42 project about identifying leaves
Following this keras tutorial by Victor Zhou, we implemented a basic CNN with one Conv2D, one MaxPolling and some fully-connected layers before the softmax.
This CNN could easily be improved as we slightly overfit on some models. Adding a Drop layer should stop overfitting.
More layers (activation of ) will also improve accuracy but will take more time for training.
We chose not to go further for the CNN as we already reached the 90% accuracy needed to validate the project (94% with hard voting).
model1_pred = [0.9, 0.5, 0.002]
model2_pred = [0.1, 0.8, 0.003]
model3_pred = [0.1, 0.8, 0.003]
avg for each class then pick best soft_vote_avg = [0.37, 0.7, 0.0027] soft_vote_pred = 1
pick best out of all pred
hard_vote_pred = 0