How to initialize the category center

[tfboys-lzz]

How to initialize the category center

[TFboys-lzz]

Sorry I didn't see it until now.
For calculating category centers, you first need pretrain the model on the source dataset,
and then you can calculate the category center by

MPSCL/domain_adaptation/train_UDA.py

Lines 101 to 127 in bb58bde

	def update_class_center_iter(cla_src_feas,batch_src_labels,class_center_feas,m):
	'''
	batch_src_feas : n*c*h*w
	barch_src_labels: n*h*w
	'''
	batch_src_feas = cla_src_feas.detach()
	batch_src_labels = batch_src_labels.cuda()
	n,c,fea_h,fea_w = batch_src_feas.size()
	batch_y_downsample = label_downsample(batch_src_labels, fea_h, fea_w) # n*fea_h*fea_w
	batch_y_downsample = batch_y_downsample.unsqueeze(1) # n*1*fea_h*fea_w
	batch_class_center_fea_list = []
	for i in range(5):
	fea_mask = torch.eq(batch_y_downsample,i).float().cuda() #n*1*fea_h*fea_w
	class_feas = batch_src_feas * fea_mask # n*c*fea_h*fea_w
	class_fea_sum = torch.sum(class_feas, [0, 2, 3]) # c
	class_num = torch.sum(fea_mask, [0, 1, 2, 3])
	if class_num == 0:
	batch_class_center_fea = class_center_feas[i,:].detach()
	else:
	batch_class_center_fea = class_fea_sum/class_num
	batch_class_center_fea = batch_class_center_fea.unsqueeze(0) # 1 * c
	batch_class_center_fea_list.append(batch_class_center_fea)
	batch_class_center_feas = torch.cat(batch_class_center_fea_list,dim=0) # n_class * c
	class_center_feas = m * class_center_feas + (1-m) * batch_class_center_feas
	return class_center_feas

[dongdongtong]

Sovled. Thanks for your kindly reply.