Code for Stanford CS224D: deep learning for natural language understanding
The interesting part is Assignment 3, implementing recursive (tree-structured) neural networks using tf.while_loop for 16x faster training and 8x faster inference.
Assignments 1 and 2 are the standard solutions for class assignments, with tiny variations.