imagenet

TTAC++ on ImageNet

TTAC++ on ImageNet under common corruptions.

Requirements

• To install requirements:

  pip install -r requirements.txt
  

• To download dataset:

  We need to firstly download the validation set and the development kit (Task 1 & 2) of ImageNet-1k on here, and put them under data folder.

  The structure of the data folder should be like

  data
  |_ ILSVRC2012_devkit_t12.tar
  |_ ILSVRC2012_img_val.tar
  

• To create the corruption dataset

  python utils/create_corruption_dataset.py
  

  The issue Frost missing after pip install can be solved following here.

  Finally, the structure of the data folder should be like

  data
  |_ ILSVRC2012_devkit_t12.tar
  |_ ILSVRC2012_img_val.tar
  |_ val
      |_ n01440764
      |_ ...
  |_ corruption
      |_ brightness.pth
      |_ contrast.pth
      |_ ...
  |_ meta.bin
  

Pre-trained Models

Here, we use the pretrain model provided by torchvision.

Results

We mainly conduct our experiments under the sTTT (N-O) protocol, which is more realistic and challenging.

• run TTAC++ on ImageNet-C under the sTTT (N-O) protocol.

  bash scripts/run_ttac2_no.sh
  

  The following results are yielded by the above script (classification errors) under the snow corruption:

  Method	ImageNet-C (Level 5)
  Test	82.22
  TTAC	44.56
  TTAC++	43.40

• run TTAC on ImageNet-C under the N-O-SF without any source information including source statistics collected from training set.

  Note: In this work, we endeavor to mitigate the dependence of previous work on source statistics from training set. We derive the approximated source domain distribution via gradient descent as implemented in utils/find_prototypes.py.

  bash scripts/run_ttac2_no_sf.sh
  

  The following results are yielded by the above script (classification errors) under the snow corruption:

  Method	ImageNet-C (Level 5)
  Test	82.22
  TTAC++	43.60