rnn_model.py 文件源码

def rnn_model(self):
        # BasicLSTMCell ????LSTM?????????forget_bias(????1)?????????????????????????
        # ??????????????????peep-hole????????
        # BasicLSTMCell ????? rnn.python.ops?? core_rnn_cell_impl.py
        cell = rnn.BasicLSTMCell(num_units=self.n_units)

        # MultiRNNCell ????????????????????RNN????????????????????
        # ??????????True    state_is_tuple = True
        # ???????LSTM??????????????????
        multi_cell = rnn.MultiRNNCell([cell]*self.n_layers)

        # we only need one output so get it wrapped to out one value which is next word index
        # ? rnn_cell ??????????? output_size?????size ?????Output_projection?rnn_cell
        cell_wrapped = rnn.OutputProjectionWrapper(multi_cell, output_size=1)

        # get input embed
        # tf.random_uniform(shape, minval, maxval, dtype, seed, name) ? ???? n*n????????minval ? maxval ??
        embedding = tf.Variable(initial_value=tf.random_uniform([self.vocab_size, self.n_units], -1.0, 1.0))

        # tf.nn.embedding_lokkup(embedding, inputs_id) : ??inputs_id??embedding??????????input_ids=[1,3,5],?
        # ??embedding????1,3,5????????????
        inputs = tf.nn.embedding_lookup(embedding, self.inputs)
        # what is inputs dim??

        # add initial state into dynamic rnn, if I am not result would be bad, I tried, don't know why
        if self.labels is not None:

        # zero_state ? ?????
            initial_state = cell_wrapped.zero_state(int(inputs.get_shape()[0]), tf.float32)
        else:
            initial_state = cell_wrapped.zero_state(1, tf.float32)

        # dynamic_rnn ?????????????batch?????????????????batch????????????????
        # dynamic_rnn ? rnn ??
        outputs, states = tf.nn.dynamic_rnn(cell_wrapped, inputs=inputs, dtype=tf.float32, initial_state=initial_state)
        outputs = tf.reshape(outputs, [int(outputs.get_shape()[0]), int(inputs.get_shape()[1])])

        # truncated_normal : ???????????
        w = tf.Variable(tf.truncated_normal([int(inputs.get_shape()[1]), self.vocab_size]))
        b = tf.Variable(tf.zeros([self.vocab_size]))

        logits = tf.nn.bias_add(tf.matmul(outputs, w), b)
        return logits, states