Learning phrase grounding from captioned images through InfoNCE bound on mutual information - View it on GitHub