ResNet50-PyTorch

The ResNet50 model training and inference in PyTorch

Project Overview

This project implements a foundational ResNet50 model using PyTorch. The model is trained on a custom dataset of healthy and diseased(rust, scab) apple leaves. This README provides instructions to set up, run, and evaluate the model.

Prerequisites

The below dependencies are to be installed before running the training model. For the versions used, please refer the requirements.txt file.

• Python 3.7 or higher
• PyTorch
• torchvision
• Numpy

Setup

1. Clone the repository using git clone <URL>
2. Check all the dependencies to be installed.
3. Train the model on a dataset : python resnet50_model.py
4. Model inference and evaluation : Once the model is saved, go ahead and experiment with the model by fine-tuning on new data or different datasets. The model in this repo saved_model/resnet50_apple_leaf_disease.pth can also be used. Some examples are given below :

i. Load the saved model :

# Initialize model
num_classes = <number_of_classes>
model = ResNet50(num_classes=num_classes)

# Load the saved state dictionary
model.load_state_dict(torch.load('resnet50_apple_leaf_disease.pth'))
model.eval()
print("Model loaded and set to evaluation mode.")

ii. Print the model architecture :

print(model)
for name, param in model.named_parameters():
    print(name, param.data)

iii. Check accuracy :

accuracy = evaluate_model(model, test_loader)
print(f"Accuracy : {accuracy:.2f}%")

Further Exploration

The current model has a total accuracy of 93.97%. This can be further optimized to meet particular project requirements using data augmentation, fine-tuning the model and other advanced optimizers.