I noticed the input class probability is in the range of 1~100, so I divided it with y = 100. I was able to get some test result right, and some smaller by a factor of 100, and others within reasonable range. What could I have done wrong?

Yes it is yolo_filter_boxes. It didnât pass the unit test.

What I did:

step 1: element-wise multiply of âbox_confidenceâ and âbox_class_probsâ
step 2: use tf.math.argmax to get âbox_classesâ from âbox_scoresâ, axis = -1
use tf.math.reduce_max to get âbox_class_scoresâ from âbox_scoresâ, axis = -1
step 3: direct compare if âbox_class_scoresâ is smaller than âthresholdâ
(here I noticed data inside âbox_class_scoresâ is in range of 0-100, so I must have done something wrong before this step?)

I also have done something very similar to this but I get an error Cannot convert 0.5 to EagerTensor of dtype int64 If I divide the probabilities by 100 i get a shape of 324,