CIFAR Image Classifier

Introduction

In this project, multiple CNN architectures are used to train image classification models on the CIFAR dataset. Model compression methods like network pruning are also explored to reduce the size of the neural network by removing redundant parameters which leads to improved computational efficiency during inference.

Performance

Dataset	Model	Pruning Method	Pruning Ratio (%)	Params (M)	Top1 Acc (%)
CIFAR10	VGG16	-	-	33.65	94.71
CIFAR10	VGG16	L1 Norm Uniform	10	28.74	92.19
CIFAR10	MobileNetV1	-	-	3.22	91.53
CIFAR10	MobileNetV2	-	-	2.3	94.44
CIFAR10	ResNet18	-	-	11.17	95.42
CIFAR100	VGG16	-	-	34.02	73.95
CIFAR100	VGG16	L1 Norm Uniform	10	29.07	63.92
CIFAR100	MobileNetV1	-	-	3.31	68.18
CIFAR100	MobileNetV2	-	-	2.41	75.96
CIFAR100	ResNet18	-	-	11.22	77.32

Getting Started

1. Clone the repository to the desired directory:

git clone https://github.com/brondaeh/CIFAR-Image-Classifier.git

2. Install the required libraries to your environment:

pip install -r requirements.txt

3. Modify the configuration parameters in config.yaml.
   • Set enable_gpu to True if you have a cuda enabled GPU for training.
   • By default L1 norm pruning is only applicable for VGG16, keep pruning_flag to False for other models.
4. Start training and testing by running main.py.

File Structure

CIFAR-Image-Classifier
├── Model_Details
│   ├── mobilenetv1_details.md
│   ├── mobilenetv2_details.md
│   ├── resnet18_details.md
│   └── vgg16_details.md
├── Models
│   ├── mobilenetv1.py
│   ├── mobilenetv2.py
│   ├── resnet.py
│   └── vgg.py
├── Pruner
│   ├── Pruning_Criterion
│   │   ├── KMean
│   │   │   └── KMean_base.py
│   │   ├── L1norm
│   │   │   └── L1norm.py
│   │   ├── Taylor
│   │   │   └── Taylor.py
│   ├── pruning_engine_base.py
│   └── pruning_engine.py
├── Pruning_Functions
│   └── prune_vgg.py
├── config.yaml
├── main.py
├── README.md
├── requirements.txt
├── utils.py

References

1. CNN-Pruning-Engine
2. pytorch-cifar
3. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
4. MobileNetV2: Inverted Residuals and Linear Bottlenecks
5. Deep Residual Learning for Image Recognition
6. Very Deep Convolutional Networks for Large-Scale Image Recognition