rn_gru_encode.py 文件源码

def _forward_step(self, x_t, s_t):
        """Input vector/matrix x(t) and state matrix s(t)."""

        # Gradient clipping
        E, a, U, W, b, V, c = [ th.gradient.grad_clip(p, -5.0, 5.0) for p in self.params.values() ]

        # Initialize state to return
        s_next = T.zeros_like(s_t)

        # Vocab-to-state encoding layer
        inout = T.tanh(T.dot(x_t, E) + a)

        # Loop over GRU layers
        for layer in range(self.hyper.layers):
            # 3 matrices per layer
            L = layer * 3
            # Get previous state for this layer
            s_prev = s_t[layer]
            # Update gate
            z = T.nnet.hard_sigmoid(T.dot(inout, U[L]) + T.dot(s_prev, W[L]) + b[L])
            # Reset gate
            r = T.nnet.hard_sigmoid(T.dot(inout, U[L+1]) + T.dot(s_prev, W[L+1]) + b[L+1])
            # Candidate state
            h = T.tanh(T.dot(inout, U[L+2]) + T.dot(r * s_prev, W[L+2]) + b[L+2])
            # New state
            s_new = (T.ones_like(z) - z) * h + z * s_prev
            s_next = T.set_subtensor(s_next[layer], s_new)
            # Update for next layer or final output (might add dropout here later)
            inout = s_new

        # Final output
        o_t = T.dot(inout, V) + c
        return o_t, s_next

    # Regularization cost