search

yikaiw / TokenFusion

[CVPR 2022] Code release for "Multimodal Token Fusion for Vision Transformers"
https://arxiv.org/pdf/2204.08721.pdf
MIT License
164 stars 15 forks source link

loss2d_class_error 100.0000 #16

Open keithAND2020 opened 9 months ago

keithAND2020 commented 9 months ago

Hi! Is loss2d_class_error 100.0000 normal?

Thanks

[12/21 19:28:05] group-free INFO: Train: [2][30/95] 0head_neg_ratio 0.6538 0head_pos_ratio 0.3462 1head_neg_ratio 0.6538 1head_pos_ratio 0.3462 2head_neg_ratio 0.6538 2head_pos_ratio 0.3462 3head_neg_ratio 0.6538 3head_pos_ratio 0.3462 4head_neg_ratio 0.6538 4head_pos_ratio 0.3462 grad_norm 3.1559 last_neg_ratio 0.6538 last_pos_ratio 0.3462 points_hard_topk4_neg_ratio 0.9848 points_hard_topk4_pos_ratio 0.0152 points_hard_topk4_upper_recall_ratio 1.0000 proposal_neg_ratio 0.6538 proposal_pos_ratio 0.3462
[12/21 19:28:05] group-free INFO: grad_norm: 31.558673858642578 [12/21 19:28:05] group-free INFO: loss 14.4250 loss2d_class_error 100.0000 loss2d_loss_bbox 0.2392 loss2d_loss_ce 0.9980 loss2d_loss_giou 0.6201 query_points_generation_loss 0.0068 sum_heads_box_loss 6.3427 sum_heads_objectness_loss 0.4403 sum_heads_sem_cls_loss 11.6525
[12/21 19:28:05] group-free INFO: last_box_loss 0.9117 last_center_loss 0.5684 last_heading_cls_loss 2.4032 last_heading_reg_loss 0.0199 last_neg_ratio 0.6538 last_objectness_loss 0.0636 last_pos_ratio 0.3462 last_sem_cls_loss 1.6602 last_size_reg_loss 0.0831
[12/21 19:28:05] group-free INFO: proposal_box_loss 0.8211 proposal_center_loss 0.4900 proposal_heading_cls_loss 2.3392 proposal_heading_reg_loss 0.0211 proposal_neg_ratio 0.6538 proposal_objectness_loss 0.0593 proposal_pos_ratio 0.3462 proposal_sem_cls_loss 1.4890 proposal_size_reg_loss 0.0760
[12/21 19:28:05] group-free INFO: 4head_box_loss 0.9417 4head_center_loss 0.5890 4head_heading_cls_loss 2.4262 4head_heading_reg_loss 0.0201 4head_neg_ratio 0.6538 4head_objectness_loss 0.0636 4head_pos_ratio 0.3462 4head_sem_cls_loss 1.7456 4head_size_reg_loss 0.0900
[12/21 19:28:05] group-free INFO: 3head_box_loss 0.9164 3head_center_loss 0.5662 3head_heading_cls_loss 2.4262 3head_heading_reg_loss 0.0206 3head_neg_ratio 0.6538 3head_objectness_loss 0.0620 3head_pos_ratio 0.3462 3head_sem_cls_loss 1.7198 3head_size_reg_loss 0.0870
[12/21 19:28:05] group-free INFO: 2head_box_loss 0.9181 2head_center_loss 0.5710 2head_heading_cls_loss 2.4070 2head_heading_reg_loss 0.0204 2head_neg_ratio 0.6538 2head_objectness_loss 0.0644 2head_pos_ratio 0.3462 2head_sem_cls_loss 1.6655 2head_size_reg_loss 0.0860
[12/21 19:28:05] group-free INFO: 1head_box_loss 0.9180 1head_center_loss 0.5718 1head_heading_cls_loss 2.4213 1head_heading_reg_loss 0.0200 1head_neg_ratio 0.6538 1head_objectness_loss 0.0640 1head_pos_ratio 0.3462 1head_sem_cls_loss 1.6628 1head_size_reg_loss 0.0841
[12/21 19:28:05] group-free INFO: 0head_box_loss 0.9157 0head_center_loss 0.5666 0head_heading_cls_loss 2.4174 0head_heading_reg_loss 0.0201 0head_neg_ratio 0.6538 0head_objectness_loss 0.0633 0head_pos_ratio 0.3462 0head_sem_cls_loss 1.7096 0head_size_reg_loss 0.0873
[12/21 19:28:05] group-free INFO: Mean IoU2d 0.47 [12/21 19:28:05] group-free INFO: loss2d_class_error 100.0000 loss2d_loss_bbox 0.2392 loss2d_loss_ce 0.9980 loss2d_loss_giou 0.6201

yikaiw commented 8 months ago

loss2d_class_error does not matter as long as loss2d_loss_giou keeps increasing. loss2d_loss_giou will be stable in about 120 epochs (we stop learning the 2d part after 120 epochs, https://github.com/yikaiw/TokenFusion/blob/main/object-detection-3d/models/loss_helper.py#L368C43-L368C43).