def forward(self, inputs):
x, lstm_state = inputs
# LSTM state consists of c and h.
c, h = jnp.split(lstm_state, 2, axis=-1)
# Dense layer on the concatenation of x and h.
w, b = self.weights
y = jnp.dot(jnp.concatenate([x, h], axis=-1), w) + b
# i = input_gate, j = new_input, f = forget_gate, o = output_gate
i, j, f, o = jnp.split(y, 4, axis=-1)
new_c = c * fastmath.sigmoid(f) + fastmath.sigmoid(i) * jnp.tanh(j)
new_h = jnp.tanh(new_c) * fastmath.sigmoid(o)
return new_h, jnp.concatenate([new_c, new_h], axis=-1)
def forward(self, inputs):
x, gru_state = inputs
# Dense layer on the concatenation of x and h.
w1, b1, w2, b2 = self.weights
y = jnp.dot(jnp.concatenate([x, gru_state], axis=-1), w1) + b1
# Update and reset gates.
u, r = jnp.split(fastmath.sigmoid(y), 2, axis=-1)
# Candidate.
c = jnp.dot(jnp.concatenate([x, r * gru_state], axis=-1), w2) + b2
new_gru_state = u * gru_state + (1 - u) * jnp.tanh(c)
return new_gru_state, new_gru_state