Unable to use checkpoints for prediction/detection

Hi all,

I trained YOLO and FASTER RCNN detectors on face image datasets with 100 epochs. During the training, checkpoints have been saved. After the training I tried to use a checkpoint as a network for prediction. As a result MATLAB behaved in a different way depending on the trained algorithm.

1) for FASTER I was able to use a checkpoint as a network for bounding box detection

2) for YOLO I was not able to make a detection and received the following error:

Unable to use a detector with batch normalization layers before training is complete. Use trainYOLOv2ObjectDetector to train the detector for one epoch to complete training.

Both models use batch normalization layers, but FASTER RCNN has this layer locked for training by default. My questions are following:

q1) how to checkpoints from YOLO to perform detection?

q2) are the checkpoints from FASTER RCNN a proper networks for detection?

regards,

Marcin

Answer 1

[Srivardhan Gadila]

YOLOv2 detector having batch normalization layers cannot be used for detection when loaded from a saved checkpoint. This is a known issue and the concerned parties might be working on it. As a workaround, train the detector loaded from the checkpoint with minimum learning rate for atleast 1 epoch, then the detector obtained can be used for Object detection.

The checkpoints saved during training of the fasterRCNN using the trainFasterRCNNObjectDetector can be used for detection just like a completely trained fasterRCNN detector.