The offline train is same to online train or not ?

qixi666666 commented 3 years ago

the show of train only to offline train , is that mean online train is same to offline train?

dy-gdut commented 3 years ago

As for how to train my data set, I have the folder of video frame and the label of the corresponding folder. How can I make train_videofolder.txt and val_videofolder.txt

chongyangwang-song commented 3 years ago

@qixi666666 I trained online model using offline code, the outputs seem to be right

qixi666666 commented 3 years ago

@qixi666666 I trained online model using offline code, the outputs seem to be right

@chongyangwang-song Maybe this is little difference between offline and online which are trained with offline code, but I think the true way should be chance some code in "temporal_shift.py". For example , delete "out[:, 1:, fold: 2 fold] = x[:, :-1, fold: 2 fold] # shift right"

qixi666666 commented 3 years ago

As for how to train my data set, I have the folder of video frame and the label of the corresponding folder. How can I make train_videofolder.txt and val_videofolder.txt

i think some tool in tsn can do it

hanson-eye commented 3 years ago

@qixi666666 I trained online model using offline code, the outputs seem to be right

@chongyangwang-song hithere， we are doing this work right now,but we have trouble with train the online mobilenet mdel, would u mind sharing your code about training tsm onnline with us? appreciate!

NB-Xie commented 3 years ago

@qixi666666 I trained online model using offline code, the outputs seem to be right

@chongyangwang-song hithere， we are doing this work right now,but we have trouble with train the online mobilenet mdel, would u mind sharing your code about training tsm onnline with us? appreciate!

Hithere, did you solve this issue? same problem here

NB-Xie commented 3 years ago

@qixi666666 I trained online model using offline code, the outputs seem to be right

Hi there, may I know: did you use the inference model they provided in /online_demo , where the input size seems to be (1,3,224,224) (frame by frame) , or build your own inference model based on training code, where the inputs of (num_segment, 3, 224, 224) ?

thank you! @chongyangwang-song

chongyangwang-song commented 3 years ago

嗨，在线版本的意思就是逐帧推理。他的输入只能是（1，3,224,224），直接用online_demo即可。离线版本输入会是（num_segments,3,224,224），t时刻的帧信息可以通过shift融合t+1时刻的信息。对于在线版本，推理当前帧的时候是不知道未来的一帧的，所以t时刻的帧只能和t前面的帧通过shift融合，这是离线和在线的差异。详细内容可以看原论文。

------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2021年7月13日(星期二) 中午11:10 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

@qixi666666 I trained online model using offline code, the outputs seem to be right

Hi there, may I know: did you use the inference model they provided in /online_demo , where the input size seems to be (1,3,224,224) (frame by frame) , or build your own inference model based on training code, where the inputs of (num_segment, 3, 224, 224) ?

thank you! @chongyangwang-song

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

NB-Xie commented 3 years ago

谢谢您的回复！

请问能知道您这样训练的细节吗？我应该从offline版本训练，然后将训练出来的网络参数导入到online版本的判别器吗？这样的话请问有些参数比如num_segments这里在offline训练中怎么调整呢？

我按照这个训练会有这样的问题： https://github.com/mit-han-lab/temporal-shift-module/issues/39#issuecomment-820022829

------------------ 原始邮件 ------------------ 发件人: "mit-han-lab/temporal-shift-module" @.>; 发送时间: 2021年7月13日(星期二) 晚上8:29 @.>; @.**@.>; 主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

嗨，在线版本的意思就是逐帧推理。他的输入只能是（1，3,224,224），直接用online_demo即可。离线版本输入会是（num_segments,3,224,224），t时刻的帧信息可以通过shift融合t+1时刻的信息。对于在线版本，推理当前帧的时候是不知道未来的一帧的，所以t时刻的帧只能和t前面的帧通过shift融合，这是离线和在线的差异。详细内容可以看原论文。

------------------ 原始邮件 ------------------ 发件人: @.>;
发送时间: 2021年7月13日(星期二) 中午11:10 收件人: @.>;
抄送: @.>; @.>;
主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

@qixi666666 I trained online model using offline code, the outputs seem to be right

Hi there, may I know: did you use the inference model they provided in /online_demo , where the input size seems to be (1,3,224,224) (frame by frame) , or build your own inference model based on training code, where the inputs of (num_segment, 3, 224, 224) ?

