Computes the combined loss for a particular transformation

Takes in numpy examples and labels and tensor-ifies and cuda's them if necessary

Adds the name of a 'data series' where each data series is a list of data-entries, where each data-entry is of the form ((epo

A fast approximation of the cumulative distribution function of the standard normal.

Special method to set an attribute if it doesn't exist in this object yet. throws error if this attr already exists ARGS:

Builds PGD examples for the given examples with l_inf bound and given step size. Is almost identical to the BIM attack, except

Returns the per-example norm of the examples, keeping along the specified dimension. e.g. if examples is NxCxHxW, applying this fxn w

Evaluate a function without caching intermediate activations, allowing for reduced memory at the expense of extra compute in the backward pas

If comm is provided, average all numerical stats across that comm

A counter for the `thop` package to count the operations in an attention operation. Meant to be used like: macs, params = thop.pr

Defaults for image and classifier training.

This takes in the minibatch self's parameters were tuned for and outputs a variable of the perturbation applied to the images ARG

Wraps `score_fn` so that the model output corresponds to a real time-dependent score function. Args: sde: An `sde_lib.SDE` object that repr

Create the drift and diffusion functions for the reverse SDE

Compute the loss function. Args: model: A score model. batch: A mini-batch of training data. Returns: loss: A scalar that

Copy model parameters into a (differently-shaped) list of full-precision parameters.

Compute the KL divergence between two gaussians. Shapes are automatically broadcasted, so batches can be compared to scalars, among othe

Sample x_{t-1} from the model at the given timestep. :param model: the model to sample from. :param x: the current tensor at

Generate one sample from the prior distribution, $p_T(x)$.

Diffuse the data for a given number of diffusion steps. In other words, sample from q(x_t | x_0). :param x_start: the initi

Create the reverse-time SDE/ODE. Args: score_fn: A time-dependent score-based model that takes x and t and returns the score. probabi

Create the drift and diffusion functions for the reverse SDE

Propagates changes of the 'use_gpu' parameter down to the attack ARGS: use_gpu : bool - if True, the attack uses the GPU, ow it d

Sets the verbosity or loglevel for training. Is called in .train method so this method doesn't need to be explicitly called.

https://stackoverflow.com/a/1094933 answer by Sridhar Ratnakumar

Create sinusoidal timestep embeddings. :param timesteps: a 1-D Tensor of N indices, one per batch element. These may b

Flatten a torch tensor `x` and convert it to numpy.

Modifies the NN weights of self.classifier_net by training with the specified parameters s ARGS: data_loader: torch.u

Construct a Variance Preserving SDE. Args: model: diffusion model score_type: [guided_diffusion, score_sde, ddpm]

Construct a Variance Preserving SDE. Args: model: diffusion model score_type: [guided_diffusion, score_sde, ddpm]

Construct a Variance Preserving SDE. Args: model: diffusion model score_type: [guided_diffusion, score_sde, ddpm]

Subroutine to run the specified attack on a minibatch and append the results to inputs/labels. NOTE: THIS DOES NOT MUTATE inputs

Returns a list of indices of valid adversarial examples ARGS: example_logits: Variable/Tensor (Nx#Classes) - output logits for a

Uses the distance_fxn and carlini_loss to create a loss function to be optimized ARGS: initial_lambda : float - which

Loads the numpy array of examples given the random 'prefix'

Extract values from a 1-D numpy array for a batch of indices. :param arr: the 1-D numpy array or a func. :param timesteps: a tensor of i

Extract values from a 1-D numpy array for a batch of indices. :param arr: the 1-D numpy array or a func. :param timesteps: a tensor of i

Extract values from a 1-D numpy array for a batch of indices. :param arr: the 1-D numpy array or a func. :param timesteps: a tensor of i

Finds the class folders in a dataset. Ensures: No class is a subdirectory of another.

Loads the numpy array of labels given the random 'prefix'

DANGEROUS TO BE CALLED OUTSIDE OF THIS FILE!!!

DANGEROUS TO BE CALLED OUTSIDE OF THIS FILE!!!

DANGEROUS TO BE CALLED OUTSIDE OF THIS FILE!!!

Does one step of optimization

Get the prior KL term for the variational lower-bound, measured in bits-per-dim. This term can't be optimized, as it only de

Modifies params going forward