Created by: LihangLiu

使用layers.embedding的时候，发现is_sparse=True 比 is_sparse=False 还慢很多。 代码如下，环境为fluid1.1。net函数中用到embedding。

import paddle
import paddle.fluid as fluid
from paddle.fluid import layers
from paddle.fluid.param_attr import ParamAttr
import numpy as np
import time

def fluid_create_lod_tensor(array, lod, place):
    tensor = fluid.LoDTensor()
    tensor.set(np.array(array), place)
    tensor.set_lod(lod)
    return tensor

def net():
    x = layers.data(name='x', shape=[-1, 1], dtype='int64')
    xavier_initializer = fluid.initializer.Xavier()
    param_clip = fluid.clip.GradientClipByValue(1.0)
    embed = layers.embedding(x,
                    size=[1000000, 16],
                    param_attr=ParamAttr(name='embed',
                                        initializer=xavier_initializer,
                                        gradient_clip=param_clip
                                        ),
                    is_sparse=False)
                    # is_sparse=True)
    out = layers.fc(embed, size=1)
    loss = layers.reduce_mean(layers.square(out))
    opt = fluid.optimizer.AdamOptimizer(
        learning_rate=0.01, beta1=0.9, beta2=0.999)
    opt.minimize(loss)
    return loss

                                
def train():
    myprogram = fluid.Program()
    with fluid.program_guard(myprogram):
        out = net()
    test_program = myprogram.clone(for_test=True)
    
    place = fluid.CPUPlace()
    exe = fluid.Executor(place)
    exe.run(fluid.default_startup_program())

    # data
    x_data = np.ones([10, 1]).astype('int64')
    x_lod = []
    x_tensor = fluid_create_lod_tensor(x_data, x_lod, place)

    time_list = []
    for _ in range(10):
        s = time.time()
        results = exe.run(myprogram, feed={'x': x_tensor}, fetch_list=[out])
        time_list.append(time.time() - s)
    print ('time', np.mean(time_list))

                                
if __name__ == "__main__":
    train()