Method forward

lesson7-Detection/src/models/yolo_training.cpp:171–245

}

std::vector<torch::Tensor> YOLOLossImpl::forward(torch::Tensor input, std::vector<torch::Tensor> targets)
{
	//inputΪbs, 3 * (5 + num_classes), 13, 13
	//һ��������ͼƬ
	auto bs = input.size(0);
	//������ĸ�
	auto in_h = input.size(2);
	//������Ŀ�
	auto in_w = input.size(3);

	//���㲽����ÿһ���������Ӧԭ����ͼƬ�϶��ٸ����ص�
	//���������Ϊ13x13�Ļ���ԭͼ416x416������£�һ��������Ͷ�Ӧԭ����ͼƬ�ϵ�32�����ص�
	auto stride_h = image_size[1] / in_h;
	auto stride_w = image_size[0] / in_w;

	//�������ijߴ�������������С����ʽ
	//�������������������϶�Ӧ�Ŀ���
	auto scaled_anchors = anchors.clone();
	scaled_anchors.select(1, 0) = scaled_anchors.select(1, 0) / stride_w;
	scaled_anchors.select(1, 1) = scaled_anchors.select(1, 1) / stride_h;

	//bs, 3 * (5 + num_classes), 13, 13->bs, 3, 13, 13, (5 + num_classes)
	auto prediction = input.view({bs, int(num_anchors / 2), bbox_attrs, in_h, in_w}).permute({0, 1, 3, 4, 2}).contiguous();
	//��predictionԤ����е���
	auto conf = torch::sigmoid(prediction.select(-1,4));//  # Conf
	auto pred_cls = torch::sigmoid(prediction.narrow(-1, 5, num_classes));  //Cls pred.

	//�ҵ���Щ������ڲ���������
	auto temp = get_target(targets, scaled_anchors, in_w, in_h, ignore_threshold);
	auto BoolType = torch::ones(1).to(torch::kBool).to(device).options();
	auto FloatType = torch::ones(1).to(torch::kFloat).to(device).options();
	auto mask = temp[0].to(BoolType); 
	auto noobj_mask = temp[1].to(device);
	auto t_box = temp[2];  
	auto tconf = temp[3];  
	auto tcls = temp[4];  
	auto box_loss_scale_x = temp[5];
	auto box_loss_scale_y = temp[6];

	auto temp_ciou = get_ignore(prediction, targets, scaled_anchors, in_w, in_h, noobj_mask);
	noobj_mask = temp_ciou[0];
	auto pred_boxes_for_ciou = temp_ciou[1];


	mask = mask.to(device);
	noobj_mask = noobj_mask.to(device);
	box_loss_scale_x = box_loss_scale_x.to(device);
	box_loss_scale_y =  box_loss_scale_y.to(device);
	tconf = tconf.to(device);
	tcls = tcls.to(device);
	pred_boxes_for_ciou = pred_boxes_for_ciou.to(device);
	t_box = t_box.to(device);


	auto box_loss_scale = 2 - box_loss_scale_x * box_loss_scale_y;
	auto ciou = (1 - box_ciou(pred_boxes_for_ciou.index({mask}), t_box.index({ mask })))* box_loss_scale.index({ mask });
	auto loss_loc = torch::sum(ciou / bs);