You can deploy this model and start earning money today!

sandipautane/
ResNet50

Views

No views yet

ResNet50 for CIFAR-100

A PyTorch implementation of ResNet50 optimized for CIFAR-100 classification with a complete training pipeline.

Features

ResNet50 Architecture: Bottleneck-based ResNet with [3, 4, 6, 3] layer configuration
CIFAR-100 Optimized: 3x3 stem convolution (no maxpool) for 32x32 input images
Complete Training Pipeline: Command-line training script with data loading, augmentation, and checkpointing
Data Augmentation: Random cropping and horizontal flipping for improved generalization
Learning Rate Scheduling: Multi-step learning rate decay at epochs 60 and 80
Checkpoint Management: Automatic saving of best models and regular checkpoints

Installation

pip install -r requirements.txt

Quick Start

Training ResNet50 on CIFAR-100

bash

# Basic training (100 epochs, default parameters)
python train.py

# Custom training parameters
python train.py --epochs 200 --batch-size 64 --lr 0.05 --plot

# Resume from checkpoint
python train.py --resume ./checkpoints/best_model.pth

Using the Model in Code

python

from models.resnet import ResNet50

# Create model for CIFAR-100 (100 classes)
model = ResNet50(num_classes=100)

# Forward pass
import torch
x = torch.randn(1, 3, 32, 32)  # Batch of 1, 3 channels, 32x32 images
output = model(x)  # Shape: [1, 100]

Training Options

Argument	Default	Description
`--epochs`	100	Number of training epochs
`--batch-size`	128	Batch size for training
`--lr`	0.1	Learning rate
`--momentum`	0.9	SGD momentum
`--weight-decay`	5e-4	Weight decay for regularization
`--resume`	""	Path to checkpoint to resume from
`--save-dir`	"./checkpoints"	Directory to save model checkpoints
`--plot`	False	Generate training plots

Expected Performance

With default training settings, you can expect:

Training Accuracy: ~95-98%
Test Accuracy: ~70-75%
Training Time: ~2-3 hours on a modern GPU

File Structure

├── models/
│   ├── __init__.py
│   └── resnet.py          # ResNet implementation
├── train.py               # Training script
├── requirements.txt       # Dependencies
└── README.md             # This file

Dependencies

PyTorch >= 2.1
Torchvision >= 0.16
Matplotlib >= 3.5

Checkpoints

The training script automatically saves:

Best Model: checkpoints/best_model.pth (highest test accuracy)
Regular Checkpoints: checkpoints/checkpoint_epoch_X.pth (every 10 epochs)
Training Plots: checkpoints/training_plots.png (if --plot flag used)

Model Architecture

The ResNet50 implementation uses:

Stem: 3x3 convolution (stride=1, padding=1) for CIFAR-100 compatibility
Bottleneck Blocks: [3, 4, 6, 3] configuration across 4 stages
Global Average Pooling: 4x4 adaptive pooling before classification
Classification Head: Linear layer with 100 outputs for CIFAR-100

sandipautane/ResNet50