questions on custom_loss

[SaisPas]

In frontend.py, could anyone explain a bit for this two line in function custom_loss()?

        cell_y = tf.transpose(cell_x, (0,2,1,3,4))
        cell_grid = tf.tile(tf.concat([cell_x,cell_y], -1), [self.batch_size, 1, 1, self.nb_box, 1])

why to transpose and why to concat this two?

[Jenia Golbstein]

I think this is a mistake, since when the input image isn't square (W != H) then this code crashes.
I changed these lines to:

    cell_x = tf.to_float(tf.reshape(tf.tile(tf.range(GRID_W), [GRID_H]), (1, GRID_H, GRID_W, 1, 1)))
    cell_y = tf.to_float(tf.reshape(tf.tile(tf.range(GRID_W), [GRID_H]), (1, GRID_H, GRID_W, 1, 1)))
    cell_grid = tf.tile(tf.concat([cell_x,cell_y], -1), [BATCH_SIZE, 1, 1, 5, 1])

and it worked.
Moreover, make sure to fix the cv2.resize inputs in "aug_image" method in "BatchGenerator" class.
should be
image = cv2.resize(image, (self.config['IMAGE_W'], self.config['IMAGE_H']))
instead of
image = cv2.resize(image, (self.config['IMAGE_H'], self.config['IMAGE_W']))

I've opened a PR for that

[Kien PM]

Dear @Golbstein ,
As I understood, according to original code cell_x and cell_y are the different (cell_y is transposed of cell_x)

cell_x = tf.to_float(tf.reshape(tf.tile(tf.range(self.grid_w), [self.grid_h]), (1, self.grid_h, self.grid_w, 1, 1)))
cell_y = tf.transpose(cell_x, (0,2,1,3,4))

But according to your code, cell_x and cell_y are the same

    cell_x = tf.to_float(tf.reshape(tf.tile(tf.range(GRID_W), [GRID_H]), (1, GRID_H, GRID_W, 1, 1)))
    cell_y = tf.to_float(tf.reshape(tf.tile(tf.range(GRID_W), [GRID_H]), (1, GRID_H, GRID_W, 1, 1)))
    cell_grid = tf.tile(tf.concat([cell_x,cell_y], -1), [BATCH_SIZE, 1, 1, 5, 1])

Is that right? And does it have any side effect?
Thank you so much!

[Ebenezer Scrooge]

Same issue, looking forward to further solution.