search

SHI-Labs / OneFormer

OneFormer: One Transformer to Rule Universal Image Segmentation, arxiv 2022 / CVPR 2023
https://praeclarumjj3.github.io/oneformer
MIT License
1.46k stars 129 forks source link

losses are zero during training on custom dataset #56

Closed sahityabonumaddi closed 1 year ago

sahityabonumaddi commented 1 year ago

this is the output after running oneformer model on custom dataset for panoptic segmentation dice loss and contrastive losses are zero after 4999 iterations. any help @praeclarumjj3

wandb: Run summary: wandb: data_time 0.00161 wandb: eta_seconds 72805.84595 wandb: lr 0.0001 wandb: time 0.46956 wandb: train/loss_ce 0.00051 wandb: train/loss_ce_0 0.00188 wandb: train/loss_ce_1 0.00058 wandb: train/loss_ce_2 0.00071 wandb: train/loss_ce_3 0.00056 wandb: train/loss_ce_4 0.00058 wandb: train/loss_ce_5 0.00056 wandb: train/loss_ce_6 0.0005 wandb: train/loss_ce_7 0.00065 wandb: train/loss_ce_8 0.00066 wandb: train/loss_contrastive 0.0 wandb: train/loss_dice 0.0 wandb: train/loss_dice_0 0.0 wandb: train/loss_dice_1 0.0 wandb: train/loss_dice_2 0.0 wandb: train/loss_dice_3 0.0 wandb: train/loss_dice_4 0.0 wandb: train/loss_dice_5 0.0 wandb: train/loss_dice_6 0.0 wandb: train/loss_dice_7 0.0 wandb: train/loss_dice_8 0.0 wandb: train/loss_mask 0.0 wandb: train/loss_mask_0 0.0 wandb: train/loss_mask_1 0.0 wandb: train/loss_mask_2 0.0 wandb: train/loss_mask_3 0.0 wandb: train/loss_mask_4 0.0 wandb: train/loss_mask_5 0.0 wandb: train/loss_mask_6 0.0 wandb: train/loss_mask_7 0.0 wandb: train/loss_mask_8 0.0 wandb: train/total_loss 0.00719

praeclarumjj3 commented 1 year ago

Hi @sahityabonumaddi, you might need to check for any bugs in your code while setting up the custom dataset. Or there might be some overfitting. I can't help you just by looking at the zero-loss values.