Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Question over atrous rate for mobilenet #5919

Closed
CrazySherman opened this issue Dec 17, 2018 · 5 comments
Labels

Comments

@CrazySherman
Copy link

@CrazySherman CrazySherman commented Dec 17, 2018

# When using 'mobilent_v2', we set atrous_rates = decoder_output_stride = None.

I tried to use some atrous rate for mobilenet_V2 and the result was like disaster.
I was wondering whether you can provide some insights over why mobilenet base architecture cannot leverate atrous convolution -- I mean afterall it's applying a couple of atrous convolutions on feature maps right? why would that makes a huge different, as opposed to Xception net that ended up benefiting from atrous convolution?

@aquariusjay

This comment has been minimized.

Copy link
Contributor

@aquariusjay aquariusjay commented Dec 17, 2018

Hi CrazySherman,

ASPP should still work for MobileNet-V2 backbone.
We do not use it because we target at faster inference speed instead of high performance when using MobileNet-V2.

@CrazySherman

This comment has been minimized.

Copy link
Author

@CrazySherman CrazySherman commented Dec 17, 2018

thx for the explanation, I tried mobilenet_v2 backboned with dice loss on kaggle's airbus dataset https://www.kaggle.com/c/airbus-ship-detection/leaderboard and achieved pretty amazing results. I think this framework can extend to detect small/sparse objects very well, as opposed to mask-rcnn, which indeed performs horrible at least for this dataset

@aquariusjay

This comment has been minimized.

Copy link
Contributor

@aquariusjay aquariusjay commented Dec 17, 2018

Hi CrazySherman,

Cool! Thanks for letting us know the experimental results!
Good job!

@msymp msymp self-assigned this Dec 28, 2018
@msymp msymp added the type:docs label Dec 28, 2018
@msymp msymp closed this Dec 28, 2018
@karanchahal

This comment has been minimized.

Copy link
Contributor

@karanchahal karanchahal commented Feb 8, 2019

@CrazySherman what was your command to train a mobile net ? Especially all the atrous rates etc ?

@mei123hao

This comment has been minimized.

Copy link

@mei123hao mei123hao commented Nov 8, 2019

I also use mobilenetv2 as backbone. I set atrous_rates as [6,12,18],output stride is 16,and it report error like this:
ValueError: Total size of new array must be unchanged for concat_projection/weights lh_shape: [(1, 1, 512, 256)], rh_shape: [(1, 1, 1280, 256)],
but it can start training when i set atrous_rates as None.What is my problem?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
5 participants
You can’t perform that action at this time.