This project is a sandbox project! This projects will be used for experiments!

Image Classifier using TensorFlow

How to Use

Installation

1. This project can be run only linux environment.
2. This project is recommended to run on virtual environment include docker.
3. Command pip install -r requirements.txt to install dependency.

Commands

• Train: python3 train.py option=<configuration file (optional)>

  • Used to obtain weights.
  • The option means argument for indivisual training options.

• Resume: python3 train.py resume=<checkpoint path>

  • The <checkpoint path> means a path of checkpoint_path and checkpoint_name in option.yaml

• Predict: python3 classify.py weights=path/to/checkpoint_path/checkpoint_name image=path/to/image(s/dir) epoch=[best|last|<an epoch number>](optional) output_format=[csv|stdout](optional)

• Export: python3 export.py path=path/to/checkpoint_path/checkpoint_name epoch=[best|last|<an epoch number>] image_size=<image_size>

  • Export weights to saved_model and TFLite
  • The weights files will be saved onto path/to/checkpoint_path/checkpoint_name/export.

• Select specific GPU (example): CUDA_VISIBLE_DEVICES=[cpu|<GPU numbers (example: 0,1,2,3)>] python3 train.py option=example.yaml

  • The default uses all GPUs.
  • If you have multiple GPUs and want to select specific GPU, you have to use this command form or set the CUDA_VISIBLE_DEVICES variable.

Dataset File Structure

dataset/
 ├ train/
 │   ├ class1/
 |   |   └ image1.xxx
 │   └ class2/
 |       └ image2.xxx
 └ val/
     ├ class1/
     |   └ image3.xxx
     └ class2/
         └ image4.xxx

Options

Option application order

1. cfg/settings.yaml (default settings)
2. option.yaml (user global settings)
3. configuration file (copied from option.yaml)

option.yaml

• Dataloader options
  • Image crop options
  • Image translate options
  • Color adjustment options
  • Quality adjustment options
  • Other dataloader options
• Train options
  • Checkpoint path
  • Training options (optimizer, loss, etc.)
  • Learning rate scheduler options
  • Model option

Model Configurations

• ResNet Models

  • resnet/18.yaml
  • resnet/34.yaml
  • resnet/50.yaml
  • resnet/110.yaml
  • resnet/152.yaml

• CSPResNetS models

  • cspresnets/18.yaml
  • cspresnets/34.yaml
  • cspresnets/50.yaml
  • cspresnets/101.yaml
  • cspresnets/152.yaml

• CSPResNetP models

  • cspresnetp/18.yaml
  • cspresnetp/34.yaml
  • cspresnetp/50.yaml
  • cspresnetp/101.yaml
  • cspresnetp/152.yaml

• S/P suffix of CSPResNet is meaning of implementation method. S is split by channels, P is projection by channales

• Other Models

  • vgg16.yaml
  • googlenet.yaml

Experiment

Experimental Environment

• CPU: AMD Ryzen 7900
• Memory: DDR5 128GB
• GPU: RTX 5060 Ti 16GB @ 2.0GHz x2
• OS: Ubuntu 24.04 (docker image: nvidia/cuda:12.8.1-cudnn8-devel-ubuntu24.04)
• Python: CPython 3.12
• TensorFlow version: 2.20.0

Experimetnal Results

Model	Accuracy	Params	FLOPs
ResNet18	67.15%	11.70M	3.63G
ResNet34	69.34%	21.81M	7.34G
ResNet50	72.36%	26.73M	8.64G
ResNet101	72.76%	44.65M	15.64G
ResNet152	73.17%	60.34M	23.08G
CSPResNetS18	66.88%	7.28M	1.88G
CSPResNetS34	69.19%	9.82M	2.92G
CSPResNetS50	71.81%	19.77M	5.52G
CSPResNetS101	73.13%	24.27M	7.27G
CSPResNetS152	73.94%	28.22M	9.14G
CSPResNetP18	68.16%	7.64M	2.10G
CSPResNetP34	69.85%	10.18M	3.03G
CSPResNetP50	71.03%	25.33M	7.00G
CSPResNetP101	71.59%	29.83M	8.76G
CSPResNetP152	71.91%	33.78M	10.62G

Trained and evaluated in ImageNet2012 dataset.