search

jfzhang95 / pytorch-deeplab-xception

DeepLab v3+ model in PyTorch. Support different backbones.
MIT License
2.9k stars 779 forks source link

Has someone tried efficientnet as backbone? #198

Open Tonyboy999 opened 3 years ago

Tonyboy999 commented 3 years ago

Efficientnet was considered as the best CNN in 2019, and many competitors got great rank with unet + efficientnet. I tried to use efficientnet as the backbone of this project but got poor results.

gfotedar commented 3 years ago

I'm facing similar issues. Its worse than resnet. Were you able to figure out why?

Tonyboy999 commented 3 years ago

Mine is not worse than resnet,it achieves similar effect as drn and a bit better than resnet. Just a little bit，but takes more time training.

发自我的iPhone

------------------ Original ------------------ From: Gaurav Fotedar @.> Date: Tue,Apr 20,2021 8:46 AM To: jfzhang95/pytorch-deeplab-xception @.> Cc: GuangzhouTony @.>, Author @.> Subject: Re: [jfzhang95/pytorch-deeplab-xception] Has someone tried efficientnet as backbone? (#198)

I'm facing similar issues. Its worse than resnet. Were you able to figure out why?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

gfotedar commented 3 years ago

Did you make any changes to efficientnet or its parameters ( dropout/drop_connect ) apart from removing the last layers?

Tonyboy999 commented 3 years ago

Almost not，I only tune the model structure by changing the low level feature layer in efficientnet. I simply extracted the feature map

发自我的iPhone

------------------ Original ------------------ From: Gaurav Fotedar @.> Date: Wed,Apr 21,2021 2:35 PM To: jfzhang95/pytorch-deeplab-xception @.> Cc: GuangzhouTony @.>, Author @.> Subject: Re: [jfzhang95/pytorch-deeplab-xception] Has someone tried efficientnet as backbone? (#198)

Did you make any changes to efficientnet or its parameters ( dropout/drop_connect ) apart from removing the last layers?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

Tonyboy999 commented 3 years ago

I think the effect of efficientnet may differ from different datasets.

发自我的iPhone

------------------ Original ------------------ From: Gaurav Fotedar @.> Date: Wed,Apr 21,2021 2:35 PM To: jfzhang95/pytorch-deeplab-xception @.> Cc: GuangzhouTony @.>, Author @.> Subject: Re: [jfzhang95/pytorch-deeplab-xception] Has someone tried efficientnet as backbone? (#198)

Did you make any changes to efficientnet or its parameters ( dropout/drop_connect ) apart from removing the last layers?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

Tonyboy999 commented 3 years ago

I think the effect of efficientnet may differ from different datasets

发自我的iPhone

------------------ Original ------------------ From: Gaurav Fotedar @.> Date: Wed,Apr 21,2021 2:35 PM To: jfzhang95/pytorch-deeplab-xception @.> Cc: GuangzhouTony @.>, Author @.> Subject: Re: [jfzhang95/pytorch-deeplab-xception] Has someone tried efficientnet as backbone? (#198)

Did you make any changes to efficientnet or its parameters ( dropout/drop_connect ) apart from removing the last layers?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.