深度学习笔记(七):如何用Mxnet来将神经网络可视化

最新推荐文章于 2024-08-14 18:30:34 发布

ZZY_dl

最新推荐文章于 2024-08-14 18:30:34 发布

阅读量272

点赞数

分类专栏：深度学习文章标签：神经网络可视化 mxnet

本文链接：https://blog.csdn.net/m0_51004308/article/details/113750483

版权

深度学习专栏收录该内容

19 篇文章 23 订阅

订阅专栏

在这里插入图片描述

文章目录

一:环境依赖

我自己的环境在这篇博客上https://blog.csdn.net/m0_51004308/article/details/112318578
下面是我为了实现可视化需要安装新的环境依赖包

step1：pip install mxnet-cu100 -i https://pypi.douban.com/simple
step2：pip install graphviz==0.8.4
并把F:\Anaconda3\envs\cv_car\Lib\site-packages\graphviz添加到电脑环境中
step3: pip install python-graphviz==0.16
https://graphviz.org/download/到这里下载Graphviz，我需要的是windows的版本所以我下的是下图这个

我安装在C:\Program Files\Graphviz，然后右键此电脑–高级系统设置–环境变量–在用户变量中添加C:\Program Files\Graphviz\bin\dot.exe，在系统变量中添加下面这些

二:具体代码

这个代码是实现MTCNN的四个级联网络，分别是P-Net、R-Net、O-Net、L-Net

import mxnet as mx



def pnet():
    data = mx.sym.Variable('data')
    conv1 = mx.sym.Convolution(data=data, kernel=(3, 3), num_filter=10, name='conv1')
    prelu1 = mx.sym.LeakyReLU(data=conv1, act_type='prelu', name='prelu1')
    pool1 = mx.sym.Pooling(data=prelu1, kernel=(2, 2), stride=(2, 2), pool_type='max', \
                           pooling_convention='full', name='pool1')
    conv2 = mx.sym.Convolution(data=pool1, kernel=(3, 3), num_filter=16, name='conv2')
    prelu2 = mx.sym.LeakyReLU(data=conv2, act_type='prelu', name='prelu2')
    conv3 = mx.sym.Convolution(data=prelu2, kernel=(3, 3), num_filter=32, name='conv3')
    prelu3 = mx.sym.LeakyReLU(data=conv3, act_type='prelu', name='prelu3')
    score = mx.sym.Convolution(data=prelu3, kernel=(1, 1), num_filter=2, name='score')
    prob = mx.sym.SoftmaxActivation(data=score, mode='channel', name='prob')
    bbox_pred = mx.sym.Convolution(data=prelu3, kernel=(1, 1), num_filter=4, name='bbox_pred')
    landmark_pred = mx.sym.Convolution(data=prelu3, kernel=(1, 1), num_filter=10, name='landmark_pred')
    out = mx.sym.Group([prob, bbox_pred, landmark_pred])
    return out


def rnet():
    data = mx.sym.Variable('data')
    conv1 = mx.sym.Convolution(data=data, kernel=(3, 3), num_filter=28, name='conv1')
    prelu1 = mx.sym.LeakyReLU(data=conv1, act_type='prelu', name='prelu1')
    pool1 = mx.sym.Pooling(data=prelu1, kernel=(3, 3), stride=(2, 2), pool_type='max', \
                           pooling_convention='full', name='pool1')
    conv2 = mx.sym.Convolution(data=pool1, kernel=(3, 3), num_filter=48, name='conv2')
    prelu2 = mx.sym.LeakyReLU(data=conv2, act_type='prelu', name='prelu2')
    pool2 = mx.sym.Pooling(data=prelu2, kernel=(3, 3), stride=(2, 2), pool_type='max', \
                           pooling_convention='full', name='pool2')
    conv3 = mx.sym.Convolution(data=pool2, kernel=(2, 2), num_filter=64, name='conv3')
    prelu3 = mx.sym.LeakyReLU(data=conv3, act_type='prelu', name='prelu3')
    fc = mx.sym.FullyConnected(data=prelu3, num_hidden=128, name='fc')
    prelu4 = mx.sym.LeakyReLU(data=fc, act_type='prelu', name='prelu4')
    score = mx.sym.FullyConnected(data=prelu4, num_hidden=2, name='score')
    prob = mx.sym.SoftmaxActivation(data=score, name='prob')
    bbox_pred = mx.sym.FullyConnected(data=prelu4, num_hidden=4, name='bbox_pred')
    landmark_pred = mx.sym.FullyConnected(data=prelu4, num_hidden=10, name='landmark_pred')
    out = mx.sym.Group([prob, bbox_pred, landmark_pred])
    return out


