Closed cestbonsuliu closed 3 months ago
You need to read the paper to understand what is being done.
Then in the fine-tuning stage, the entire MAE should lose the reconstruction capability because the encoder should fine-tune its embedding output to fit the downstream data, regardless of what the decoder expects.
If you pre-train on your dataset, then the MAE will be able to learn to reconstruct your data, but the same goes. Fine-tuning to something will break the reconstruction. No fine-tuning if you need the reconstruction.
Be careful you are using a dataset big enough for pre-training.
thanks
I am trying to perform anomaly localization after fine-tuning on the MvTec dataset. Here are some of my settings:
The file structure of the MvTec dataset is as follows:
The "train" folder contains normal samples, while the "val" folder contains abnormal samples.
My fine-tuning command is:
Then, I tested the following image:
The obtained results are as follows:
Can anyone help me with this issue?