404
LorenzoDag opened this issue · comments
In README RexNet 1.3, 1.5 and 2.0 redirect to 404 error.
ReXNet-lite is currently released only for the x1.0 model. If you need those models, please let me know.
Ok, thank you.
Just a question. I want to try to use this pretrained model to perform facial exprespression recognition. To obtain this, once i've loaded the weigths i need to add dense layer and train the new model. Is that right?
…________________________________
Da: Greg Dongyoon Han ***@***.***>
Inviato: giovedì 29 luglio 2021 11:48
A: clovaai/rexnet ***@***.***>
Cc: D'AGOSTINO LORENZO ***@***.***>; Author ***@***.***>
Oggetto: Re: [clovaai/rexnet] 404 (#24)
ReXNet-lite is currently released only for the x1.0 model. If you need those models, please let me know.
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub<https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fclovaai%2Frexnet%2Fissues%2F24%23issuecomment-888973588&data=04%7C01%7Cs1085177%40studenti.univpm.it%7Ce594c089dbe648cf6cba08d952760d42%7C117b418dfb21416fa85f1e9ff725bf2c%7C0%7C0%7C637631489264963792%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=rIuPeCmg5%2FguQMIrSG%2BLR9aEdta%2FZCWE1jBuOx4kUCU%3D&reserved=0>, or unsubscribe<https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FALQ444JSCVF47FTXHBW57OLT2EPXTANCNFSM5BEN4VPQ&data=04%7C01%7Cs1085177%40studenti.univpm.it%7Ce594c089dbe648cf6cba08d952760d42%7C117b418dfb21416fa85f1e9ff725bf2c%7C0%7C0%7C637631489264973784%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=o5AFqkDLHjJ9f5oc%2BilsedEwyzahhZiGGk917I29VJA%3D&reserved=0>.
Thank you for using our backbones!
Pretrained backbones are generally used replacing the last dense layer with a new one initialized randomly, and training is then performed only with the replaced dense layer by freezing some or entire convolution layers. However, sometimes, adding a single or couple of dense layers after the final dense layer of the backbone (as you asked) achieves better results on particular downstream tasks after training. Please try both ways and pick a better one for your own task.
Ok. Thank you for helping me, and thank you for you work too.
Ottieni Outlook per Android<https://aka.ms/ghei36>
…________________________________
From: Greg Dongyoon Han ***@***.***>
Sent: Thursday, July 29, 2021 6:55:58 PM
To: clovaai/rexnet ***@***.***>
Cc: D'AGOSTINO LORENZO ***@***.***>; Author ***@***.***>
Subject: Re: [clovaai/rexnet] 404 (#24)
Thank you for using our backbones!
Pretrained backbones are generally used replacing the last dense layer with a new one. Then, training is done only for the replaced dense layer with freezing some or entire convolution layers. However, sometimes, adding a single or couple of dense layers after the final dense layer of the backbone (as you asked) achieves better results on particular downstream tasks after training. Please try both ways and pick a better one for your own task.
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub<https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fclovaai%2Frexnet%2Fissues%2F24%23issuecomment-889307621&data=04%7C01%7Cs1085177%40studenti.univpm.it%7C808fcdd7002342fcf6a008d952b1be09%7C117b418dfb21416fa85f1e9ff725bf2c%7C0%7C0%7C637631745620119658%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=IyD7Rq0p7cWgliCP%2BkQGC4cSZXEDglZ%2B8869lgJPdTs%3D&reserved=0>, or unsubscribe<https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FALQ444PZJNHFUGDTOB4GEKTT2GBZ5ANCNFSM5BEN4VPQ&data=04%7C01%7Cs1085177%40studenti.univpm.it%7C808fcdd7002342fcf6a008d952b1be09%7C117b418dfb21416fa85f1e9ff725bf2c%7C0%7C0%7C637631745620129654%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=n5BceVVbOSvBCOIhzZw%2Bi%2F1yRl4piJTxycVxYKXZY8E%3D&reserved=0>.
excuse me sir,
is your model trained with 3 channels images?
…________________________________
Da: Greg Dongyoon Han ***@***.***>
Inviato: giovedì 29 luglio 2021 18:55
A: clovaai/rexnet ***@***.***>
Cc: D'AGOSTINO LORENZO ***@***.***>; Author ***@***.***>
Oggetto: Re: [clovaai/rexnet] 404 (#24)
Thank you for using our backbones!
Pretrained backbones are generally used replacing the last dense layer with a new one. Then, training is done only for the replaced dense layer with freezing some or entire convolution layers. However, sometimes, adding a single or couple of dense layers after the final dense layer of the backbone (as you asked) achieves better results on particular downstream tasks after training. Please try both ways and pick a better one for your own task.
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub<https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fclovaai%2Frexnet%2Fissues%2F24%23issuecomment-889307621&data=04%7C01%7Cs1085177%40studenti.univpm.it%7C808fcdd7002342fcf6a008d952b1be09%7C117b418dfb21416fa85f1e9ff725bf2c%7C0%7C0%7C637631745620119658%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=IyD7Rq0p7cWgliCP%2BkQGC4cSZXEDglZ%2B8869lgJPdTs%3D&reserved=0>, or unsubscribe<https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FALQ444PZJNHFUGDTOB4GEKTT2GBZ5ANCNFSM5BEN4VPQ&data=04%7C01%7Cs1085177%40studenti.univpm.it%7C808fcdd7002342fcf6a008d952b1be09%7C117b418dfb21416fa85f1e9ff725bf2c%7C0%7C0%7C637631745620129654%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=n5BceVVbOSvBCOIhzZw%2Bi%2F1yRl4piJTxycVxYKXZY8E%3D&reserved=0>.
@LorenzoDag Yes, please refer to the code.