lstm-cell — Lyft case

Lyft-inspired ETA model processes a sequence of trip events with a compact recurrent unit. You need to perform one LSTM cell update from the already-activated input, forget and output gates, the candidate values and the previous cell state.

Compute one LSTM cell update from already activated gate values: i, f, o are the input, forget and output gate activations (in $(0,1)(0,1)$), g is the candidate cell value and c_prev is the previous cell state. The new cell state is $c=f\odot c_{prev}+i\odot gc=f\backslash odot\; c\_\{prev\}+i\backslash odot\; g$ and the new hidden state is $h=o\odot \tanh (c)h=o\backslash odot\backslash tanh(c)$.

Implement solve(i,f,o,g,c_prev).

Returns. Return a tuple (h, c) of NumPy vectors.

Examples

Example 1

Input:

solve([0.5, 0.5], [0.9, 0.9], [1.0, 1.0], [0.2, -0.2], [0.0, 0.0])

Output:

([0.099668, -0.099668], [0.1, -0.1])

Example 2

Input:

solve([0.0], [1.0], [1.0], [5.0], [2.0])

Output:

([0.964028], [2.0])