Calculate the loss

The loss calculation, which should be overrideen when the attack change the loss calculation (e.g., ATA, etc.)

generating the matrices: \\mathcal{L}, \\mathcal{H} :return: self.matrix_low = \\mathcal{L}, self.matrix_high = \\mathcal{H}

生成变换矩阵 generating the matrices: \\mathcal{L}, \\mathcal{H} :return: self.matrix_low = \\mathcal{L}, self.matrix_high = \\math

生成变换矩阵 generating the matrices: \\mathcal{L}, \\mathcal{H} :return: self.matrix_low = \\mathcal{L}, self.matrix_high = \\math

生成变换矩阵 generating the matrices: \\mathcal{L}, \\mathcal{H} :return: self.matrix_low = \\mathcal{L}, self.matrix_high = \\math

Apply Mixed-Frequency Input transformation. - With probability (1 - diversity_prob), bypass and return x (no transform). - Us

Create a function to give the output of the score-based model. Args: model: The score model. train: `True` for training and `False` f

The momentum calculation

The momentum calculation

Return the dataset's normalization layer

Probability flow ODE sampler with the black-box ODE solver. Args: sde: An `sde_lib.SDE` object that represents the forward SDE. shape: A se

x: (C, H, W)· returns (C), (C), (C)

Create a Predictor-Corrector (PC) sampler. Args: sde: An `sde_lib.SDE` object representing the forward SDE. shape: A sequence of integers.

Calculate the gradients of the sampled points

Calculate momentum for robust token updates

Calculate the sampled gradients

Create a loss function for training with arbirary SDEs. Args: sde: An `sde_lib.SDE` object that represents the forward SDE. train: `True` f

Legacy code to reproduce previous results on SMLD(NCSN). Not recommended for new work.

This matches the implementation in Denoising Diffusion Probabilistic Models: From Fairseq. Build sinusoidal embeddings. This matches

Model checkpoints can be download from: https://drive.google.com/drive/folders/14unrFNAkglh4y-Jsas29WOpmMydt8-bm?usp=sharing

Calculate update direction

Calculate the gradient variance

2-dimentional Inverse Discrete Cosine Transform (IDCT)

The inverse to 2D DCT-II, which is a scaled Discrete Cosine Transform, Type III Our definition of idct is that idct_2d(dct_2d(x)) ==

The inverse to 2D DCT-II, which is a scaled Discrete Cosine Transform, Type III Our definition of idct is that idct_2d(dct_2d(x)) ==

The inverse to 2D DCT-II, which is a scaled Discrete Cosine Transform, Type III Our definition of idct is that idct_2d(dct_2d(x)) ==

Sample from p(x_{t-1} | x_t)

Initialize robust token perturbations