Car-Brand-Classification

A PyTorch model for a car brand classification based on ResNet152. The cars dataset contains 16,185 images of 196 classes of cars. (11,185 for training; 5,000 for testing)

Hardware

• Intel(R) Core(TM) i5-9600K CPU @ 3.70GHz
• NVIDIA GeForce RTX 2080 Ti

Environment

• Microsoft win10
• Python 3.7.3
• Pytorch 1.7.0
• CUDA 10.2

Install Packages

• pandas, matplotlib

pip install pandas
pip install matplotlib

• pytorch: see https://pytorch.org/get-started/locally/

Data Preparation

Download the given dataset from kaggle.

dataset
  +- testing_data / testing_data
  +- training_data / training_data
  +- training_labels.csv

And run command python data_prepare.py to reorganize the train and valid data structure as below:

train/
├── class1
│   ├── aaa.jpg
│   ├── bbb.jpg
│   └── ccc.jpg	
├── class2
│   ├── ddd.jpg
│   ├── eee.jpg
│   └── fff.jpg	
│        .
│        .
│        .
└── classN
    ├── xxx.jpg
    ├── yyy.jpg
    └── zzz.jpg	

Training

• split the training data into 9:1 for train and valid
• set the parameters for ResNet152
• change the desired number of epochs
• start training the model

Run the program

1. create your working directory and run command git clone https://github.com/chia56028/Car-Brand-Classification.git
2. put the organized training dataset into the cloned folder and run command python data_prepare.py to do the data preparation
3. run command python hw1.py to train

※ get more info by python hw1.py --help

usage: hw1.py [-h] [-r WORKING_DIR] [-tr TRAINING_DIR] [-te TESTING_DIR]
              [-l LABEL_PATH] [-n MODEL_NAME] [-t IS_TRAIN] [-p IS_PREDICT]
              [-e EPOCH] [-lr LEARNING_RATE] [-d DEVICE]

optional arguments:
  -h, --help            show this help message and exit
  -r WORKING_DIR, --root WORKING_DIR
                        path to dataset
  -tr TRAINING_DIR, --train_dir TRAINING_DIR
                        path to training set
  -te TESTING_DIR, --test_dir TESTING_DIR
                        path to testing set
  -l LABEL_PATH, --label LABEL_PATH
                        path to label file
  -n MODEL_NAME, --model_name MODEL_NAME
                        name the model
  -t IS_TRAIN, --train IS_TRAIN
                        train
  -p IS_PREDICT, --predict IS_PREDICT
                        predict
  -e EPOCH, --epochs EPOCH
                        num of epoch
  -lr LEARNING_RATE, --learning_rate LEARNING_RATE
                        set learning rate
  -d DEVICE, --device DEVICE

References

• Transfer Learning for Computer Vision Tutorial
• Deep CARs— Transfer Learning With Pytorch