def onet():
    data = mx.sym.Variable('data')
    conv1 = mx.sym.Convolution(data=data, kernel=(3, 3), num_filter=32, name='conv1')
    prelu1 = mx.sym.LeakyReLU(data=conv1, act_type='prelu', name='prelu1')
    pool1 = mx.sym.Pooling(data=prelu1, kernel=(3, 3), stride=(2, 2), pool_type='max', \
                           pooling_convention='full', name='pool1')
    conv2 = mx.sym.Convolution(data=pool1, kernel=(3, 3), num_filter=64, name='conv2')
    prelu2 = mx.sym.LeakyReLU(data=conv2, act_type='prelu', name='prelu2')
    pool2 = mx.sym.Pooling(data=prelu2, kernel=(3, 3), stride=(2, 2), pool_type='max', \
                           pooling_convention='full', name='pool2')
    conv3 = mx.sym.Convolution(data=pool2, kernel=(3, 3), num_filter=64, name='conv3')
    prelu3 = mx.sym.LeakyReLU(data=conv3, act_type='prelu', name='prelu3')
    pool3 = mx.sym.Pooling(data=prelu3, kernel=(2, 2), stride=(2, 2), pool_type='max', \
                           pooling_convention='full', name='pool3')
    conv4 = mx.sym.Convolution(data=pool3, kernel=(2, 2), num_filter=128, name='conv4')
    prelu4 = mx.sym.LeakyReLU(data=conv4, act_type='prelu', name='prelu4')
    fc = mx.sym.FullyConnected(data=prelu4, num_hidden=256, name='fc')
    prelu5 = mx.sym.LeakyReLU(data=fc, act_type='prelu', name='prelu5')
    score = mx.sym.FullyConnected(data=prelu5, num_hidden=2, name='score')
    prob = mx.sym.SoftmaxActivation(data=score, name='prob')
    bbox_pred = mx.sym.FullyConnected(data=prelu5, num_hidden=4, name='bbox_pred')
    landmark_pred = mx.sym.FullyConnected(data=prelu5, num_hidden=10, name='landmark_pred')
    out = mx.sym.Group([prob, bbox_pred, landmark_pred])
    return out


def lnet():
    data = mx.sym.Variable('data')
    conv1 = mx.sym.Convolution(data=data, kernel=(3, 3), num_filter=50, num_group=5, name='conv1')
    prelu1 = mx.sym.LeakyReLU(data=conv1, act_type='prelu', name='prelu1')
    pool1 = mx.sym.Pooling(data=prelu1, kernel=(3, 3), stride=(2, 2), pool_type='max', \
                           pooling_convention='full', name='pool1')
    conv2 = mx.sym.Convolution(data=pool1, kernel=(3, 3), num_filter=100, num_group=5, name='conv2')
    prelu2 = mx.sym.LeakyReLU(data=conv2, act_type='prelu', name='prelu2')
    pool2 = mx.sym.Pooling(data=prelu2, kernel=(3, 3), stride=(2, 2), pool_type='max', \
                           pooling_convention='full', name='pool2')
    conv3 = mx.sym.Convolution(data=pool2, kernel=(2, 2), num_filter=200, num_group=5, name='conv3')
    prelu3 = mx.sym.LeakyReLU(data=conv3, act_type='prelu', name='prelu3')
    conv4 = mx.sym.Convolution(data=prelu3, kernel=(3, 3), stride=(3, 3), num_filter=100, num_group=5, name='conv4')
    prelu4 = mx.sym.LeakyReLU(data=conv4, act_type='prelu', name='prelu4')
    conv5 = mx.sym.Convolution(data=prelu4, kernel=(1, 1), num_filter=50, num_group=5, name='conv5')
    prelu5 = mx.sym.LeakyReLU(data=conv5, act_type='prelu', name='prelu5')
    conv6 = mx.sym.Convolution(data=prelu5, kernel=(1, 1), num_filter=10, num_group=5, name='conv6')
    out = mx.sym.Reshape(conv6, shape=(-1, 10))
    return out

if __name__ == '__main__':
    p = pnet()
    r = rnet()
    o = onet()
    l = lnet()
    mx.viz.plot_network(p, shape={'data': (1, 3, 12, 12)}).render('tmp/pnet')
    mx.viz.plot_network(r, shape={'data': (1, 3, 24, 24)}).render('tmp/rnet')
    mx.viz.plot_network(o, shape={'data': (1, 3, 48, 48)}).render('tmp/onet')
    mx.viz.plot_network(l, shape={'data': (1, 15, 24, 24)}).render('tmp/lnet')