Project: 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.

Tain set - DOWNLOAD. Just download it and upload it to Colab. Don’t use curl since it manages to mess up zipped files!

For your submission create a new github repo and upload code/notebooks and the final model (.h5 file). Next, try to go to the link flatland evaluation and you should see the message ‘Welcome to Flatland!’. This means that the evaluation service is running and you can submit your own model by calling https://us-central1-aiprimer.cloudfunctions.net/flatland?model_link=[PATH TO YOUR .h5] (be patient, it can take a while).

Evaluation script and corrects labels as follows:

import numpy as np

data = np.load('flatland_train.npz')
X = data['X']
y = data['y']

y[y != 0] -= 2    # Correct labels so that triangle is mapped to class 1
X = X / 255.      # Scale down to range [0, 1]

# Construct and train your model (don't forget train/test split and other tricks)
model = ...

# Save the model and upload it to git
model.save('model.h5')

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

Leaderboard

Submissions will end up in a leaderboard that will be shared during the lectures. Each model will be evaluated on two test sets - one that closely matches the training set and a slightly more advanced one whose nature will be revealed at the end of the course. To know your grade look at the advanced one!

Passing benchmarks:

Points (out of 4)	Hint	Test set (adv.)
0	???	<70%
2	FNN	>70%
3	CNN	>80%
4	???	>90%

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