本文记录参加“深度学习7日入门-CV疫情特辑”课程中的“人流密度估计”比赛所用的算法，虽然并没有达到很好的名次，但是初识了PaddlePaddle，更重要的是还有免费V100 GPU算力可以蹭，正好后面会有一些项目，想要在AI Studio上用免费GPU进行验证，所以在此记录下这次比赛所用的算法，顺便攒一攒算力卡~~

    本次用的是MSCNN，基于人流密度图来估计人流数量的算法。MSCNN可以说是人流密度估计方法的始祖（古董）了，但是很有助于理解人流密度估计关键模型的原理，何况本人也是刚刚接触Paddle，就先拿来练练手吧。

直接上代码：

def MSCNN(x):
    x = fluid.layers.conv2d(input=x, num_filters=64, filter_size=9, padding=4, act='relu')
    x1_1 = fluid.layers.conv2d(input=x, num_filters=64, filter_size=9, padding=4, act='relu')
    x1_2 = fluid.layers.conv2d(input=x, num_filters=64, filter_size=7, padding=3, act='relu')
    x1_3 = fluid.layers.conv2d(input=x, num_filters=64, filter_size=5, padding=2, act='relu')
    x1_4 = fluid.layers.conv2d(input=x, num_filters=64, filter_size=3, padding=1, act='relu')
    x = fluid.layers.concat(input=[x1_1, x1_2, x1_3, x1_4], axis=1)
    x = fluid.layers.batch_norm(input=x, act=None)

    x = fluid.layers.pool2d(input=x, pool_size=2, pool_stride=2, pool_type='max')

    x2_1 = fluid.layers.conv2d(input=x, num_filters=128, filter_size=9, padding=4, act='relu')
    x2_2 = fluid.layers.conv2d(input=x, num_filters=128, filter_size=7, padding=3, act='relu')
    x2_3 = fluid.layers.conv2d(input=x, num_filters=128, filter_size=5, padding=2, act='relu')
    x2_4 = fluid.layers.conv2d(input=x, num_filters=128, filter_size=3, padding=1, act='relu')
    x = fluid.layers.concat(input=[x2_1, x2_2, x2_3, x2_4], axis=1)
    x = fluid.layers.batch_norm(input=x, act=None)
    x3_1 = fluid.layers.conv2d(input=x, num_filters=128, filter_size=9, padding=4, act='relu')
    x3_2 = fluid.layers.conv2d(input=x, num_filters=128, filter_size=7, padding=3, act='relu')
    x3_3 = fluid.layers.conv2d(input=x, num_filters=128, filter_size=5, padding=2, act='relu')
    x3_4 = fluid.layers.conv2d(input=x, num_filters=128, filter_size=3, padding=1, act='relu')
    x = fluid.layers.concat(input=[x3_1, x3_2, x3_3, x3_4], axis=1)
    x = fluid.layers.batch_norm(input=x, act=None)

    x = fluid.layers.pool2d(input=x, pool_size=2, pool_stride=2, pool_type='max')

    x4_1 = fluid.layers.conv2d(input=x, num_filters=64, filter_size=7, padding=3, act='relu')
    x4_2 = fluid.layers.conv2d(input=x, num_filters=64, filter_size=5, padding=2, act='relu')
    x4_3 = fluid.layers.conv2d(input=x, num_filters=64, filter_size=3, padding=1, act='relu')
    x = fluid.layers.concat(input=[x4_1, x4_2, x4_3], axis=1)
    x = fluid.layers.batch_norm(input=x, act='relu')

    x5_1 = fluid.layers.conv2d(input=x, num_filters=64, filter_size=7, padding=3, act='relu')
    x5_2 = fluid.layers.conv2d(input=x, num_filters=64, filter_size=5, padding=2, act='relu')
    x5_3 = fluid.layers.conv2d(input=x, num_filters=64, filter_size=3, padding=1, act='relu')
    x = fluid.layers.concat(input=[x5_1, x5_2, x5_3], axis=1)
    x = fluid.layers.batch_norm(input=x, act='relu')

    x = fluid.layers.pool2d(input=x, pool_size=2, pool_stride=2, pool_type='max')

    x = fluid.layers.conv2d(input=x, num_filters=1000, filter_size=1, act='relu')
    x = fluid.layers.conv2d(input=x, num_filters=1, filter_size=1, act=None)
    x = fluid.layers.sigmoid(x)
    x = fluid.layers.relu(x)

    print('output x.shape:', x.shape)

    return x

可以看到，这个实际上可以理解成是多尺度的一个VGG，并且通过使用合适的filter_size，padding，使得卷积层的输入输出的宽高不变，所以输入网络的图像，最终的大小就由pooling层决定。因为使用了3个2*2的pooling，所以实际上输入的图像，输出的宽高是降采样了8倍，也就是宽高变为原来的八分之一。

技术心得的提交时间快截至了，就先写这么多，后面再来分享。。。。