This repository contains code and documentation for training, testing and evaluating deep learning models [of various architectures] on a vehicle/cars dataset. The dataset contains 196 classes of cars which constitute different make, models, colors and other details of cars.
This repository used keras with tensorflow as backend.
Following are the required libraries for repo. Omit "gpu" for tensorflow, if you want only the cpu version.
opencv-4.1.0
tensorflow-1.13.0-rc0
keras-2.2.4
numpy-1.16
mat4py (to read .mat annotation files)
scikit-learn
scikit-image (for data augmentation)A virtual environment is highly recommended to install these packages. That will make the installation process clean and isolated.
pip install opencv-python==4.1.0
pip install tensorflow-gpu==1.13.0-rc0
pip install keras==2.2.4
pip install numpy
pip install mat4py
pip install sklearn
pip install scikit-image- Please note the this repo is only tested on Windows 10.
- To use first clone the repository.
- Please run all the commands inside the 1_code/ folder of the repository.
- Create a folder named trainedModels in the 2_outputs directory of repo.
- Download the trained model from this link. Place the downloaded file in 2_outputs/trainedModels/ folder in the project.
- In order to test the trained model, this notebook shows simple demo code.
- Please note that all the trained models should reside in 2_outputs/trainedModels/ directory. It is highly recommended to give relative paths to the models.
- In order to evaluate[get metrics], use the evaluate method in evaluate.py. Please do not forget to place the images folder and annotations mat file in the 0_data folder of the project. Remember to match the annotation matfile exactly with the given matfiles from the dataset.
- Create a folder named preTrainedModels in the 2_outputs directory of repo.
- To train models from scratch or fine_tune/transfer_learn, please look at the train.py file.
- For the resnet152 model , I have partially used implementation of this gist. In order to transfer-learn using resnet152, please download pre-trained model from here and place inside the 2_outputs/preTrainedModels directory.
- Please adjust the parameters in config.json file for training.
- Here is a short sample snippet, which can be used for training models.
from train import TrainModel
# path relative to the folder "0_data".
matfile_name = 'devkit/cars_train_annos.mat'
# Percentage of dataset used for validation
validation_perc = 0.2
# Uncomment if you want to fine-tune from already trained model.
# fine_tune_path = '../2_outputs/trainedModels/model_022_0.128.hdf5'
fine_tune_path = None
# Instantiate training class with arguments
trainModel = TrainModel(matfile_name, validation_perc, custom_model=False,
fine_tune_path=fine_tune_path)
# Start training
trainModel.train()If you wish to see the training graphs, [training/val loss, training/val acc], invoke following command to initiate tensorboard.
tensorboard --logdir ../2_ouputs/tensorBoard- All the images are resized to (224,224) in order to gain optimal performance with transfer learning.
- The training has been done by transfer learning/fine tuning from already existing models (e.g. VGG16), which have been pre-trained on millions of imagenet dataset.
- A small pitfall in the dataset is that labels start from 1, so it has been handled by making it start from 0.
- Since the dataset is small, in order to improve generalizability, data augmentation has been performed. Can be seen in the imageGenerator.py.
- Training and experimenting with various architectures with given compute resources, an accuracy of 65% has been obtained on a 20% validation set from training data.
- The most straightforward way of improvement is by increasing the dataset size.
- One can also experiment with more diverse set of architectures.
Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.
Please make sure to update tests as appropriate.