Skip to content

Latest commit

 

History

History
130 lines (102 loc) · 7.63 KB

README.md

File metadata and controls

130 lines (102 loc) · 7.63 KB
Raw

📈 CLIP

Implementation of the CLIP algorithm for Lipschitz regularization of neural networks, proposed in CLIP: Cheap Lipschitz Training of Neuronal Networks [1]. Feel free to use it and please refer to our paper when doing so.

@inproceedings{bungert2021clip,
  title={CLIP: Cheap Lipschitz training of neural networks},
  author={Bungert, Leon and Raab, Ren{\'e} and Roith, Tim and Schwinn, Leo and Tenbrinck, Daniel},
  booktitle={International Conference on Scale Space and Variational Methods in Computer Vision},
  pages={307--319},
  year={2021},
  organization={Springer}
}
}

✔️ Benchmarks

Regularization Clean Noise PGD-L2 (epsilon = 2.0)
None 97.1 34.5 31.2
CLIP95 96.4 29.8 71.4
CLIP97 97.0 34.6 72.4

The table above shows the accuracy in [%] for a fully connected model (two hidden layers with 200 and 80 neurons) for the MNIST dataset. We evaluate the accuracy (averaged over 3 runs) on the clean test set and the accuracy under different adversarial attacks. The runs can be reproduced with the configurations below.

💡 Method Description

The CLIP Algorithm proposes a regularization for controlling the Lipschitz constant of a neural network. For a neural network

parametrized by weights the goal is to minimize a loss term of the form

where

  • denotes the training set,
  • denotes a loss function.

The Lipschitz constant of the net w.r.t. the input space variable is defined as

In the algorithm we approximate this constant via difference quotients on a finite subset

In order to ensure that the tuples in this set correspond to point with that realize a high difference quotient the set is updated iteratively via a gradient ascent scheme.

🔧 Usage

Our code is implemented Python and utilizes PyTorch [2]. An example how to use the code is provided in the file main.py. Therein the dictionary conf_arg specifies the configuration of a run. The specification conf = cf.plain_example(data_file, use_cuda=False, download=False) will train an unregularized network, whereas conf = cf.clip_example(data_file, use_cuda=False, download=False) utilizes our algorithm. Executing main.py will throw an error if the dataset is not available. You can change download=False to download=True for automatically downloading the dataset.

CUDA Settings

  • conf_arg['use_cuda']: Boolean that specifies wether the model should be trained on the CPU, default: False.

Regularization

You can specify the following options for the regularizer:

  • conf_arg['regularization']: Specifies which regularization to use:
    • "global_lipschitz": activates the CLIP regularization as described above (Default),
    • "none": deactivates any kind of additional regularization.
  • conf_arg['reg_iters']: The number of gradient ascent steps for the Lipschitz set update, default: reg_iters=1.
  • conf_arg['reg_lr']: Step size for the gradient ascent scheme, default: reg_lr=1.0.
  • conf_arg['reg_interval']: Specifies in which interval the regularization is applied, default: reg_interval=1.
  • conf_arg['reg_max']: Specifies the maximum value of a Lipschitz constant that is allowed to enter the backprop. Note, that large Lipschitz regularization terms yield numerical instabilities, default: reg_max=5e3.
  • conf_arg['reg_init']: Specifies the initialization strategy of the Lipschitz set.
    • "plain": Splits a batch of the loader equally and assigns each half to u and v (Default).
    • "partial_random": Assigns u to a batch of the loader and sets v = u + delta, where delta is sampled from gaussian distribution.
    • "noise": Chooses u,v as samples from the normal distribution.
  • conf_arg['goal_accuracy']:
  • conf_arg['lamda']: The regularization parameter, note that we use this spelling, because lambda is blocked keyword. Secondly lamda is actually the correct spelling of the greek letter, default: lamda=0.0.
  • conf_arg['lamda_incremental']: The incremental update for the regularization parameter as described in [1].

Adversarial Attack

General:

  • gauss_attack(nl=1.0):

  • fgsm(model, loss, epsilon=0.3):

  • pgd(model, loss, epsilon=None, alpha=None, alpha_mul=1.0, restarts=1, attack_iters=7, norm_type="l2"):

Datasets

The example loads the data via helper methods which then call the standard dataloaders provided by PyTorch.

  • conf_arg['data_file']: Specifies the path to the dataset you wish to use.
  • conf_arg['data_set']:: Name of the dataset.
    • "MNIST" (Default, [3])
    • "Fashion-MNIST" ([4]).

The dataloaders are then created by

train_loader, valid_loader, test_loader = get_data_set(conf.dataset, conf.data_file, conf.batch_size)

Note: The download flags for the torchvision dataset methods are all set to False by default and can be set to True in the main.pyscript as described above. You can easily substitute this by your other dataloaders as long the three loaders train_loader, valid_loader, test_loader are specified.

Model

The example loads a simple fully connected net from the file model.py.

  • conf_arg['model']: Specifies the model that should be loaded from model.py.
    • "fc" (Default, fully connected model, currently only possibility).
  • conf_arg['activation_function']: Specifies the activation function for the net.
    • "ReLU" (Default),
    • "sigmoid".

The model is then loaded via

model = models.fully_connected([784, 400, 200, 10], conf.activation_function)
model.to(conf.device)

where [784, 400, 200, 10] denotes the layer dimensions. Alternatively, you can use an arbitrary PyTorch model, i.e., a subclass of nn.Module.

📝 References

[1] Leon Bungert, René Raab, Tim Roith, Leo Schwinn, Daniel Tenbrinck. "CLIP: Cheap Lipschitz Training of Neuronal Networks." arXiv preprint arXiv:2103.12531 (2021). https://arxiv.org/abs/2103.12531

[2] The Pytorch website https://pytorch.org/, see also their git https://github.com/pytorch/pytorch

[3] The MNIST dataset http://yann.lecun.com/exdb/mnist/

[4] The Fashion-MNIST dataset https://research.zalando.com/welcome/mission/research-projects/fashion-mnist/