Flatland

Task is slightly inspired by the book Flatland. You will have to classify images by ‘calculating’ a number of corners a figure in that image has using deep learning.

Train set contains pictures of the following shapes: circles, triangles, squares, pentagons, and hexagons.

1. DOWNLOAD train set and upload it to Colab. Don’t use curl since it manages to mess up zipped files!

import pickle, gzip

X, y = pickle.load(gzip.open(path_to_flatland_train_data, 'rb'))

2. Train the model and do your magic.

3. Create a new github repo and upload code/notebooks and the final model.

4. DOWNLOAD test set and run your model on it. Save the result as single long text using ''.join([str(round(p)) for p in pred]). So it should something like 6645533... but have 10k integers.

5. Finally go to evaluation app and submit your test set results and github link. Your submitions will end in the leaderboard.

Passing benchmarks:

Points (out of 4)	Hint	Test set
0	???	<80%
2	FNN	>80%
3	CNN	>90%
4	???	>95%

Additional task: extra 0.5 points are added if you manage to make a model that takes less than 500kb.

For faster training you can use colab, just change it to GPU mode by setting it at Edit -> Notebook settings -> Hardware accelerator.