search

rwightman / efficientdet-pytorch

A PyTorch impl of EfficientDet faithful to the original Google impl w/ ported weights
Apache License 2.0
1.58k stars 293 forks source link

[FEATURE] Set max detection points, dets per image, and data annotation instances per image constants in config #120

Closed MichaelMonashev closed 3 years ago

MichaelMonashev commented 3 years ago

How to set number of detections?

MichaelMonashev commented 3 years ago

I have to change MAX_DETECTIONS_PER_IMAGE in the efficientdet-pytorch code to predict all objects. How to increase this limit without your code patching?

rwightman commented 3 years ago

There are module level constants defined for those limits right now, they are currently set at comfortable values for COCO, VOC, OI, etc

They can be changed on the anchor side and dataset side https://github.com/rwightman/efficientdet-pytorch/blob/master/effdet/anchors.py#L46 https://github.com/rwightman/efficientdet-pytorch/blob/master/effdet/data/loader.py#L12

Low priority TODO to move them into config.

Ekta246 commented 3 years ago

Does this means if I have only 1 class, 1 object per image i.e, 1 bounding box, I should substitute the MAX_NUM_INSTANCES=1 in the dataloader?

If I do not do so will it hurt my training/validation accuracy?

MichaelMonashev commented 3 years ago

Does this means if I have only 1 class, 1 object per image i.e, 1 bounding box, I should substitute the MAX_NUM_INSTANCES=1 in the dataloader?

No.

Ekta246 commented 3 years ago

The changing of the dets_per image and max_num_instances varies the validation loss drastically. I am still finding the reason of the overfitting and validation loss divergence.

It would be of huge help is someone can throw light on this

MichaelMonashev commented 3 years ago

Thank you, @rwightman .