mintusf Goto Github PK

auto_object_removal

An application to automatically remove selected objects from images and videos

dnn-in-numpy

Implementation of DNN algorithm with using only the numpy library. Model is trained using batch gradient descent and tested on MNIST dataset achieving 94% accuracy on test set. Dropout and L2 are implemented as regularization algorithms. Optimizer can be chosen between gradient descent and Adam. The model is fully scallable, which means that regularization parameter as well as numbers of hidden layers and nodes can be set to any value. All functions, including forward propagation, back propagation, cross entropy loss calculation, dropout and training algorithm are written without the tensorflow library. Different activation function can be used: sigmoid, Relu or hyperbolic tangent.

keras-yolo3

A Keras implementation of YOLOv3 (Tensorflow backend)

kernel-k-means-with-dimensions-reduction

The following program applied k-means algorithm for unsupervised learning with sklearn circles dataset. To create a decision boundary in the case of non-linear dataset, Gaussian kernels are utilized. In order to make the model computationally less expensive, dimensions reduction using principal component analysis was applied. All functions, including model building and trainins, kernels calculation and dimensions reduction were written from scratch without using any built-in sklearn functions.

land_cover_segmentation

Segmentation for land cover using raw satellite data (Sentinel 2La)

land_cover_tracking

Land cover tracking Web App

mintusf

recommender-system-with-collaborative-filtering

The program has two main functionalities. The first one is building a regularized recommender system with collaborative filtering and learning it with a "MovieLens" dataset (https://grouplens.org/datasets/movielens/) with 1 millions movie ratings. The second feature is finding recommendations for a new user, which is done by implementation of a content-based recommender system. A user, after inputting some ratings to the "Movies_fornewuser.txt" file, can use provided model, which was learnt using the first feature, to find new movies recommendations. In order to use, please read README.md.

semisupervised_classification

Case study for semi-supervised learning using imbalanced CIFAR 1- dataset

sentiment_analysis

A sentiment analysis model trained on Amazon reviews dataset.

traffic-signs-recognition

The project includes two solutions for a traffic signs recognition. Original dataset is available at https://www.kaggle.com/valentynsichkar/traffic-signs-preprocessed. The first solution utilizes tensorflow framework to build a scalable Deep Neural Network to recognize 43 different traffic signs. Its accuracy with the cross validation set reaches 90%. Algorithm implements batch normalization, learning rate decay and dropout. For cost function minimization it uses minibatch Adam optimizer. In the second solution, convolutional layers are used and a model based on VGG-16 architecture is built. Its accuracy with the cross validation set reaches 95%. Algorithm implements batch normalization, learning rate decay and dropout. For cost function minimization it uses minibatch Adam optimizer.