Car Brand Classification

Tasks-

• ML Training pipeline to train the pre-trained CNN to classify 3 different car brands.
• Inference Script to run the prediction on a random picture from a given URLs and provide their matching probabilities on each of the brands
• Classification model, returns the most similar image to sample S from the training data set for a given a random sample S

Dependency installation

conda env update --file environment.yaml 
source activate car_brand

or

pip install -r requirements.txt 

Data download

Automated web scraping script to downloads the data from the google

cd utils/
python data_scraper.py ---car-brand audi

Data preparation

Once the images are downloaded, split the train, val. test data set as mentioned below.

Project directory tree should be look like this:

$ROOT/
├── /data
│   ├── test
│   │   ├── audi
│   │   ├── benz
│   │   └── bmw
│   ├── train
│   │   ├── audi
│   │   ├── benz
│   │   └── bmw
│   └── val
│       ├── audi
│       ├── benz
│       └── bmw
|
├── /inference_data
├── /output
├── /utils
│   |── driver
│   |    └── geckodriver
│   └── data_scraper.py
├── /config.json 
├── /model_interface.ipynb
└── /model_training.ipynb

Model Training

model_training.ipynb

Open this in jupyter notebook to train the model.

Script Functionalities-

1. Loads the dataset according the data split
2. Build model
3. Train model
4. Save model
5. Test prediction
6. convert the H5 to pb file

Model Inference

model_inference .ipynb

Script Functionalities-

1. Downloads the random picture from the given URls
2. Loads the data and saved H5 model
3. Inference on random downloaded test samples
4. Returns the most similar image to sample S from the training data set

Results

Sr. No	Model	Model Size	epochs	optimizer	train Acc	Val ACC
1	MobileNet2	14 MB	30	Adam	70.05	84.38
2	ResNet50	98 MB	30	Adam	34.47	39.58

Contributors

• Sharat Gujamagadi

Contact

• Sharat Gujamagadi (sharatcg7@gmail.com)