High Confidence Loss on first epoch

[mlpotter]

Is a high confidence loss for a new dataset to be expected on the first epoch?

[glenn-jocher]

@mlpotter this high conf loss is created on purpose, as the conf loss is multiplied by 64 in the compute_loss() function. This was done as the result of a hyperparameter search from a few months ago #2 (comment), though this was done on the incorrect mAP metric used back then, so it should probably be redone.

cls loss is also divided by 4, so the natural conf and cls losses at epoch 0 batch 0 are really about 2 and 12 rather than 139 and 3.29. You can run your own evaluation of these loss component weights, or reset them to 1 (which would align more closely with the darknet loss function as I understand it), but in our tests we saw much worse results this way.

yolov3/utils/utils.py

Lines 264 to 277 in 09b02d2

	# Compute losses
	k = 1 # nT / bs
	if len(b) > 0:
	pi = pi0[b, a, gj, gi] # predictions closest to anchors
	tconf[b, a, gj, gi] = 1 # conf
	lxy += k * MSE(torch.sigmoid(pi[..., 0:2]), txy[i]) # xy loss
	lwh += k * MSE(pi[..., 2:4], twh[i]) # wh loss
	lcls += (k / 4) * CE(pi[..., 5:], tcls[i]) # class_conf loss
	# pos_weight = FT([gp[i] / min(gp) * 4.])
	# BCE = nn.BCEWithLogitsLoss(pos_weight=pos_weight)
	lconf += (k * 64) * BCE(pi0[..., 4], tconf) # obj_conf loss
	loss = lxy + lwh + lconf + lcls