Submission Rules

The submission for this challenge consists of a single .h5 file containing the parameters of a EfficientNetLite0 neural network. This neural network architecture was selected due to is small size and efficiency, which makes it possible to run in space on board of OPS-SAT.

The easiest way to create a submission is by using Tensorflow and importing the neural network architecture like so:

from efficientnet_lite import EfficientNetLiteB0

using the Python module efficientnet_lite.py that we provide as part of our starter-kit on Gitlab.

Instantiate your model like this

# input_shape is (height, width, number of channels)
input_shape = (200, 200, 3)   
num_classes = 8
model = EfficientNetLiteB0(classes=num_classes, 
                           weights=None, 
                           input_shape=input_shape, 
                           classifier_activation=None)

and start training. Once you are done, you can generate a submission using

model.save_weights('submission.h5')

which you can now upload on our submission-page.

Users that prefer PyTorch over Tensorflow can make use of the conversion utility found in our starter-kit to generate equivalent .h5-files for submission. However, the evaluation (inference on unseen tiles of test-set) server-side is done using Tensorflow and Tensorflow Lite as OPS-SAT is setup to only run quantized Tensorflow Lite models.

After float16 quantization and inference, the Scoring of your submission is computed and you will enter our Leaderboard.

More details about the quantization of the model and the server-side computation of the final score(s) can be found in a dedicated notebook within our starter-kit.