thank you! @chongyangwang-song

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

chongyangwang-song commented 3 years ago

是的，先在offline训练，然后把参数导入到online版本，在offline训练时segment参数对精度的影响似乎不是很大，我设置的是8。我复现过demo中的手势识别，由于当时没搞懂tvm，我就把tvm加速去掉了，只用torch推理，识别结果是正确的。

------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2021年7月14日(星期三) 上午8:59 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

谢谢您的回复！

请问能知道您这样训练的细节吗？  我应该从offline版本训练，然后将训练出来的网络参数导入到online版本的判别器吗？这样的话请问有些参数比如num_segments这里在offline训练中怎么调整呢？

我按照这个训练会有这样的问题： https://github.com/mit-han-lab/temporal-shift-module/issues/39#issuecomment-820022829

------------------ 原始邮件 ------------------ 发件人: "mit-han-lab/temporal-shift-module" @.>; 发送时间: 2021年7月13日(星期二) 晚上8:29 @.>; @.**@.>; 主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

嗨，在线版本的意思就是逐帧推理。他的输入只能是（1，3,224,224），直接用online_demo即可。离线版本输入会是（num_segments,3,224,224），t时刻的帧信息可以通过shift融合t+1时刻的信息。
对于在线版本，推理当前帧的时候是不知道未来的一帧的，所以t时刻的帧只能和t前面的帧通过shift融合，这是离线和在线的差异。
详细内容可以看原论文。

------------------&nbsp;原始邮件&nbsp;------------------
发件人: @.&gt;;
发送时间: 2021年7月13日(星期二) 中午11:10
收件人: @.&gt;;
抄送: @.&gt;; @.&gt;;
主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

@qixi666666 I trained online model using offline code, the outputs seem to be right

Hi there, may I know:
did you use the inference model they provided in /online_demo , where the input size seems to be (1,3,224,224) (frame by frame) ,
or build your own inference model based on training code, where the inputs of (num_segment, 3, 224, 224) ?

thank you! @chongyangwang-song

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub, or unsubscribe.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

NB-Xie commented 3 years ago

谢谢！那请问需要把bi-direction换成uni-direction进行训练吗？

------------------ 原始邮件 ------------------ 发件人: "mit-han-lab/temporal-shift-module" @.>; 发送时间: 2021年7月14日(星期三) 上午9:12 @.>; @.**@.>; 主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

是的，先在offline训练，然后把参数导入到online版本，在offline训练时segment参数对精度的影响似乎不是很大，我设置的是8。我复现过demo中的手势识别，由于当时没搞懂tvm，我就把tvm加速去掉了，只用torch推理，识别结果是正确的。

------------------ 原始邮件 ------------------ 发件人: @.>;
发送时间: 2021年7月14日(星期三) 上午8:59 收件人: @.>;
抄送: @.>; @.>;
主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

谢谢您的回复！

请问能知道您这样训练的细节吗？&nbsp;
我应该从offline版本训练，然后将训练出来的网络参数导入到online版本的判别器吗？这样的话请问有些参数比如num_segments这里在offline训练中怎么调整呢？

我按照这个训练会有这样的问题：
https://github.com/mit-han-lab/temporal-shift-module/issues/39#issuecomment-820022829

------------------&nbsp;原始邮件&nbsp;------------------
发件人: "mit-han-lab/temporal-shift-module" @.&gt;;
发送时间:&nbsp;2021年7月13日(星期二) 晚上8:29
@.&gt;;
@.**@.&gt;;
主题:&nbsp;Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

嗨，在线版本的意思就是逐帧推理。他的输入只能是（1，3,224,224），直接用online_demo即可。离线版本输入会是（num_segments,3,224,224），t时刻的帧信息可以通过shift融合t+1时刻的信息。
对于在线版本，推理当前帧的时候是不知道未来的一帧的，所以t时刻的帧只能和t前面的帧通过shift融合，这是离线和在线的差异。
详细内容可以看原论文。

------------------&amp;nbsp;原始邮件&amp;nbsp;------------------
发件人: @.&amp;gt;;
发送时间: 2021年7月13日(星期二) 中午11:10
收件人: @.&amp;gt;;
抄送: @.&amp;gt;; @.&amp;gt;;
主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

@qixi666666 I trained online model using offline code, the outputs seem to be right

