about loss

[HonestyBrave]

hello, tztztztztz:
thankyou for your opensource code, it's very good!
but i have a question about loss, i read some blog about the loss, the yolov2 loss is like the picture, two red box in the picture is not find in your loss.py, is i didn't find or you do not write?

[tztztztztz]

Hi,

Thank you for using this repo!
In fact, the two losses in the red boxes ARE INCLUDED in the loss function.

yolov2.pytorch/loss.py

Lines 218 to 219 in 014559c

	box_loss = 1 / b * cfg.coord_scale * F.mse_loss(delta_pred_batch * box_mask, box_target * box_mask, reduction='sum') / 2.0
	iou_loss = 1 / b * F.mse_loss(conf_pred_batch * iou_mask, iou_target * iou_mask, reduction='sum') / 2.0

For box losses, the losses for both obj_coord and noobj_coord are included in a single tensor. Since the loss for noobj_coord is actually 0, I use a box_mask to mask out losses of noobj_coord.

For conf losses (iou_loss), the loss weights for conf_obj and conf_noobj are set by iou_mask

yolov2.pytorch/loss.py

Line 61 in 014559c

iou_mask = delta_pred_batch.new_ones((bsize, H * W, num_anchors, 1)) * cfg.noobject_scale

yolov2.pytorch/loss.py

Line 159 in 014559c

iou_mask[b, cell_idx, argmax_anchor_idx, :] = cfg.object_scale