Some questions about the traininig time and the finally results about ddad HOT 11 CLOSED

Training the model on 15 categories of MVTec for 2 days is realistic since diffusion models are expensive to train.
Some hyperparameter tuning is necessary to achieve optimal results as mentioned in the config file and readme. We will soon publish the best settings for each category.

from ddad.

I also met this question about the results of carprt, there is a huge gap between the report results and the one I reproduced, I try epochs 1000, 1500, 2000 , AD epoch 1,2 3 and many hyperparameter

from ddad.

For carpet w = 0 and DA_chp =0 would be the best. However, I will publish checkpoints and settings very soon.

from ddad.

For carpet w = 0 and DA_chp =0 would be the best. However, I will publish checkpoints and settings very soon.

There is likely no AD_chp in config.yaml, you mean DA_epochs or others?

from ddad.

The DA_epochs indicates the number of iterations for fine-tuning the feature extractor. And DA_chp indicates the checkpoint you load the checkpoint. For carpet setting them to zero would results in the best.

from ddad.

The DA_epochs indicates the number of iterations for fine-tuning the feature extractor. And DA_chp indicates the checkpoint you load the checkpoint. For carpet setting them to zero would results in the best.

AD_chp=0 means I onlt need to train one iteration to fine-tuning the feature extractor for carpet?

from ddad.

It means a pretrained feature extractor outperforms a fine-tuned one.

from ddad.

It means a pretrained feature extractor outperforms a fine-tuned one.

No fine tuning? directly use pretrained feature extractor?

from ddad.

It means a pretrained feature extractor outperforms a fine-tuned one.

No fine tuning? directly use pretrained feature extractor?

Exactly.

from ddad.

It means a pretrained feature extractor outperforms a fine-tuned one.

No fine tuning? directly use pretrained feature extractor?

Exactly.

I directly use pretrained feature extractor and set w=0, I reproduced the best results is:
Image AUROC: 91.7
Pixel AUROC: 93.9
PRO: 81.2

from ddad.

checkpoints are published

from ddad.