Importance-sample timesteps for a batch. :param batch_size: the number of timesteps. :param device: the torch device to save

Randomly sample an example x' in B_epsilon(x)

Select operations based on learned probabilities

Store REST tokens for query, key, and value

Setup hyperparameters based on model type

Create a list of timesteps to use from an original diffusion process, given the number of timesteps we want to take from equally-sized portio

Save the current parameters for restoring later. Args: parameters: Iterable of `torch.nn.Parameter`; the parameters to be temp

Swin-S @ 224x224, trained ImageNet-1k

Converts a torch Tensor into an image Numpy array Input: 4D(B,(3/1),H,W), 3D(C,H,W), or 2D(H,W), any range, RGB channel order Output: 3D(

Calculate the probability trace for selected operations

Admix the input for Admix Attack

Random transform the input images

Random transform the input images

Use an arbitrary style transfer network to transform the images into different domains Mix up the generated images added by random no

Random transform the input images

Use DCT to transform the input image from spatial domain to frequency domain, Use IDCT to transform the input image from frequency do

Use DCT to transform the input image from spatial domain to frequency domain, Use IDCT to transform the input image from frequency do

Use DCT to transform the input image from spatial domain to frequency domain, Use IDCT to transform the input image from frequency do

Update adversarial perturbation

Update target parameters to be closer to those of source parameters using an exponential moving average. :param target_params: the targe

Update robust tokens using gradient information

Update the reweighting using losses from a model. Sub-classes should override this method to update the reweighting using lo

Update the reweighting using losses from a model. Call this method from each rank with a batch of timesteps and the correspo

Fused `upsample_2d()` followed by `tf.nn.conv2d()`. Padding is performed only once at the beginning, not between the operations. The f

Ported from JAX.

Get a numpy array of weights, one per diffusion step. The weights needn't be normalized, but must be positive.

Args: windows: (num_windows*B, window_size, window_size, C) window_size (int): Window size H (int): Height of image

Apply attack operations to attention and FFN modules

Add normalization layer with mean and std in training configuration

Replace the forward_features method of VisionTransformer with our custom version This enables injection of robust tokens during forward pass