Models: Object Detection API - Feature Request: RetinaNet with Focal Loss

[apacha]

I found some hints that indicate, that RetinaNet + Focal Loss might already be implemented in the Object Detection API, but I couldn't find any official documentation on how to use it in order to achieve the same that is done with this keras implementation: https://github.com/fizyr/keras-retinanet

Is it already implemented and ready to be used, but just undocumented? If so, it would be great to have this documented somewhere.

[xiaoyongzhu]

I am also interested in proposing this feature request. @apacha I agree that some hints suggest someone has been working on that for a while, especially there is a whole folder talking about RetinaNet under TPU folder (https://github.com/tensorflow/tpu/tree/master/models/official/retinanet) and looks like the TPU version is already running. I suspect the code can be run using regular GPU as well.

[austinmw]

Hi, I see that there's a support added message for RetinaNet from July 13th, but don't see the config file in the model zoo. Is it ran differently than the other models?

[daavoo]

@austinmw The config file is
https://github.com/tensorflow/models/blob/master/research/object_detection/samples/configs/ssd_resnet50_v1_fpn_shared_box_predictor_640x640_coco14_sync.config

And the name in the model zoo:
ssd_resnet_50_fpn_coco ☆

[pkulzc]

That's right, it now available in model zoo. Closing this request.

[tr-ips-bahubali]

@pkulzc is there any ipynb file which has a sample code to use the ssd_resnet_50_fpn_coco model?

[pkulzc]

All models in model zoo (except for tflite models) can be run in the notebook.

[tr-ips-bahubali]

Thanks !!

[zubairahmed-ai]

@austinmw The config file is
https://github.com/tensorflow/models/blob/master/research/object_detection/samples/configs/ssd_resnet50_v1_fpn_shared_box_predictor_640x640_coco14_sync.config

And the name in the model zoo:
ssd_resnet_50_fpn_coco ☆

So this ssd_resnet_50_fpn_coco is actually a RetinaNet model with ResNet base trained on coco dataset?
Why isn't it named like so?

[austinmw]

@zubairahmed-ai I believe because RetinaNet is just SSD with the addition of a FPN base network and a new loss function

[zubairahmed-ai]

@zubairahmed-ai I believe because RetinaNet is just SSD with a different loss function

Is that so, I dont know better, so I can just use this config instead? https://github.com/tensorflow/models/blob/master/research/object_detection/samples/configs/ssd_resnet50_v1_fpn_shared_box_predictor_640x640_coco14_sync.config

[austinmw]

@zubairahmed-ai Yeah

# SSD with Resnet 50 v1 FPN feature extractor, shared box predictor and focal
# loss (a.k.a Retinanet).

[zubairahmed-ai]

@zubairahmed-ai Yeah

# SSD with Resnet 50 v1 FPN feature extractor, shared box predictor and focal
# loss (a.k.a Retinanet).

Yes, yes I got it buddy, I will train my model with this now. Thanks a bunch

[nathantsoi]

Hi all, I tried to train this model (at b3158fb0183809400e9e7f8092dd541201b1c4d4) using the trainval35k coco split, fine tuning from http://download.tensorflow.org/models/resnet_v1_50_2016_08_28.tar.gz

However, it seems like there is a problem with classification. Can anyone else confirm this result?