Save input image to given location, add poison_delta if necessary.

Torch->PIL pipeline as in torchvision.utils.save_image.

https://discuss.pytorch.org/t/dropout-at-test-time-in-densenet/6738/6.

https://discuss.pytorch.org/t/dropout-at-test-time-in-densenet/6738/6.

Attack within given constraints with task as in _objective.

Attack within given constraints with task as in _objective.

Attack within given constraints with task as in _objective. This is effectively a slight mixing. with mixing factor lmb = 1 - eps / 255.

Attack within given constraints with task as in _objective.

Attack within given constraints with task as in _objective.

Attack within given constraints with task as in _objective.

Attack within given constraints with task as in _objective.

Attack within given constraints with task as in _objective.

Attack within given constraints with task as in _objective.

Attack within given constraints with task as in _objective.

Attack within given constraints with task as in _objective.

This function will be evaluated on all GPUs.

This function will be evaluated on all GPUs.

This function will be evaluated on all GPUs.

This function will be evaluated on all GPUs.

This function will be evaluated on all GPUs.

This function will be evaluated on all GPUs.

This function will be evaluated on all GPUs.

This function will be evaluated on all GPUs.

This function will be evaluated on all GPUs.

This function will be evaluated on all GPUs.

This function will be evaluated on all GPUs.

This function will be evaluated on all GPUs.

Compute function on all models and join computations. Distributed hmm

Compute function on all models. Function has arguments that are possibly sequences of length args.ensemble

r"""Compute function on the given optimization problem, defined by criterion \circ model. Function has arguments: model, criterion

Compute loss. This is just a normal loss for cutout.

Compute loss. Here the loss is computed as worst-case estimate over the trials.

Carlini-Wagner loss for brewing [Liam's version].

CW variant 2. This is assert-level equivalent.

Control distributed poison brewing, no-op in single network training.

Switch everything into evaluation mode.

Switch everything into evaluation mode.

run cutout.

run cutmix.

Live Patch ... :> ...

Freezes all parameters and then unfreeze the last layer.

Freezes all parameters and then unfreeze the last layer.

Return only the target and its id. Args: index (int): Index Returns: tuple: (target, idx) where target is cl

Return only the target and its id. Args: index (int): Index Returns: tuple: (target, idx) where target is cl

Return only the target and its id. Args: index (int): Index Returns: tuple: (target, idx) where target is cl

Return only the target and its id.

Compute the gradient of criterion(model) w.r.t to given data.

Compute the gradient of criterion(model) w.r.t to given data.

Construct the central argument parser, filled with useful defaults. The first block is essential to test poisoning in different scenarios. Th

Choose targets from some label which will be poisoned toward some other chosen label, by modifying some subset of the training data within som

Choose targets from some label which will be poisoned toward some other chosen label, by modifying some subset of the training data within som

Choose targets from some label which will be poisoned toward some other chosen label. Using the subset of the training data within some bound

Save output to a csv table.

Reset scheduler objects to initial state.

Reset scheduler object to initial state.

Reset scheduler object to initial state.

Switch pytorch into a deterministic computation mode.

Step through a model epoch. Optionally minimize target loss during this. This function is limited because it assumes that defs.batch_size, de

Sync gradients of given variable across all workers.

Decide and print GPU / CPU / hostname info.