Bird Species Classification with PyTorch

This project focuses on building and training deep learning models to classify bird species from images using PyTorch. It includes two different training approaches:

1. A custom convolutional neural network (CNN) built from scratch.
2. A pretrained transfer learning model based on ResNet-50.

The goal is to compare a fully custom model with a modern pretrained architecture and understand the tradeoffs between them in terms of accuracy, training time, and generalization.

---

Project Structure

├── model.py # Custom CNN architecture
├── pretrained_model.py # ResNet50-based transfer learning model
├── train.py # Training loop and dataset loading
├── predict.py # Run inference on a single image
├── bird_cnn_[retrained].pth # Saved model checkpoint
└── README.md

---

Models

1. Custom CNN (model.py)

This model is implemented entirely from scratch using standard convolutional building blocks:

• Convolution layers
• Batch normalization
• ReLU activations
• Max pooling
• Fully connected classifier

The architecture progressively increases feature depth (32 → 256 channels) while reducing spatial resolution. Images are expected to be resized to 128×128, which results in a final feature map of 8×8 before classification.

This model is useful for:

• Understanding how CNNs work internally
• Learning how architectural choices affect performance
• Training when pretrained weights are not desired

---

2. Pretrained Model (pretrained_model.py)

This model uses ResNet-50, pretrained on ImageNet, as a feature extractor. The original classification head is replaced with a custom classifier tailored to the bird species dataset.

Key ideas here:

• The backbone has already learned general visual features (edges, textures, shapes).
• Only the final layers need to adapt to bird species classification.
• Training is faster and typically more accurate with smaller datasets.

This approach demonstrates transfer learning, which is widely used in real-world machine learning systems.

---

Machine Learning Concepts Used

Convolutional Neural Networks (CNNs)

CNNs are designed to process images by learning spatial hierarchies of features. Early layers learn simple patterns like edges, while deeper layers capture more complex structures.

Transfer Learning

Instead of training a large network from scratch, we reuse a pretrained model and fine-tune it. This:

• Reduces training time
• Improves performance on limited data
• Leverages knowledge from large-scale datasets

Freezing and Unfreezing Layers

During training with a pretrained model:

• Frozen layers do not update their weights and act as fixed feature extractors.
• Unfrozen layers are trainable and adapt to the new task.

This helps prevent overfitting early in training and allows controlled fine-tuning later.

Overfitting Prevention

Several techniques are used to improve generalization:

• Dropout in the classifier
• Data normalization and resizing
• Validation set monitoring

---

Dependencies

This project requires the following Python packages:

• Python 3.8+
• PyTorch
• torchvision
• Pillow
• NumPy

Install dependencies

Using pip:

pip install torch torchvision pillow numpy

If you have an NVIDIA GPU, make sure your PyTorch installation matches your CUDA version. You can verify GPU support with:

python -c "import torch; print(torch.cuda.is_available())"

---

Training

Train from scratch (custom CNN)

python train.py

Training scripts load image datasets using more_data, apply transformations, and optimize the model using cross-entropy loss.

Models are trained on GPU when available:

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

Checkpoints store:

• Model weights
• Class-to-index mapping
• Class names for inference

---

Inference

The predict.py script loads a saved checkpoint and runs inference on a single image. The output is the predicted bird species based on the trained classifier.

Example:

python predict.py path/to/image.jpg

The output will include a list of probable birds ranked by the softmax evaluator.

---

Summary

This project demonstrates both foundational and modern approaches to image classification:

• Building a CNN from scratch to understand the fundamentals
• Applying transfer learning for better real-world performance
• Using PyTorch best practices for training and inference
• Together, these models provide a solid introduction to practical computer vision workflows.

---

Dataset Acknowledgement

This project uses bird species image data for training and evaluation.
Huge thanks to the authors and maintainers of the following public dataset repository: UmairPirzada/BIRDS_SPECIES--IMAGE-CLASSIFICATION