Adding explicit_padding to EfficientDet model

[qraleq]

Prerequisites

1. The entire URL of the file you are using

https://github.com/tensorflow/models/blob/master/research/object_detection/models/ssd_efficientnet_bifpn_feature_extractor.py

2. Describe the feature you request

Currently, in ssd_efficientnet_bifpn_feature_extractor.py

https://github.com/tensorflow/models/blob/83b2ad3bf9557ec684c80d5e1760887f847b6ab6/research/object_detection/models/ssd_efficientnet_bifpn_feature_extractor.py#L101-L102

it says that using the use_explicit_padding parameter is unsupported by EfficientNetBiFPN.

I would like to be able to use the explicit_padding option with the EfficientDet model. Not sure if this is a limitation of the EfficientDet model, or it just hasn't been implemented yet.

3. Additional context

In order to make the model invariant to the input image size, I usually use a combination of FlexibleAnchorGenerator and explicit_padding. I previously used this successfully in the TF1 models such as ssd_mobilenet_v2_fpn. Now I'm trying to achieve the same behavior using TF2 OD API and EfficientDet model, but I found that use_explicit_padding is not supported in the EfficientDet model.

I tried to implement this on my own, but when I add the padding layers to efficientnet_model.py similar to how it's already done in mobilenet_v2.py, I get errors when loading the weights from the ModelZoo checkpoint, and I'm not sure how to fix it. Additionally, the structure or the ssd_efficientnet_feature_extractor.py and bidirectional_feature_pyramid_generators.py is quite different from the corresponding files for ssd_mobilenet_v2_fpn, which makes it difficult to figure out where the padding has to be applied.

4. Are you willing to contribute it? (Yes or No)

Yes, but will need some guidance.

[qraleq]

@tombstone @pkulzc Can you please check this and provide some guidance, please?

[matejcrnac]

Has anyone managed to solve this issue?