Created by: lexy0093

组网：

        label = fluid.layers.data(name='label', shape=[1], dtype='int64', lod_level=0)
        index = fluid.layers.data(name='index', shape=[1], dtype='int64', lod_level=1)
        feature = fluid.layers.data(name='feature', shape=[1], dtype='float32', lod_level=1)

        droped = fluid.layers.dropout(x=feature, dropout_prob=self.dropout_prob)
        emb = fluid.layers.embedding(input=index, size=[self.vocab_dim, self.emb_dim], is_sparse=True)
        stk = fluid.layers.expand(feature, expand_times = [1, 256])
        mm = fluid.layers.elementwise_mul(x=emb, y=stk)
        mm_fc = fluid.layers.sequence_pool(input=mm, pool_type='sum')
        fc1 = fluid.layers.relu(x=mm_fc)

        fc2 = fluid.layers.fc(input=fc1, size=self.hid_dim, act='relu')
        pred = fluid.layers.fc(input=[fc1, fc2], size=self.class_dim, act='softmax')

特征维度：80w， batch_size=1024，每轮训练包括：train data 和test data，训练开始每个batch执行大概稳定25s，快结束时batch运行时延增加，并且在test结束后，第二轮batch时延也加大，严重影响运行效率。 日志： http://10.73.173.35:8910/fileview.html?type=logsdir&path=/&instance=0.app-user-20190716103040-38241--mcdnn-fluid_paddlecloud