R-CNN designed to handle the Standford Cars dataset.
Largely copied from: https://github.com/foamliu/Car-Recognition
However adjusted to require slightly less pre-processing, to be slightly less memory heavy, to be slightly more user-friendly, to allow url as input for making predictions (as opposed to direct file upload), and to include jupyter notebook files ready to deploy in the cloud.

To begin, extract all files into a common directory. In the same directory create a sub-directory called 'datasets'.
In here, you must put the training and test images, however you will want to combine all images into two files:
train_cars.h5 and test_cars.h5.
To create these two files, use the 'h5write.py' script found here: https://github.com/EvanEames/h5_ReadWrite (note that you will also need the 'devkit/train_perfect_preds.txt' and 'devkit/test_perfect_preds.txt' files to create the h5 files. These txt files should be packaged with the cars dataset).
Once you have created them, place these h5 files in the datasets directory, and then run CarModel.py.

You will also need to download Resnet152 weights pre-trained on the CIFAR-10 dataset for transfer learning. Just google "resnet152_weights_tf.h5" and you should be able to find them. 

Once the weights are well trained, use the guess.py file to make predictions based on url images.