Skip to content

Instantly share code, notes, and snippets.

Embed
What would you like to do?
Multi-task Deep Learning Experiment using fastai Pytorch
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
@Light--

This comment has been minimized.

Copy link

@Light-- Light-- commented Jul 10, 2020

thanks for sharing....but, is that really a multi-task model? what's the architecture of the network really? Does it have a classical MTL structure which including a backbone and multi-head? or just only use resnet50.....sorry, could you explanin a little bit more...?

@icmpnorequest

This comment has been minimized.

Copy link

@icmpnorequest icmpnorequest commented Aug 10, 2020

thanks for sharing....but, is that really a multi-task model? what's the architecture of the network really? Does it have a classical MTL structure which including a backbone and multi-head? or just only use resnet50.....sorry, could you explanin a little bit more...?

It's a classical multi-task learning architecture, which uses hard parameters sharing method. For example, $loss_{total} = \lambda_{age} * loss_{age} + \lambda_{gender} * loss_{gender}$. Hope it may help you

@Light--

This comment has been minimized.

Copy link

@Light-- Light-- commented Aug 11, 2020

$loss_{total} = \lambda_{age} * loss_{age} + \lambda_{gender} * loss_{gender}$

@icmpnorequest
as far as i know, only calculating the loss together doesn't make the model to have a multi-task structure, you are doing the multi-objective learning without a multi-task model structure, right?
if the model structure is really using the classical multi-task structure as you said, could you tell me which paper/webpage you refer to?
thanks.

@icmpnorequest

This comment has been minimized.

Copy link

@icmpnorequest icmpnorequest commented Aug 12, 2020

$loss_{total} = \lambda_{age} * loss_{age} + \lambda_{gender} * loss_{gender}$

@icmpnorequest
as far as i know, only calculating the loss together doesn't make the model to have a multi-task structure, you are doing the multi-objective learning without a multi-task model structure, right?
if the model structure is really using the classical multi-task structure as you said, could you tell me which paper/webpage you refer to?
thanks.

Maybe you could refer to the paper [19 CVPR] Multimodal Age and Gender Classification Using Ear and Profile Face Images

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment