Loss is negative when finetune pretrained model on megadepth

[skill-diver]

Hi Author. Thank you for the great work at first.
I get a issue when training your pretrained model on megadepth, I find the loss is negative.

Here's the training scripts I used:

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m torch.distributed.run --nproc_per_node=4 train.py
--train_dataset "68_400 @ MegaDepth(ROOT='megadepth', split='train', resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], aug_crop='auto', aug_monocular=0.005, transform=ColorJitter, n_corres=8192, nneg=0.5)"
--test_dataset "1_000 @ MegaDepth(ROOT='megadepth', split='val', resolution=(512,336), seed=777, n_corres=1024)"
--model "AsymmetricMASt3R(pos_embed='RoPE100', patch_embed_cls='ManyAR_PatchEmbed', img_size=(512, 512), head_type='catmlp+dpt', output_mode='pts3d+desc24', depth_mode=('exp', -inf, inf), conf_mode=('exp', 1, inf), enc_embed_dim=1024, enc_depth=24, enc_num_heads=16, dec_embed_dim=768, dec_depth=12, dec_num_heads=12, two_confs=True, desc_conf_mode=('exp', 0, inf))"
--train_criterion="ConfLoss(Regr3D(L21, norm_mode='avg_dis'), alpha=0.2)"
--test_criterion "Regr3D(L21, norm_mode='?avg_dis', gt_scale=True, sky_loss_value=0) + -1.*MatchingLoss(APLoss(nq='torch', fp=torch.float16), negatives_padding=12288)"
--pretrained "/checkpoints/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric.pth"
--lr 0.0001 --min_lr 1e-06 --warmup_epochs 8 --epochs 50 --batch_size 2 --accum_iter 2
--save_freq 1 --keep_freq 5 --eval_freq 1 --print_freq=10 --disable_cudnn_benchmark
--output_dir "checkpoints/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric"

Is the negative loss caused by my training set up or it's normal?

[yocabon]

1: yes negative confidence loss is normal
2: note that you are not training with the matching loss, but it's there in the validation.