We read every piece of feedback, and take your input very seriously.
To see all available qualifiers, see our documentation.
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
Many DETR models use group normalization within channel mapper (e.g. DINO, DETA, etc.).
The implementation of channel mapper in this repo defaults bias = True, which is not best practice since group normalization has its own bias term.
bias = True
In contrast, the channel mapper implementation in MMDetection relies on mmcv.ConvModule, which turns bias to False if a norm layer is attached (https://mmcv.readthedocs.io/en/2.x/api/generated/mmcv.cnn.ConvModule.html).
mmcv.ConvModule
Suggest either turning bias to False by default, or updating the model configs that use ChannelMapper with group norm to have bias=False
ChannelMapper
group norm
bias=False
In practice I get ~0.3 - 0.6 AP improvement from this
The text was updated successfully, but these errors were encountered:
No branches or pull requests
Many DETR models use group normalization within channel mapper (e.g. DINO, DETA, etc.).
The implementation of channel mapper in this repo defaults
bias = True
, which is not best practice since group normalization has its own bias term.In contrast, the channel mapper implementation in MMDetection relies on
mmcv.ConvModule
, which turns bias to False if a norm layer is attached (https://mmcv.readthedocs.io/en/2.x/api/generated/mmcv.cnn.ConvModule.html).Suggest either turning bias to False by default, or updating the model configs that use
ChannelMapper
withgroup norm
to havebias=False
In practice I get ~0.3 - 0.6 AP improvement from this
The text was updated successfully, but these errors were encountered: