Learning Phrase Representations: One Full GRU Step
Advanced
The problem, from first principles
Putting the pieces together gives one full GRU step. The update gate decides how much of the candidate to take, and the candidate already uses the reset gate. The output blends the old state and the candidate.
From theory to code
Implement gru_step(x, h, Wz, Uz, Wr, Ur, W, U), computing the update gate, the reset gate, the candidate, and the new state.
Constraints
- All weight matrices have the shapes shown in the docstring.
Hints
Hint 1
Compute z, r, h_tilde, then return (1 - z) * h + z * h_tilde.
Keywords
research-paperstransformersllmrecurrentgru
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.