CIFAR10-Adversarial-Perturbations.ipynb -- Load pretrained CIFAR-10 models not included and incorrectly named

[miscpeeps]

CIFAR10-Adversarial-Perturbations.ipynb example jupyter notebook and tutorial reference mitll_cifar_l2_1_0.pt and mitll_cifar_nat.pt as pretrained CIFAR-10 models. These models are not included in the standard rai-toolbox[mushin] install (perhaps due to licensing or desire to have most up-to-date models?).

Models download from urls at robustness Github are named cifar_l2_1_0.pt and cifar_nat.pt and will cause the following error on In[10] of CIFAR10-Adversarial-Perturbations.ipynb:

FileNotFoundError                         Traceback (most recent call last)
Input In [8], in <cell line: 3>()
      1 # Load pretrained model that was trained using a robust approach (i.e., adversarial training)
      2 ckpt_robust = "mitll_cifar_l2_1_0.pt"
----> 3 model_robust = load_model(ckpt_robust)
      4 model_robust.eval();
      6 # Load pretrained model that was trained with standard approach

Input In [7], in load_model(ckpt)
      2 def load_model(ckpt):
----> 3     base_model = load_from_checkpoint(
      4         model = resnet50(),
      5         ckpt = ckpt,
      6         weights_key="state_dict",
      7     )
      9     normalizer = transforms.Normalize(
     10         mean=[0.4914, 0.4822, 0.4465],
     11         std=[0.2023, 0.1994, 0.2010],
     12     )
     14     model = nn.Sequential(normalizer, base_model)

File ~/dev/rai-toolbox-james/responsible-ai-toolbox/src/rai_toolbox/mushin/_utils.py:60, in load_from_checkpoint(model, ckpt, weights_key, weights_key_strip, model_attr)
     57     ckpt = Path.home() / ".torch" / "models" / ckpt
     58 log.info(f"Loading model checkpoint from {ckpt}")
---> 60 ckpt_data: Dict[str, Any] = torch.load(ckpt, map_location="cpu")
     62 if weights_key is not None:
     63     assert weights_key in ckpt_data

File ~/anaconda3/envs/rai-toolbox-james/lib/python3.9/site-packages/torch/serialization.py:699, in load(f, map_location, pickle_module, **pickle_load_args)
    696 if 'encoding' not in pickle_load_args.keys():
    697     pickle_load_args['encoding'] = 'utf-8'
--> 699 with _open_file_like(f, 'rb') as opened_file:
    700     if _is_zipfile(opened_file):
    701         # The zipfile reader is going to advance the current file position.
    702         # If we want to actually tail call to torch.jit.load, we need to
    703         # reset back to the original position.
    704         orig_position = opened_file.tell()

File ~/anaconda3/envs/rai-toolbox-james/lib/python3.9/site-packages/torch/serialization.py:231, in _open_file_like(name_or_buffer, mode)
    229 def _open_file_like(name_or_buffer, mode):
    230     if _is_path(name_or_buffer):
--> 231         return _open_file(name_or_buffer, mode)
    232     else:
    233         if 'w' in mode:

File ~/anaconda3/envs/rai-toolbox-james/lib/python3.9/site-packages/torch/serialization.py:212, in _open_file.__init__(self, name, mode)
    211 def __init__(self, name, mode):
--> 212     super(_open_file, self).__init__(open(name, mode))

FileNotFoundError: [Errno 2] No such file or directory: '/home/scott/.torch/models/mitll_cifar_l2_1_0.pt'

Models must be renamed and manually copied to /home/{$USER}/.torch/models to proceed with tutorial.

[jgbos]

This should be updated in our notebook. By default the loader will check in your home directory under .torch/models, but this notebook should be explicit with the location. You should actually just provide the full path to the model and name, e.g.,

ckpt_robust = "/home/scott/.torch/models/cifar_nat.pt"

[rsokl]

Closed by #69 and #73