metrics:

• Precision, Recall, F1-score (weighted and macro as required).
• Detailed classification report (per-class precision/recall/f1).

6. Requirements & Setup:

• Python 3.x
• PyTorch and torchvision
• NumPy, Matplotlib, scikit-learn
• Install dependencies: pip install torch torchvision numpy matplotlib scikit-learn

7. Usage:

• Train the model and evaluate: run the main training script (e.g., python train.py or your notebook).
• Make single-sample predictions using the provided prediction helper function.
• View plots generated by evaluation utilities to analyze model performance.

8. Troubleshooting:

• GPU not used: Ensure CUDA is available and PyTorch is installed with CUDA support.
• Slow training: Reduce dataset size, increase batch size, or use GPU.
• Poor accuracy: Try increasing model capacity, use data augmentation, or tune learning rate.

9. Future Improvements:

• Add data augmentation (random crop, flips, brightness/contrast) to improve generalization.
• Use more advanced architectures or pretrained backbones (transfer learning).
• Implement model checkpoints and training resume functionality.
• Provide a simple web demo for live webcam classification and visualization.

The code is available here: GitHub Repository.