About training and testing details about vkd

[angpo]

Hi,

You can find the supplementary materials on arXiv (precisely, at the end of the main paper).
Additionally, commands.txt contains the bash scripts we used to train the models. If you refer to the values of hyperparameters, Section 4.1 of the main paper proposes a brief discussion about Implementation details.

If you need any further information please feel free to write us.

A.

[bigmoking]

Okay, thank you so much. Angelo Porrello <[email protected]> 于2020年7月16日周四 下午10:09写道：

…

Hi, You can find the supplementary materials on arXiv <https://arxiv.org/abs/2007.04174> (precisely, at the end of the main paper). Additionally, commands.txt <https://github.com/aimagelab/VKD/blob/master/commands.txt> contains the bash scripts we used to train the models. If you refer to the values of hyperparameters, Section 4.1 of the main paper proposes a brief discussion about Implementation details. If you need any further information please feel free to write us. A. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ALQIOSLZWQ2LTEKMULXCITTR34C23ANCNFSM4O4IVVRA> .

[bigmoking]

Hi： I have the following questions ： One: multiple views is camera views？ Two；why teacher network gets good performance？In the second training stage, teacher network also trained? Three: what is self-distillation? My level is limited. I would appreciate any more interruptions. Yours Angelo Porrello <[email protected]> 于2020年7月16日周四 下午10:21写道：

…

Closed #1 <#1>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ALQIOSMU54IGQVAYDIATYBLR34EFBANCNFSM4O4IVVRA> .

[angpo]

Hi,

1. Yes. Each image in the input set comes from a different camera.
2. No. The teacher network is fixed during the second stage; we train only the student one.
3. In general terms, Knowledge distillation relies on transferring knowledge from one network to another. In the model-compression area, a compact neural network is trained to mimic the outputs of a larger (and slower) model. Instead, the idea behind self-distillation is to teach a student network comprising the same architecture of its teacher. If you wish to know more about self-distillation, you can find here a list of paper on the topic.

A.

[bigmoking]

okay! Thank you so much for your answers , which solves a lot of my confusion. Angelo Porrello <[email protected]> 于2020年7月19日周日 下午5:52写道：

…

Reopened #1 <#1>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ALQIOSJIEBWRDDB6TOUCURTR4K64ZANCNFSM4O4IVVRA> .