Skip to content

Homeworks for the course Artificial Neural Networks & Deep Learning @ PoliMi.

Notifications You must be signed in to change notification settings

jtonglet/Deep-Learning-Homework

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

40 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Artificial Neural Networks and Deep Learning homeworks

This repository contains the jupyter notebooks used for the two homeworks of the Artificial Neural Networks & Deep Learning course given by Professors Matteuci and Boracchi at Politecnico di Milano.

Homework 1 : Leaf Classification 🍂🍃🍁🌿

Codalab

Overview

The first homework took the form of an image classification competition on Codalab. We were supplied with the train and validation set. Then, we had to submit our models on the Codalab platform to evaluate them on the hidden test set. The dataset contains pictures of leaves belonging to 14 different species and is characterized by a strong class imbalance. In particular, tomato leaves are significantly overrepresented, as shown on the following picture :

distribution

To solve the problem, we used a weighted ensemble of 5 pre-trained models (EfficientNetB3, VGG16, ResNet50, Xception and DenseNet121). The models were fine-tuned on the train set. The ensemble achieved an accuracy of 90.57% on the hidden test set.

Homework 2 : Forecasting 🔮🧙‍♂️⏳

Codalab

Overview

The second homework consisted in a multivariate time series forecasting task. The dataset includes 7 time series. The goal was to design and implement forecasting models to learn how to exploit past observations in the input sequence to correctly predict the next 864 time steps. Predictions were evaluated with the RMSE on a hidden test set. By using the N-BEATS model we achieved a final RMSE score of 3.79.

The forecast for the first tree time series are shown below :

forecast values

Reference

Oreshkin, B. N., Carpov, D., Chapados, N., & Bengio, Y. (2019). N-BEATS: Neural basis expansion analysis for interpretable time series forecasting.