Someone explain the sess.run in yolo.py?

[doubleLLL3]

I don't really understand the [self.boxes, self.scores, self.classes] in self.generate().
My purpose is to use yolo4.pb to do some prediction just in tensorflow, but here [self.boxes, self.scores, self.classes] do some tricks in keras and I don't know how to do this exchange using yolo4.pb.

out_boxes, out_scores, out_classes = self.sess.run(
            [self.boxes, self.scores, self.classes],
            feed_dict={
                self.yolo_model.input: image_data,
                self.input_image_shape: [image.size[1], image.size[0]]
                # K.learning_phase(): 0
            })

Thanks for the coming help!

I use amir-abdi/keras_to_tensorflow convert the yolo4.h5 to yolo4.pb. The format of outputs are different compared to using yolo4.h5.

[doubleLLL3]

This is what I use to run the pb model, I confirm that the input and output name are correct.

pred_sbbox, pred_mbbox, pred_lbbox = sess.run(
            [return_tensors[1], return_tensors[2], return_tensors[3]],
            feed_dict={return_tensors[0]: image_data})

and one part of pred_sbbox (its shape is (1, 52, 52, 255), 255 is 3 * 85) is like this:

It's strange! The 4th column should be the score, but they are all negative, so do the right columns.
Why so many negetive value?
Thanks for the coming help!

One of the correct results I think are like this:

This is from yolo3, and pred_sbbox's shape is (1,52,52,3,85)