Hi there, may I know:
did you use the inference model they provided in /online_demo , where the input size seems to be (1,3,224,224) (frame by frame) ,
or build your own inference model based on training code, where the inputs of (num_segment, 3, 224, 224) ?

thank you! @chongyangwang-song

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub, or unsubscribe.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub, or unsubscribe.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

chongyangwang-song commented 3 years ago

bi-direction和uni-direction 是什么意思

------------------ 原始邮件 ------------------ 发件人: "mit-han-lab/temporal-shift-module" @.>; 发送时间: 2021年7月14日(星期三) 上午9:21 @.>; @.**@.>; 主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

谢谢！那请问需要把bi-direction换成uni-direction进行训练吗？

------------------ 原始邮件 ------------------ 发件人: "mit-han-lab/temporal-shift-module" @.>; 发送时间: 2021年7月14日(星期三) 上午9:12 @.>; @.**@.>; 主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

是的，先在offline训练，然后把参数导入到online版本，在offline训练时segment参数对精度的影响似乎不是很大，我设置的是8。我复现过demo中的手势识别，由于当时没搞懂tvm，我就把tvm加速去掉了，只用torch推理，识别结果是正确的。

------------------&nbsp;原始邮件&nbsp;------------------
发件人: @.&gt;;
发送时间: 2021年7月14日(星期三) 上午8:59
收件人: @.&gt;;
抄送: @.&gt;; @.&gt;;
主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

谢谢您的回复！

请问能知道您这样训练的细节吗？&amp;nbsp;
我应该从offline版本训练，然后将训练出来的网络参数导入到online版本的判别器吗？这样的话请问有些参数比如num_segments这里在offline训练中怎么调整呢？

我按照这个训练会有这样的问题：
https://github.com/mit-han-lab/temporal-shift-module/issues/39#issuecomment-820022829

------------------&amp;nbsp;原始邮件&amp;nbsp;------------------
发件人: "mit-han-lab/temporal-shift-module" @.&amp;gt;;
发送时间:&amp;nbsp;2021年7月13日(星期二) 晚上8:29
@.&amp;gt;;
@.**@.&amp;gt;;
主题:&amp;nbsp;Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

嗨，在线版本的意思就是逐帧推理。他的输入只能是（1，3,224,224），直接用online_demo即可。离线版本输入会是（num_segments,3,224,224），t时刻的帧信息可以通过shift融合t+1时刻的信息。    
对于在线版本，推理当前帧的时候是不知道未来的一帧的，所以t时刻的帧只能和t前面的帧通过shift融合，这是离线和在线的差异。    
详细内容可以看原论文。    

------------------&amp;amp;amp;amp;nbsp;原始邮件&amp;amp;amp;amp;nbsp;------------------    
发件人: ***@***.***&amp;amp;amp;amp;gt;;     
发送时间: 2021年7月13日(星期二) 中午11:10    
收件人: ***@***.***&amp;amp;amp;amp;gt;;     
抄送: ***@***.***&amp;amp;amp;amp;gt;; ***@***.***&amp;amp;amp;amp;gt;;     
主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)    

@qixi666666 I trained online model using offline code, the outputs seem to be right    

Hi there, may I know:    
 did you use the inference model they provided in /online_demo , where the input size seems to be (1,3,224,224) (frame by frame) ,    
 or build your  own inference model based on training code, where the inputs  of (num_segment, 3, 224, 224) ?    

thank you! @chongyangwang-song    

—    
You are receiving this because you were mentioned.    
Reply to this email directly, view it on GitHub, or unsubscribe.

—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub, or unsubscribe.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub, or unsubscribe.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

NB-Xie commented 3 years ago

哦哦，就是您说的“对于在线版本，推理当前帧的时候是不知道未来的一帧的，所以t时刻的帧只能和t前面的帧通过shift融合，这是离线和在线的差异。” 训练offline的时候要把这个双向（bi-direction）shift的代码改成单向shift（uni-direction）吗？

------------------ 原始邮件 ------------------ 发件人: "mit-han-lab/temporal-shift-module" @.>; 发送时间: 2021年7月14日(星期三) 上午9:23 @.>; @.**@.>; 主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

bi-direction和uni-direction 是什么意思

------------------ 原始邮件 ------------------ 发件人: "mit-han-lab/temporal-shift-module" @.>; 发送时间: 2021年7月14日(星期三) 上午9:21 @.>; @.**@.>; 主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

