In Homework 2 we are going to train models that can recommend hashtags based on the content of an image. We will use the dataset "HARRISON". Please refer to the paper (https://arxiv.org/abs/1605.05054) for more details about the dataset.

The original data of HARRISON can be downloaded here: (https://drive.google.com/file/d/1B9NZf42J_GpslRNlTTAxvM5c9WRVl_1Z/view?usp=sharing). You can download the images and take a look, but this is not mandatory. We already extracted the image features for you and upload to Kaggle as "harrison_features.npz".

The demo code separates the training process into two stages: word2vec training and image-to-word2vec training. The notebook of last semester can be found here: https://www.kaggle.com/kuantinglai/train-devise
You can export and import the code to this competition.

To evaluate the model performance, the top 10 hashtags (KNN=10) recommended from your model will be used to calculate the hamming loss of ground truth.  Get your computer ready and have fun!