This repo contains the code used to train a few shot model on the Fine-Grained Visual Classification (FGVC) of Aircraft dataset. It also contains code of a simple web-service that can be used to demonstrate a model's prediction results with user uploaded images, as well as code to containerise the web-service into a Docker image.

A sample model may be found in the docker\app folder; the file is named model.h5. This model was created by fine-tuning the VGG16 pretrained model on the FGVC aircraft dataset.

Environments

The codebase was developed with Python 3.9. Install requirements as follows:

pip install -r path\to\docker\requirements.txt

Preprocessing

To train a few shot model, the following needs to be done:

Ensure all images are of similar dimensions (and converted to 3 channel images if grayscale)
Four .h5 files - train_pairs.h5, val_pairs.h5 and train_labels.h5, val_labels.h5 - are to be created
- The pairs.h5 files contain pairs of images as np arrays to be fed into our siamese network, while labels.h5 files contain a single np array of 1s and 0s that correspond to the labels of the image pairs (1 if they are of the same class, and 0 otherwise)
Two .h5 files - X_test.h5 and Y_test.h5 - are to be created for testing the model
- X_test contains individual images (3 channel and similar dimensions to train/validation images) as np arrays
- Y_test is a np array that contains the class names/labels corresponding to each image in X_test
Save the files with a similar h5 dataset name (i.e. 'train_pairs' for train_pairs.h5, etc) for loading purposes

The preprocess.py utility script in the Utils folder may provide some help in padding/resizing images and pairing up the images.

Training/Testing a Few Shot Model

Make sure to have done the required preprocessing as explained above.

Then, in the command prompt, run the following:

python path\to\main\main.py

main.py is an interactive script, so follow the onscreen instructions to either create a new model, train an existing model, or evaluate it on various one-shot tasks!

Inference

In the command prompt, run the follownig:

python path\to\main\inference.py <path to images folder> <path to model.h5 file>

Do ensure that your images folder is formatted like so:

folder  
   └── query  
       ├── foo.png   
       └── bar.jpg  
   └── support  
       ├── label1_foo.png  
       ├── label1_bar.jpg  
       ├── label2_foo.png  
       └── label2_bar.png 
       ...

See the example folder for more details.
The program will pair the query images with each of the support images and output its similarity score.

Running the Web-Service

In the command prompt, run the following:

python path\to\docker\app\main.py

Containerising the Web-Service