谢谢！
那请问需要把bi-direction换成uni-direction进行训练吗？

------------------&nbsp;原始邮件&nbsp;------------------
发件人: "mit-han-lab/temporal-shift-module" @.&gt;;
发送时间:&nbsp;2021年7月14日(星期三) 上午9:12
@.&gt;;
@.**@.&gt;;
主题:&nbsp;Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

是的，先在offline训练，然后把参数导入到online版本，在offline训练时segment参数对精度的影响似乎不是很大，我设置的是8。我复现过demo中的手势识别，由于当时没搞懂tvm，我就把tvm加速去掉了，只用torch推理，识别结果是正确的。

------------------&amp;nbsp;原始邮件&amp;nbsp;------------------
发件人: @.&amp;gt;;
发送时间: 2021年7月14日(星期三) 上午8:59
收件人: @.&amp;gt;;
抄送: @.&amp;gt;; @.&amp;gt;;
主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)

谢谢您的回复！    

请问能知道您这样训练的细节吗？&amp;amp;amp;amp;nbsp;    
我应该从offline版本训练，然后将训练出来的网络参数导入到online版本的判别器吗？这样的话请问有些参数比如num_segments这里在offline训练中怎么调整呢？    

我按照这个训练会有这样的问题：    
https://github.com/mit-han-lab/temporal-shift-module/issues/39#issuecomment-820022829    

------------------&amp;amp;amp;amp;nbsp;原始邮件&amp;amp;amp;amp;nbsp;------------------    
发件人:                                                                                                                        "mit-han-lab/temporal-shift-module"                                                                                    ***@***.***&amp;amp;amp;amp;gt;;    
发送时间:&amp;amp;amp;amp;nbsp;2021年7月13日(星期二) 晚上8:29    
***@***.***&amp;amp;amp;amp;gt;;    
***@***.******@***.***&amp;amp;amp;amp;gt;;    
主题:&amp;amp;amp;amp;nbsp;Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)    

 嗨，在线版本的意思就是逐帧推理。他的输入只能是（1，3,224,224），直接用online_demo即可。离线版本输入会是（num_segments,3,224,224），t时刻的帧信息可以通过shift融合t+1时刻的信息。     
 对于在线版本，推理当前帧的时候是不知道未来的一帧的，所以t时刻的帧只能和t前面的帧通过shift融合，这是离线和在线的差异。     
 详细内容可以看原论文。     

 ------------------&amp;amp;amp;amp;amp;nbsp;原始邮件&amp;amp;amp;amp;amp;nbsp;------------------     
 发件人: ***@***.***&amp;amp;amp;amp;amp;gt;;      
 发送时间: 2021年7月13日(星期二) 中午11:10     
 收件人: ***@***.***&amp;amp;amp;amp;amp;gt;;      
 抄送: ***@***.***&amp;amp;amp;amp;amp;gt;; ***@***.***&amp;amp;amp;amp;amp;gt;;      
 主题: Re: [mit-han-lab/temporal-shift-module] The offline train is same to online train or not ? (#190)     

 @qixi666666 I trained online model using offline code, the outputs seem to be right     

 Hi there, may I know:     
  did you use the inference model they provided in /online_demo , where the input size seems to be (1,3,224,224) (frame by frame) ,     
  or build your  own inference model based on training code, where the inputs  of (num_segment, 3, 224, 224) ?     

 thank you! @chongyangwang-song     

 —     
 You are receiving this because you were mentioned.     
 Reply to this email directly, view it on GitHub, or unsubscribe.     
—    
You are receiving this because you commented.    
Reply to this email directly, view it on GitHub, or unsubscribe.

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub, or unsubscribe.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub, or unsubscribe.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

Deepakgthomas commented 3 years ago

@qixi666666 I trained online model using offline code, the outputs seem to be right

@chongyangwang-song hithere， we are doing this work right now,but we have trouble with train the online mobilenet mdel, would u mind sharing your code about training tsm onnline with us? appreciate!

@hanson-eye I am struggling to understand the online code. I just need to understand how to move the feature maps efficiently in the online network. Can you please help me with that? Thank you.

mit-han-lab / temporal-shift-module

The offline train is same to online train or not ? #190