gated_rnn_test.py 文件源码

def testComputation(self):
    batch_size = 2
    hidden_size = 4
    inputs = tf.placeholder(tf.float32, shape=[batch_size, hidden_size])
    prev_cell = tf.placeholder(tf.float32, shape=[batch_size, hidden_size])
    prev_hidden = tf.placeholder(tf.float32, shape=[batch_size, hidden_size])
    lstm = snt.LSTM(hidden_size)
    _, next_state = lstm(inputs, (prev_hidden, prev_cell))
    next_hidden, next_cell = next_state
    lstm_variables = lstm.get_variables()
    param_map = {param.name.split("/")[-1].split(":")[0]:
                 param for param in lstm_variables}

    # With random data, check the TF calculation matches the Numpy version.
    input_data = np.random.randn(batch_size, hidden_size)
    prev_hidden_data = np.random.randn(batch_size, hidden_size)
    prev_cell_data = np.random.randn(batch_size, hidden_size)

    with self.test_session() as session:
      tf.global_variables_initializer().run()
      fetches = [(next_hidden, next_cell),
                 param_map[snt.LSTM.W_GATES],
                 param_map[snt.LSTM.B_GATES]]
      output = session.run(fetches,
                           {inputs: input_data,
                            prev_cell: prev_cell_data,
                            prev_hidden: prev_hidden_data})

    next_state_ex, gate_weights_ex, gate_biases_ex = output
    in_and_hid = np.concatenate((input_data, prev_hidden_data), axis=1)
    real_gate = np.dot(in_and_hid, gate_weights_ex) + gate_biases_ex
    # i = input_gate, j = next_input, f = forget_gate, o = output_gate
    i, j, f, o = np.hsplit(real_gate, 4)
    real_cell = (prev_cell_data / (1 + np.exp(-(f + lstm._forget_bias))) +
                 1 / (1 + np.exp(-i)) * np.tanh(j))
    real_hidden = np.tanh(real_cell) * 1 / (1 + np.exp(-o))

    self.assertAllClose(real_hidden, next_state_ex[0])
    self.assertAllClose(real_cell, next_state_ex[1])