I use my own data, but got this problem :(
[*] Model Built
Traceback (most recent call last):
File "trainCPM.py", line 38, in
model.train()
File "../net/CPM.py", line 224, in train
_train_batch = next(self.generator)
File "../dataset/datagen_para_branch.py", line 59, in generator
gen = self.datagen.generator(*args, **kwargs)
File "../dataset/datagen.py", line 459, in generator
return self._aux_generator(*args, **kwargs)
TypeError: _aux_generator() got an unexpected keyword argument 'norm'

您好！
非常感谢你的分享！
我是一个刚入门的小白。在运行您的程序 predict.py 时报错了。
if name=='main':
img_names = ['test_pic.jpg']
j = predict(img_names, 'model/model.ckpt-99.data-00000-of-00001', debug=True, is_name=True)
print(j)

请问下，是我的参数设置的哪里有问题吗？

前辈，请问代码的TensorFlow的ckpt模型您有保存吗？我想用来预测看一下效果，跪求。

你好，请问可以自己定义关键点位置和做数据么？如果可以，能给些指导么？

hello, i find the url of trained model is invalid, can you re-upload an available link?

看了前辈的文章和代码，我受益匪浅，十分感谢！
我看您的code在训练中使用了pretrained_model='../model/mobilev1_1.0.npy'； pretrained_model="../model/vgg19.npy"这个两个pretrained模型，但是我不知道他们在哪，看了您分享的链接，没有这两个文件，希望前辈指教

在_computer_acc 中有一个除以91的操作，这个91 是怎么来的呐？

您好！感谢您的分享！
我对net文件夹里的文件的具体含义稍微有一些疑惑：

1. PPNnMask.py指代的是R-CNN吗？
2. MobileCPM的结构与原本的CPM的区别在哪里呢？
   由于我是刚入门计算机视觉的新手，提出的问题可能比较粗浅，望谅解！
   再次感谢您的分享！

Hello,
Thanks for your sharing! However, the site offered for downloading your model seems to be unaccessible currently. Is it Okay for you to put them on Google or BaiduYun Drive? Thanks a lot!

大神您好，想针对自己的问题做个尝试，有两个问题想请教您，关于MPII标注的数据结构，是否就是 picturename minbox.x minbox.y maxbox.x maxbox.y point1.x point1.y ........
另外，我自己尝试运行trainMobileCPM，发现缺少了‘import dataset.datagen_para_branch as paral’，添加后出现RuntimeError :
RuntimeError:
An attempt has been made to start a new process before the
current process has finished its bootstrapping phase.

    This probably means that you are not using fork to start your
    child processes and you have forgotten to use the proper idiom
    in the main module:

        if __name__ == '__main__':
            freeze_support()
            ...

    The "freeze_support()" line can be omitted if the program
    is not going to be frozen to produce an executable.

这里使用你写的valPCK 验证文件执行的结果，这个结果为什么这么高？是不是那里出现了问题了呢？

情况主要出现在对大像素图片进行测试时。

def _crop_data(self, height, width, box, joints, boxp=0.05):
""" Automatically returns a padding vector and a bounding box given
the size of the image and a list of joints.
Args:
height : Original Height
width : Original Width
box : Bounding Box
joints : Array of joints
boxp : Box percentage (Use 20% to get a good bounding box)
"""
padding = [[0, 0], [0, 0], [0, 0]]
j = np.copy(joints)
if box[0:2] == [-1, -1]:
j[joints == -1] = 1e5
box[0], box[1] = min(j[:, 0]), min(j[:, 1])
crop_box = [box[0] - int(boxp * (box[2] - box[0])), box[1] - int(boxp * (box[3] - box[1])),
box[2] + int(boxp * (box[2] - box[0])), box[3] + int(boxp * (box[3] - box[1]))]
if crop_box[0] < 0: crop_box[0] = 0
if crop_box[1] < 0: crop_box[1] = 0
if crop_box[2] > width - 1: crop_box[2] = width - 1
if crop_box[3] > height - 1: crop_box[3] = height - 1
new_h = int(crop_box[3] - crop_box[1])
new_w = int(crop_box[2] - crop_box[0])
crop_box = [crop_box[0] + new_w // 2, crop_box[1] + new_h // 2, new_w, new_h]
if new_h > new_w:
bounds = (crop_box[0] - new_h // 2, crop_box[0] + new_h // 2)
if bounds[0] < 0:
padding[1][0] = abs(bounds[0])
if bounds[1] > width - 1:
padding[1][1] = abs(width - bounds[1])
elif new_h < new_w:
bounds = (crop_box[1] - new_w // 2, crop_box[1] + new_w // 2)
if bounds[0] < 0:
padding[0][0] = abs(bounds[0])
if bounds[1] > width - 1:
padding[0][1] = abs(height - bounds[1])
crop_box[0] += padding[1][0]
crop_box[1] += padding[0][0]
return padding, crop_box

你这里是 操作的padding[1]，padding[0], 最后使用了 np.pad(padding),但是 np.pad 中的pad 参数 第一个是用来填充的数字，padding[1], padding[2]才是用来存储应该pad的最小最大扩展长度

还有能否告知一下前辈运行的具体环境

请问在哪里提取mpii的人体bbox和kps坐标？
mpii测试集没有人体bbox是吗？

As you said in the comment in net/CPM.py line 262, my training crashed exactly when it tries to save the ckpt file. But this is not the problem of the TF 1.10.

If you change the name of the local variable 'n' in line 238 to something else (e.g. '_n'), it works perfectly fine even with TF 1.10. Note that another variable with the same name has already been used in the line 221

你好，我在用你的网络训练模型时，测试自己的图片出现了问题，不知道为啥我现在传不上去图片，已经将相关问题发你[email protected]邮箱了

希望得到你的回复

@mpskex
datagen_mpii.py的304行：if bounds[1] > width - 1:
其实应该是 if bounds[1] > height - 1:

不知道我的理解是否有误？

大佬可以详细说一下怎么训练mpii数据集么，下的数据集是.mat格式，就是可以详细的说一下怎么跑么