[Bug]: Setting the threshold using ManualThreshold seems to have no effect on the output of predict

[JinYuannn]

Describe the bug

Setting the threshold using ManualThreshold seems to have no effect on the output of predict.
I set the default_value in ManualThreshold to 0.99 and 0.0001, but the images saved by predict are the same.

Dataset

N/A

Model

Reverse Distillation

Steps to reproduce the behavior

model = ReverseDistillation()

threshold = ManualThreshold(default_value=0.1)

engine = Engine(
        max_epochs=50, 
        limit_train_batches=100, 
        threshold=threshold, 
        check_val_every_n_epoch=2,
        logger=True
    )

engine.predict(
        datamodule=folder_datamodule,
        model=model,
        ckpt_path="my ckpt path" )

OS information

OS information:

• OS: Windows 11
• Python version: 3.10.11
• Anomalib version: 1.2.0.dev0
• PyTorch version: 2.3.1+cu121
• CUDA/cuDNN version: 121
• GPU models and configuration: RTX 4090

Expected behavior

different threshold, different predicted mask

Screenshots

No response

Pip/GitHub

GitHub

What version/branch did you use?

No response

Configuration YAML

None

Logs

None

Code of Conduct

• I agree to follow this project's Code of Conduct

[alexriedel1]

What do your predictions look like and what do you expect? Is the code above your full code, because the model training (and validation) phase is missing.

[JinYuannn]

1. Here is my training code:

model = ReverseDistillation()
    
engine = Engine(
    max_epochs=200, 
    check_val_every_n_epoch=5,
    logger=True
)

 engine.fit(datamodule=folder_datamodule, model=model)

2. When I finish training the model and call the model using the predict method, no matter what value I change the default_value in ManualThreshold to, the 'Predicted Heat Map' and 'Predicted Mask' drawn by predict method are exactly the same. I believe that modifying the threshold should output different 'Predicted Mask'. For example, if I set the threshold very low, there should be many defects.

[JinYuannn]

@alexriedel1 Looking forward to your reply, thanks.