Method model

chatbotv3/encoder_decoder_seq2seq.py:176–238

(self, x, y, weights, biases, training=True)

        return XY, Y

    def model(self, x, y, weights, biases, training=True):
        # 注：以下的6是one_hot_word_vectors_dim
        # 取第一个样本的ABC
        encoder_inputs = tf.slice(x, [0, 0, 0], [1, self.max_seq_len, self.word_vec_dim])  # shape=(1, 8, 128)
        # 展开成2-D Tensor
        encoder_inputs = tf.unstack(encoder_inputs, self.max_seq_len, 1)  # [<tf.Tensor shape=(1, 128)>,...] 内含8个Tensor

        # 取第一个样本的<EOS>WXYZ
        decoder_inputs = tf.slice(x, [0, self.max_seq_len, 0], [1, self.max_seq_len, self.word_vec_dim])  # shape=(1, 8, 128)
        decoder_inputs = decoder_inputs[0]  # shape=(8, 128)
        # 转成解码器的输入输出形状
        decoder_inputs = tf.matmul(decoder_inputs, weights['enc2dec']) + biases['enc2dec']
        # 展开成2-D Tensor
        decoder_inputs = tf.unstack([decoder_inputs], axis=1)  # [<tf.Tensor shape=(1, 6)>,...] 内含8个Tensor

        # 取第一个样本的WXYZ
        target_outputs = tf.slice(y, [0, 0, 0], [1, self.max_seq_len, self.one_hot_word_vectors_dim])  # shape=(1, 8, 6)
        target_outputs = target_outputs[0]  # shape=(8, 6)

        # 构造网络结构：两层结构
        encoder_layer1 = rnn.BasicLSTMCell(self.n_hidden, forget_bias=1.0)
        encoder_layer2 = rnn.BasicLSTMCell(self.n_hidden, forget_bias=1.0)
        decoder_layer1 = rnn.BasicLSTMCell(self.n_hidden, forget_bias=1.0)
        decoder_layer2 = rnn.BasicLSTMCell(self.n_hidden, forget_bias=1.0)

        # 输入是8个shape=(1, 128)的Tensor，输出是8个shape=(1, 1000)的Tensor
        encoder_layer1_outputs, encoder_layer1_states = rnn.static_rnn(encoder_layer1, encoder_inputs, dtype=tf.float32, scope='encoder_layer1')
        # 输入是8个shape=(1, 1000)的Tensor，输出是8个shape=(1, 1000)的Tensor
        encoder_layer2_outputs, encoder_layer2_states = rnn.static_rnn(encoder_layer2, encoder_layer1_outputs, dtype=tf.float32, scope='encoder_layer2')
        # 取解码器输入的<EOS>
        # 输入是1个shape=(1, 6)的Tensor(<EOS>)，输出是1个shape=(1, 1000)的Tensor
        decoder_layer1_outputs, decoder_layer1_states = rnn.static_rnn(decoder_layer1, decoder_inputs[:1], initial_state=encoder_layer1_states, dtype=tf.float32, scope='decoder_layer1')
        # 输入是1个shape=(1, 1000)的Tensor，输出是1个shape=(1, 1000)的Tensor
        decoder_layer2_outputs, decoder_layer2_states = rnn.static_rnn(decoder_layer2, decoder_layer1_outputs, initial_state=encoder_layer2_states, dtype=tf.float32, scope='decoder_layer2')

        decoder_layer2_outputs_combine = []
        decoder_layer2_outputs_combine.append(decoder_layer2_outputs)
        for i in range(self.max_seq_len - 1):
            decoder_layer2_outputs = tf.unstack(decoder_layer2_outputs, axis=1)[0]
            decoder_layer2_outputs = tf.matmul(decoder_layer2_outputs, weights['hid2tar']) + biases['hid2tar'][i]
            # 输入是1个shape=(1, 6)的Tensor，输出是1个shape=(1, 1000)的Tensor
            if training:
                decoder_layer1_outputs, decoder_layer1_states = rnn.static_rnn(decoder_layer1, decoder_inputs[i+1:i+2], initial_state=decoder_layer1_states, dtype=tf.float32, scope='decoder_layer1')
            else:
                decoder_layer1_outputs, decoder_layer1_states = rnn.static_rnn(decoder_layer1, [decoder_layer2_outputs], initial_state=decoder_layer1_states, dtype=tf.float32, scope='decoder_layer1')
            # 输入是1个shape=(1, 1000)的Tensor，输出是1个shape=(1, 1000)的Tensor
            decoder_layer2_outputs, decoder_layer2_states = rnn.static_rnn(decoder_layer2, decoder_layer1_outputs, initial_state=decoder_layer2_states, dtype=tf.float32, scope='decoder_layer2')
            decoder_layer2_outputs_combine.append(decoder_layer2_outputs)

        # 下面的过程把8个shape=(1, 1000)的数组转成8个shape=(1, 1000)的Tensor
        decoder_layer2_outputs_combine = tf.unstack(decoder_layer2_outputs_combine, axis=1)[0]
        decoder_layer2_outputs_combine = tf.unstack(decoder_layer2_outputs_combine, axis=1)[0]
        decoder_layer2_outputs_combine = tf.unstack([decoder_layer2_outputs_combine], axis=1)
        # 重新对decoder_layer2_outputs赋值
        decoder_layer2_outputs = decoder_layer2_outputs_combine

        decoder_layer2_outputs = tf.unstack(decoder_layer2_outputs, axis=1)[0]  # shape=(8, 1000)
        decoder_layer2_outputs = tf.matmul(decoder_layer2_outputs, weights['hid2tar']) + biases['hid2tar']  # shape=(8, 6)