This project implements a ResNet50 model for image classification on the ImageNet-1K dataset using PyTorch Lightning. The model is trained to classify images into 1000 different categories and includes a user-friendly Gradio interface for making predictions.

• ResNet50 model from torchvision
• Training on ImageNet-1K dataset (1000 classes)
• Optimized data loading using pre-generated annotation files
• Learning rate optimization using LR Finder
• Mixed precision training (FP16)
• Distributed training across multiple GPUs
• Gradio web interface for model inference

• Hardware: AWS g6.12xLarge Instance
  • 4 GPUs
  • 48 vCPUs
  • 24GB memory per GPU
  • Total cost $17

• Data Loading: Custom annotation files for faster dataset creation instead of ImageFolder
• Learning Rate:
  • Used LR Finder for 500 steps to find optimal learning rate
  • Implemented One Cycle Policy for learning rate scheduling
  • Total training steps: 33,399
• Training Speed:
  • FP16 precision for faster training
  • Distributed Data Parallel (DDP) training across 4 GPUs
  • Batch size of 384 with gradient accumulation

• Architecture: ResNet50
• Number of classes: 1000
• Batch size: 384
• Learning rate: 0.478 (found using LR Finder)
• Training epochs: 40
• Validation frequency: Every 5 epochs

• Final Training Accuracy: 72.16%
• Final Validation Accuracy: 73.55%
• Best Validation Accuracy: 73.55%
• Average Training Time per Epoch: 908.20 seconds
• Total Training Time: 10.09 hours

Figure 1: Training and validation accuracy/loss, Learning rate, and training time per epoch

Figure 2: Accuracy per step

Hugging Face APP

The key files in this project are:

• requirements.txt: Contains the dependencies for the project
• ResnetNetwork.py: Implements the ResNet50 model using PyTorch Lightning with custom training logic and LR finder
• dataset.py: Contains the ImageNet dataset class and Lightning DataModule with data transformations
• create_annotations.py: Utility script to create CSV annotation files for efficient data loading
• config.py: Contains all configuration parameters for model training and dataset paths
• main.py: Entry point for training, sets up training configuration and launches the training process

1. Clone the repository

git clone https://github.com/walnashgit/Resnet50Imagenet.git

2. Install the dependencies using

pip install -r requirements.txt

3. Download the ImageNet dataset and place it in the data directory

4. Run the create_annotations script

python create_annotations.py

5. Run the training script

python main.py