search

chongyangwang-song / multi-person_parallel_behavior_recognition

online multi-person parallel behavior recognition
15 stars 0 forks source link

About the training of the TSM online model #1

Open zjh908491372 opened 2 years ago

zjh908491372 commented 2 years ago

I would like to know how the online model of TSM is trained, thanks. Is the online model trained with offline training? If yes, how does the input and output match?

chongyangwang-song commented 2 years ago

是用offline方式训练的，match确实是要做一些适配，把buffer的尺寸弄对就好了

---Original--- From: @.> Date: Tue, Aug 9, 2022 19:32 PM To: @.>; Cc: @.***>; Subject: [chongyangwang-song/multi-person_parallel_behavior_recognition] About the training of the TSM online model (Issue #1)

I would like to know how the online model of TSM is trained, thanks. Is the online model trained with offline training? If yes, how does the input and output match?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>

zjh908491372 commented 2 years ago

那它训练时候是按demo里的一帧一帧输入，还是直接输入多帧啊

chongyangwang-song commented 2 years ago

一帧一帧输入，保存前面帧的特征信息

---Original--- From: @.> Date: Wed, Aug 10, 2022 09:14 AM To: @.>; Cc: @.**@.>; Subject: Re: [chongyangwang-song/multi-person_parallel_behavior_recognition] About the training of the TSM online model (Issue #1)

那它训练时候是按demo里的一帧一帧输入，还是直接输入多帧啊

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

zjh908491372 commented 2 years ago

batch为1，segments为8的情况下，训练时按离线方式读取数据集，然后将8帧依次输入，最后跟离线方式一样，得出来8个output求平均计算损失吗？

chongyangwang-song commented 2 years ago

训练时是要离线训练，直接输入视频段训练。在线测试时每次输入一帧，要保存前面的帧信息，方便与当前帧做shift操作（详见官方的在线demo）

---Original--- From: @.> Date: Wed, Aug 10, 2022 09:38 AM To: @.>; Cc: @.**@.>; Subject: Re: [chongyangwang-song/multi-person_parallel_behavior_recognition] About the training of the TSM online model (Issue #1)

batch为1，segments为8的情况下，训练时按离线方式读取数据集，然后将8帧依次输入，最后跟离线方式一样，得出来8个output求平均计算损失吗？

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

zjh908491372 commented 2 years ago

现在只有一个问题了，就是训练和在线测试的模型都是同一个模型吧？还是说训练时用离线模型进行训练，然后在线模型读取离线模型的参数进行在线测试？

chongyangwang-song commented 2 years ago

用离线的方式训练。在线测试时用离线训练好的模型做测试，只不过输入方式与离线不一样了。在线测试是单帧输入

---Original--- From: @.> Date: Wed, Aug 10, 2022 10:17 AM To: @.>; Cc: @.**@.>; Subject: Re: [chongyangwang-song/multi-person_parallel_behavior_recognition] About the training of the TSM online model (Issue #1)

现在只有一个问题了，就是训练和在线测试的模型都是同一个模型吧？还是说训练时用离线模型进行训练，然后在线模型读取离线模型的参数进行在线测试？

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

zjh908491372 commented 2 years ago

也就是说训练和在线测试是同一个模型，只不过输入方式不同

chongyangwang-song commented 2 years ago

是

---Original--- From: @.> Date: Wed, Aug 10, 2022 11:15 AM To: @.>; Cc: @.**@.>; Subject: Re: [chongyangwang-song/multi-person_parallel_behavior_recognition]About the training of the TSM online model (Issue #1)

也就是说训练和在线测试是同一个模型，只不过输入方式不同

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

zjh908491372 commented 2 years ago

非常感谢