[Question] how to set Focal loss Parameters

[vkherde]

Hi @mvoelk

I am trying to use focal loss in SSD model, as per the code in DSOD_train.ipynb, lines

• loss = SSDFocalLoss(lambda_conf=10000, class_weights=class_weights)
• class_weights = np.array([0.00007205, 1.3919328 , 1.43665262, 1.30902077, 1.36668928, 1.2391509 , 1.21337629, 0.41527107, 1.1458096 , 0.29150119, 1.25713104, 0.61941517, 1.03175604, 1.21542005, 1.01947561, 0.0542007 , 1.12664538, 1.14966073, 1.12464889, 1.49998021, 1.09218961])

I want to understand how these parameters were calculated for "lamba_conf" and "class_weights" please confirm if you got a chance. Many thanks indeed

Best Regards
Vaibhav

[mvoelk]

I did not spend much time on the DSOD model.

The value for lambda is probably much too high. I took this value only for some experiments. A value of 100.0, as I used it for the SegLink models, should work find. I will also change that in the code.

The class_weight values were calculated using compute_class_weights method in the PriorUtility.

ssd_detectors/ssd_utils.py

Lines 552 to 561 in 3ef2c31

	def compute_class_weights(self, gt_util, num_samples=np.inf):
	"""Computes weighting factors for the classification loss by considering
	the inverse frequency of class instance in local ground truth.
	"""
	s = np.zeros(gt_util.num_classes)
	for i in tqdm(range(min(gt_util.num_samples, num_samples))):
	egt = self.encode(gt_util.data[i])
	s += np.sum(egt[:,-gt_util.num_classes:], axis=0)
	si = 1/s
	return si/np.sum(si) * len(s)

Roughly speaking, they are the normalized inverse of the class frequencies in the local ground truth.

[vkherde]

Thanks for the explanation @mvoelk

[mvoelk]

Just in case someone reads it... #18 may also be relevant.