branch: LTS_b2_soc 应用于人像训练会导致手臂细节消失

[MHGL]

损失配置

# soc semantic loss
        downsampled_fusion = F.interpolate(pred_fusion, scale_factor=1/8, mode='nearest')
        downsampled_pseudo_gt_fusion = downsampled_fusion.max(1)[1]
        pseudo_gt_semantic = pred_semantic.max(1)[1]
        soc_semantic_loss = F.cross_entropy(pred_semantic, downsampled_pseudo_gt_fusion.detach()) + \
                            F.cross_entropy(downsampled_fusion, pseudo_gt_semantic.detach())

        backup_fusion, backup_detail, _ = self.config.output_backup
        # sub-objectives consistency between `pred_detail` and `pred_backup_detail` (on boundaries only)
        backup_detail_loss = boundaries * F.cross_entropy(pred_detail, backup_detail.max(1)[1], weight=self.config.classes_weight, reduction='none')
        backup_detail_loss = torch.mean(backup_detail_loss)

        # sub-objectives consistency between pred_matte` and `pred_backup_matte` (on boundaries only)
        backup_fusion_loss = boundaries * F.cross_entropy(pred_fusion, backup_fusion.max(1)[1], reduction='none')
        backup_fusion_loss = torch.mean(backup_fusion_loss)

        self.config.loss = 5 * soc_semantic_loss + backup_detail_loss + backup_fusion_loss

模型表现

• 原模型手臂部分完整检测，但存在部分误检
• soc 模型0eps 误检消除，但是手臂部分未检测到

[MHGL]

试验1

• train.py +73

  # 手臂检测情况略有好转，但误检消除能力变弱，图片存在较多误检
  self.config.loss = 1 * soc_semantic_loss + backup_detail_loss + backup_fusion_loss

  试验2

• train.py +73

  # 手臂检测情况更差，误检消除能力增强，图片存在极少误检
  self.config.loss = 10 * soc_semantic_loss + backup_detail_loss + backup_fusion_loss