Application to the new scence

[zsicl92]

Hi, @heyufan1995

When I train the VISTA3D on the new dataset, I encounter the following issue, could you please help me? Thanks a lot.

部分修改的yml文件如下：

drop_label_prob: 0
drop_point_prob: 1
finetune: {activate: false, exclude_vars: null, pretrained_ckpt_name: null}
loss: {target: DiceCELoss, include_background: true, sigmoid: false, smooth_dr: 1.0e-05, smooth_nr: 0, softmax: true, squared_pred: true,
to_onehot_y: true} # 带背景训练
lr_scheduler: {target: monai.optimizers.WarmupCosineSchedule, optimizer: $@optimizer,
t_total: $@num_epochs+1, warmup_multiplier: 0.1, warmup_steps: 0}
max_backprompt: 0
max_foreprompt: 0
max_point: 0 #3
max_prompt: 0 #null
num_epochs: 300
freeze_epoch: 300
freeze_head: "point" #0 #'auto'

发现在运行的过程中，直接跳过训练

logger.debug(f"Iteration skipped due to None prompts at {ds_name}")

想请教一下，我们在制作数据的过程中，必须包含label_prompt、point、label_prompt_pseudo、point_pseudo中的一个数据么？
还是在做键映射的时候需要特殊修改？如果只有Image和label的时候，该怎么去设置呢？

[heyufan1995]

if you want to finetune auto only. Just follow https://github.com/Project-MONAI/VISTA/blob/main/vista3d/configs/finetune/train_finetune_word.yaml

Yes. label_prompt、point、label_prompt_pseudo、point_pseudo must exist at least one, otherwise there is no label to train on.

In your setting, if you want to train auto only,

drop_label_prob: 0
drop_point_prob: 1
finetune: {activate: false, exclude_vars: null, pretrained_ckpt_name: null}
loss: {target: DiceCELoss, include_background: true, sigmoid: false, smooth_dr: 1.0e-05, smooth_nr: 0, softmax: true, squared_pred: true,
to_onehot_y: true} # 带背景训练
lr_scheduler: {target: monai.optimizers.WarmupCosineSchedule, optimizer: $@optimizer,
t_total: $@num_epochs+1, warmup_multiplier: 0.1, warmup_steps: 0}
max_backprompt: null 
max_foreprompt: null # you cannot set to 0 because this defines how many foreground class are sampled
max_point: 0 
max_prompt: null #null
num_epochs: 300
freeze_epoch: 300 # If you want to freeze the encoder and the point branch, you can set it to 300. Otherwise set to 0. 
freeze_head: "point"