基于keras_Realtime_Multi-Person_Pose_Estimation进行性能分析

keras_Realtime_Multi-Person_Pose_Estimation安装

Realtime_Multi-Person_Pose_Estimation有许多版本，包括keras、tensorflow、pytorch等，几乎涵盖了所有深度学习框架，在这里由于进行性能分析，想用到keras中model.summary()的函数，所以采用了keras版本。
安装过程根据官方安装过程，甚至可以参考原版代码的安装流程，具体自行搜索。

在配置过程中，由于国内网络原因，许多模型无法流畅下载，提供两个caffemodel以及转换成keras格式的model.h5
相应文件格式如下：
/keras_Realtime_Multi-Person_Pose_Estimation/model/caffe/_trained_COCO :pose_iter_440000.caffemodel
/keras_Realtime_Multi-Person_Pose_Estimation/model/caffe/_trained_MPI :pose_iter_146000.caffemodel
/keras_Realtime_Multi-Person_Pose_Estimation/model/keras :model.h5
链接：https://pan.baidu.com/s/1nINVrBmubnqd1ye332YYrA
提取码：kfs0

对Realtime_Multi-Person_Pose_Estimation使用model.summary()

由于主要想看到模型的每一层操作以及相应操作数，采用model.summary()进行，配置完成后,在demo.ipynb中配置代码，定义model后添加，model.summary()

weights_path = "model/keras/model.h5" # orginal weights converted from caffe
#weights_path = "training/weights.best.h5" # weights tarined from scratch 

input_shape = (None,None,3)

img_input = Input(shape=input_shape)

stages = 6
np_branch1 = 38
np_branch2 = 19

img_normalized = Lambda(lambda x: x / 256 - 0.5)(img_input)  # [-0.5, 0.5]

# VGG
stage0_out = vgg_block(img_normalized)

# stage 1
stage1_branch1_out = stage1_block(stage0_out, np_branch1, 1)
stage1_branch2_out = stage1_block(stage0_out, np_branch2, 2)
x = Concatenate()([stage1_branch1_out, stage1_branch2_out, stage0_out])

# stage t >= 2
for sn in range(2, stages + 1):
    stageT_branch1_out = stageT_block(x, np_branch1, sn, 1)
    stageT_branch2_out = stageT_block(x, np_branch2, sn, 2)
    if (sn < stages):
        x = Concatenate()([stageT_branch1_out, stageT_branch2_out, stage0_out])

model = Model(img_input, [stageT_branch1_out, stageT_branch2_out])
model.load_weights(weights_path)

model.summary()

运行后即可输出相应结果。