zhengchang467 / strpm Goto Github PK

作者您好，我对您提出的STRPM模型十分感兴趣，您是否能公开或者介绍一下训练数据集的处理方法和训练过程。万分感谢

Dear authors @ZhengChang467 :
Many thanks for your excellent work. But, there is some unclear meaning in STRPM.py. What is the sense of c_att, m_att, c_net, and m_net？ Which formulas do these definitions correspond to in the paper?

I am looking forward to your reply.
Thanks

Dear authors of STRPM，

Thank you so much for sharing this fantastic work with the community! I know you are going to release the training code soon, but I am wondering if it is possible to get some details in advance about how to train STRPM on UCF sports and match the performances in the paper. Especially,

• The train / test split of the ucf sports dataset;
• The training hyperparameters that you think may be matter for the final performance, such as learning rate, batch size, max epoch/iteration number, etc. (I am asking for this information is because that I have tried the default learning rate & decay & iteration number with both batch sizes of 4 and 8 on my own ucf split, but fail to reproduce the similar performance.)

Any information would be highly appreciated. Thank you so much!

可能我对代码理解不太够，希望作者能分享下这个时空特征到底是怎么提取的呢，如果只是在图像上卷积不是只提取了空间特征吗

Dear authors of STRPM，

I'm trying to train my own dataset with your code.

From you code:
print('Loading train dataset') self.path = data_train_path with codecs.open(self.path) as f: self.file_list = f.readlines() print('Loading train dataset finished, with size:', len(self.file_list))
I think the content of train file should like:
data/ucfsport/Kicking-Side/,6
data/ucfsport/Kicking-Side/,7
data/ucfsport/Kicking-Side/,8
data/ucfsport/Kicking-Side/,9

the first line means using 6-9th to predict 10th
the second line means using 7-10th to predict 11th
and so on.

Am I right?
I appreciate your reply！

Thanks a lot!
Chu

First, thank you for your code.
Could you share the whole processed(decoded) test set or how to pre-process the videos and train-test split file ?
Especially for the SJTU4K and UCF

你好，我在看代码的时候发现STRPM_cell的实际实现增加了遗忘门用于处理attn，是类似ST-LSTM？同时想咨询一下cell中为何在最后进行residual操作的时候需要进行判断呢？

您好，感谢您的分享，我对代码有几处疑问（英文怕解释不清）

1. mask_true的作用
   在strpm.py的forward函数中，下面这段代码使用到了mask_true
   if t < self.configs.input_length:
   net = frames[:, t] #t=0 1,192,64,64 取某一幅图画
   else:
   time_diff = t - self.configs.input_length #input_length = 4 , 表示预测周期，每4张预测1张
   net = mask_true[:, time_diff] * frames[:, t] + (1 - mask_true[:, time_diff]) * x_gen
   调试后发现train时，mask_true都是1，test时mask_true都是0：
   问题1：是不是可以理解为，训练时，都输入原始图片，而测试时，输入Input_length长度的图片以后，只使用预测出的图片来循环预测？
   问题2：但既然是那样的话，在训练代码的loss计算时，为何用了所有的9张图输出，不否应该把t=0,1,2,3时的输出剔除掉来计算loss更好？

2. tau的设置
   tau的设置是不是与input_length有关?
   我的场景是用8张图来预测后面2张图，input_length=8,total_length=10, tau是不是应该设为9为好? 这块没有特别理解.

谢谢！盼复！
Qc
2020.6.20