I built an image recognition system to tackle Kaggle's invasive species competition, in which I tried to idenfity whether or not an invasive species of plant are present in an image.
I begin by finetuning the VGG network (link to a blog post where I describe the process), a neural network which has been pre-trained on the ImageNet corpus.
I then finetune 2 other pre-trained neural networks, ResNet (which I train over two notebooks, ResNet and ResNet2) and V3.
I combine the outputs of the 3 neural networks in Ensemble1, and then use scikit-learn to predict whether or not an image contains an invasive species. (link to a blog post where I describe training the other networks, and the ensembling).
Specifically, I use the optunity module to find the best algorithm, which is a SVM with a polynomial kernel.