rnn_model_numpy.py

class Inferer:
    def __init__(self, model):
        parameters = list(model.lstm.parameters())
        
        W_ii, W_if, W_ig, W_io = parameters[0].split(HIDDEN_SIZE)
        W_hi, W_hf, W_hg, W_ho = parameters[1].split(HIDDEN_SIZE)
        b_ii, b_if, b_ig, b_io = parameters[2].split(HIDDEN_SIZE)
        b_hi, b_hf, b_hg, b_ho = parameters[3].split(HIDDEN_SIZE)
        
        self.W_ii = W_ii.detach().numpy()
        self.b_ii = b_ii.detach().numpy()
        
        self.W_hi = W_hi.detach().numpy()
        self.b_hi = b_hi.detach().numpy()
        
        self.W_if = W_if.detach().numpy()
        self.b_if = b_if.detach().numpy()
        
        self.W_hf = W_hf.detach().numpy()
        self.b_hf = b_hf.detach().numpy()
        
        self.W_ig = W_ig.detach().numpy()
        self.b_ig = b_ig.detach().numpy()
        
        self.W_hg = W_hg.detach().numpy()
        self.b_hg = b_hg.detach().numpy()
        
        self.W_io = W_io.detach().numpy()
        self.b_io = b_io.detach().numpy()
        
        self.W_ho = W_ho.detach().numpy()
        self.b_ho = b_ho.detach().numpy()
        
        self.W = model.fc.weight.detach().numpy().T
        self.b = model.fc.bias.detach().numpy()
        
    def infer(self, x):
        x_t, h_t, c_t = None, np.zeros(HIDDEN_SIZE), np.zeros(HIDDEN_SIZE)
        for i in range(x.shape[0]):
            x_t = x[i]
            _, h_t, c_t = self.lstm_cell(x_t, h_t, c_t)
        r = np.dot(h_t, self.W) + self.b
        return self.sigmoid(r)
    
    def lstm_cell(self, x_t, h_tm1, c_tm1):
        i_t = self.sigmoid(
            np.dot(self.W_ii, x_t) + self.b_ii + np.dot(self.W_hi, h_tm1) + self.b_hi
        )
        f_t = self.sigmoid(
            np.dot(self.W_if, x_t) + self.b_if + np.dot(self.W_hf, h_tm1) + self.b_hf
        )
        g_t = np.tanh(
            np.dot(self.W_ig, x_t) + self.b_ig + np.dot(self.W_hg, h_tm1) + self.b_hg
        )
        o_t = self.sigmoid(
            np.dot(self.W_io, x_t) + self.b_io + np.dot(self.W_ho, h_tm1) + self.b_ho
        )
        c_t = f_t * c_tm1 + i_t * g_t
        h_t = o_t * np.tanh(c_t)
        return o_t, h_t, c_t
    
    @staticmethod
    def sigmoid(x):
        return 1 / (1 + np.exp(-x))