AI for Airplane classification (ai2r)

Training

Assuming that you have got the requirements setup on your local system, in a virtual environment or in a docker image; start by python ai2r.py -d dataset_name -m model_name --gpu 1. Alternatively you can set your desired paramters in "Train.sh" and run this bash file.

• dataset_name is the name of the dataset which has to be put in raw directory. If raw data does not exists or you want to use current ai2r dataset the datset will be downloaded and extracted in a new directory "datasets". You will see the details of this process in the prompt messages.
• model_name is the name of a model to be used. This should be defined in models directory. Currently this can be set to one of:

  inceptionv3_pretrained
  vgg16_pretrained

For example: python ai2r.py -d aircrafts -m inceptionv3_pretrained --gpu 1 will train the aircrafts dataset using inceptionv3 with trained weights from imagenet.

If this is the first time, this command will trigger downloading the aircrafts dataset and the corresponding labels in .npy format. Once download is complete the code will unzip the data and add it to a directory i.e., "datasets". The training will start with the configured parameters in "config.json". Once the training is done the results will be saved under out directory. The results include the trained model, binarized labels in pickle format, a summary of the training and a .png graph of how the training/validation progressed in different epochs

Training using docker:

• cd to the ai2r project directory.
• Specify the model and dataset to be used in Train.sh script.
• Run bash runDocker.sh gpu train. This will run a docker container with the requirements installed. It will also mount the current directory to /ai2r directory in the docker so that the docker container has access to Train.sh. The docker image will run Train.sh in the same manner as if you run ai2r.py on your system or in virtual environment.
• To use a docker image with CPU version of keras/tensorflow for training use bash runDocker.sh cpu train.

Prediction

Assuming some training is done and results are available under "out" directory, you can start by python predict.py -i path_to_image. This will scan the out directory for trained models and a prompt will ask user for the desired model to be used for prediction.

Prediction using docker:

• cd to the ai2r project directory.
• Add the path to the image you would like to run a prediction on to the Predict.sh script.
• Run bash runDocker.sh. This will run a docker container with the requirements installed. It will also mount the current directory to /ai2r directory in the docker so that the docker container has access to Predict.sh. The docker image will run Predict.sh in the same manner as if you run predict.py on your system or in virtual environment.
• Running bash runDocker.sh by default will use a GPU version docker image and run prediction on the test image specified in Predict.sh. To use a docker image with CPU version of keras/tensorflow for prediction use bash runDocker.sh cpu

Using the Flask api

The code for the Flask api is located under "api" directry. Any source code related to models training or prediction is on the root director of the project or under "src" directory. Trained models available to the api must be added to "deployment/model".

Testing Flask api

The required Postman collection and environment are available in the root directory of the project. These files can be imported into Postman to test or try out the api.

Data

The data is scraped from the web. Currently it contains samples for 13 types of airplanes mostly airlines. Data is not exactly balanced but it is good enough to fly! Here is a list of airplanes in the data:

Type	Manufacturer	Code	#Samples
A330	Airbus	a330	471
A380	Airbus	a380	747
B737	Boeing	b737	463
B757	Boeing	b757	684
Citation II	Cessna	c550	421
concord	Concord	conc	325
DC-10	Douglas	dc10	573
50	Fokker	f50	523
100	Fokker	f100	482
Falcon 900	Dassault	f900	613
Gulfstream V	glf5	Gulfstream	466
Jetstream 31	British Aerospace	js31	525
Yak-40	Yakovlev	yk40	531

Requirements:

Refer to requirements.txt