def loss_fn(self, out, annot):
tar_vector = Losses.get_tar_vector(annot)
loss_loc = Losses.get_loc_error(out, tar_vector)
loss_wh = Losses.get_w_h_error(out, tar_vector)
loss_conf = Losses.get_confidence_error(out, tar_vector)
loss_cls = Losses.get_class_error(out, tar_vector)
return loss_loc, loss_wh, loss_conf, loss_cls
for x in range(cell_num):
some_predictions_class_prob_notexist[0, y, x, :] = torch.ones((30))
unittest.TestCase().assertAlmostEqual(
0.5 (noobject_coef * 20 * cell_num * cell_num),
Losses.get_conditional_class_prob_notexist(
some_predictions_class_prob_notexist,
some_targets_class_prob_notexist).cpu().detach().numpy()[0], 2)
#0.5(noobject_coef*20objects*7*7) == 490
target_for_wh = {
(3, 3):
[[0, [torch.Tensor([0.36]).cuda(),
torch.Tensor([0.64]).cuda()], [0, 0]]]
}
predictions_wh = torch.zeros((1, cell_num, cell_num, 30))
for y in range(cell_num):
for x in range(cell_num):
if y == 3 and x == 3:
predictions_wh[0, y, x, :] = torch.Tensor([
0, 0, 0.04, 0.09, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
])
else:
predictions_wh[0, y, x, :] = torch.zeros((30))
#0.5*((0.6-0.2)^2 + (0.8-0.3)^2 ) = 0.5(0.16+0.25) = 0.5*0.41 = 0.205
unittest.TestCase().assertAlmostEqual(
0.205,
Losses.get_w_h_error(predictions_wh,
target_for_wh).cpu().detach().numpy()[0], 3)