[Task]: EfficientAD: Bad results in custom dataset with 1024*1040

[jiamfeng]

What is the motivation for this task?

Try to get results in custom datasets with segmentation task

Describe the solution you'd like

when I use the efficientAd in custom dataset with 1024*1040(h * w), the results is very bad, how can I optimizer ?

Additional context

dataset:
  name: xinyuan
  format: folder
  path: ./datasets/TS/gemo
  normal_dir: train # name of the folder containing normal images.
  abnormal_dir: test # name of the folder containing abnorm
  normal_test_dir: null # name of the folder containing normal test images.
  mask: null
  task: segmentation
  extensions: ".bmp"
  train_batch_size: 1
  eval_batch_size: 1
  mask_dir: null
  num_workers: 8
  image_size: [1024, 1040] # dimensions to which images are resized (mandatory)
  center_crop: null # dimensions to which images are center-cropped after resizing (optional)
  normalization: none # data distribution to which the images will be normalized: [none, imagenet]
  transform_config:
    train: null
    eval: null
  test_split_mode: none # options: [from_dir, synthetic]
  test_split_ratio: 0.2 # fraction of train images held out testing (usage depends on test_split_mode)
  val_split_mode: same_as_test # options: [same_as_test, from_test, synthetic]
  val_split_ratio: 0.5 # fraction of train/test images held out for validation (usage depends on val_split_mode)

model:
  name: efficient_ad
  teacher_out_channels: 384
  model_size: small # options: [small, medium]
  lr: 0.0001
  weight_decay: 0.00001
  padding: false
  pad_maps: true # relevant for "padding: false", see EfficientAd in lightning_model.py
  # generic params
  normalization_method: min_max # options: [null, min_max, cdf]

[alexriedel1]

Do you get good results for other input image resolutions?

[jiamfeng]

no, I think the results is more relative with resolution of the pretrained weights

[samet-akcay]

Hi @jiamfeng, this does not seem to be an issue, but more a question if I understand it correctly. I'm moving this to the Q&A, let's continue there. Cheers!

[abc-125]

It is worth trying to train it with the default 256*256 resolution. The feature extractor in the EfficientAD was trained to extract similar features from 256*256 as WideResNet101 from 512*512, so this might be sufficient for you.

I guess with the higher resolution, EffcientAD can be less effective because the size of one patch in its feature extractor is still 33*33, so the receptive field of every patch would be too small.

[nguyenanhtuan1008]

I got the efficientAd check point model after training . Now how do I inference it?.In my case, lightning_inference.py not work, so I trying to export it to Torch and Openvino but dont know how to do it?