I couldn't understand this piece of code.

[ramsane]

        # 3. inverting the pred box equations
        grid_size = tf.shape(y_true)[1]
        grid = tf.meshgrid(tf.range(grid_size), tf.range(grid_size))
        grid = tf.expand_dims(tf.stack(grid, axis=-1), axis=2)
        true_xy = true_xy * tf.cast(grid_size, tf.float32) - \
            tf.cast(grid, tf.float32)
        true_wh = tf.math.log(true_wh / anchors)
        true_wh = tf.where(tf.math.is_inf(true_wh),
                           tf.zeros_like(true_wh), true_wh)

It is in the YoloLoss function. https://github.com/zzh8829/yolov3-tf2/blob/master/yolov3_tf2/models.py#L259

Can someone explain it to me what's happening?

I can understand what meshgrid and expand_dim are doing. We are creating 2D coordinates like [(0,1),(0,2),(0,3),.....(3,3)]. But I couldn't understand how and why we are calculating true_wh and true_xy.

[zzh8829]

input xy contains the absolute coordinate of bbox scaled to [0, 1] this operation will translate it to xy coordinate of bbox relative to the grid it resides in

true_wh calculates the log size of the bbox. since log(0) = inf, we use tf.where to set inf back to zero

we use relative xy and log wh because the paper author determined this strategy had the best performance