Train Deep Learning Models On 25,000 images + Keywords

We have a database of 25,000 high quality images, with about 20 keywords per image. The keywords include purely descriptive elements, such as "lion", "table", or "eating", "running", but also more conceptual notions, such as "joy", "loneliness", "friendship". We need you to do 3 things: 1- Separate the conceptual keywords from the others, using a Python or R script with appropriate NLP libraries 2- Define and train a first model to predict the "concepts", based on the other keywords. 3- Define and train a second model based on the images themselves. For 2- and 3-, we expect you to use a Deep Learning library, such as Tensorflow. We will ask you to provide the model and and script used to prepare the data (keywords/images). The training should be done by you. If needed, we can provide AWS instances. However, please remember that GPUs available on AWS don't have enough memory for recent networks such as VGG. Please avoid generic bids. In order to show that you have read and understood this text, please start you bid with “Deep learning projet”. Please explain what library you intend to use. If you think you can do 2- and 3- together, please explain how, this will be still better. This is urgent, we will make a decision in the next 2-3 days.