Description

Multi-Label Image Classification with PyTorch: Image Tagging

Multi-label classification assigns to each sample a set of target labels, whereas multi-class classification makes the assumption that each sample is assigned to one and only one label out of the set of target labels.The key difference is that multi-output classification always predicts a fixed-length set of labels per sample and can be theoretically replaced with the corresponding number of separate classifiers while multi-label classification requires predicting non-fixed length subset of labels. The main purpose of this project is to demonstrate to deal with multi-label classification, let’s try to make our life a bit easier and solve one task at a time. In order to do that, we decided to keep only the 27 most frequent labels so that each label has more than 100 corresponding samples. Also, for fast training, we took the first 6000 samples from the dataset. We split them to train and test subsets using a 5 to 1 ratio. The resulting size of the training set is 5000 images, and the test set – 1000 images.The Pytorch’s Dataset implementation for the NUS-WIDE is standard and very similar to any Dataset implementation for a classification dataset.