peal.explainers.counterfactual_explainer.ACEConfig¶

class peal.explainers.counterfactual_explainer.ACEConfig[source]¶

Bases: ExplainerConfig

This class defines the config of a ACE, DiME or FastDiME explainer. This config is primarily for replicating results of related work.

Methods

Attributes

`attack_checkpoint_backward_steps`
`attack_epsilon`
`attack_iterations`
`attack_joint`
`attack_joint_checkpoint`
`attack_joint_shortcut`
`attack_method`
`attack_step`
`attempts`
`attention_resolutions`
`base_path`
`batch_size`
`channel_mult`
`chunk`
`chunks`
`class_cond`
`classifier_path`
`classifier_scales`
`clip_denoised`
`cudnn_deterministic`
`data_config`
`data_dir`
`dataset`
`diffusion_steps`
`dilation`
`dist_l1`
`dist_l2`
`dist_schedule`
`distilled_predictor`
`dropout`
`exp_name`
`explainer_type`	The type of explanation that shall be used.
`generator`
`gpu`
`guided_iterations`
`image_size`
`l1_loss`
`l2_loss`
`l_perc`
`l_perc_layer`
`learn_sigma`
`loss_fn`
`masking_threshold`
`merge_and_eval`
`merge_chunks`
`method`
`model_path`
`n_samples`
`noise_schedule`
`num_batches`
`num_channels`
`num_chunks`
`num_head_channels`
`num_heads`
`num_heads_upsample`
`num_res_blocks`
`num_samples`
`oracle_path`
`output_path`
`percentage`
`predict_xstart`
`predictor`
`query_label`
`replace_with_activation`	leaky_relu, leaky_softplus
`resblock_updown`
`rescale_learned_sigmas`
`rescale_timesteps`
`sampling_dilation`
`sampling_inpaint`
`sampling_scale`
`sampling_stochastic`
`sampling_time_fraction`
`save_images`
`save_x_t`
`save_z_t`
`scale_grads`
`seed`	The seed of all randomness to make results reproducible.
`self_optimized_masking`
`shortcut_label_name`
`start_step`
`subtype`
`target_label`
`task_label`
`task_label_name`
`timestep_respacing`
`use_checkpoint`
`use_ddim`
`use_fp16`
`use_kl`
`use_logits`
`use_new_attention_order`
`use_sampling_on_x_t`
`use_scale_shift_norm`
`use_train`
`warmup_step`
`y_target_goal_confidence`

explainer_type: str = 'ACE'¶: The type of explanation that shall be used. Options: [‘counterfactual’, ‘lrp’]

subtype: str = 'ACE'¶

loss_fn: None | str = None¶

predictor: str | None | dict = None¶

generator: None | GeneratorConfig = None¶

data_config: None | DataConfig = None¶

attack_iterations: list | int = 100¶

sampling_time_fraction: list | float = 0.3¶

dist_l1: list | float = 0.0001¶

dist_l2: list | float = 0.0¶

sampling_inpaint: list | float = 0.2¶

sampling_dilation: list | int = 17¶

timestep_respacing: list | int = 50¶

distilled_predictor: None | str | dict = None¶

attempts: int = 1¶

clip_denoised: bool = True¶

batch_size: int = 1¶

gpu: str = '0'¶

save_images: bool = False¶

num_samples: int = 500000000000¶

cudnn_deterministic: bool = False¶

base_path: str = ''¶

exp_name: str = 'example_name'¶

seed: int = 4¶: The seed of all randomness to make results reproducible.

attack_method: str = 'PGD'¶

attack_epsilon: float = 255¶

attack_step: float = 1.0¶

attack_joint: bool = True¶

attack_joint_checkpoint: bool = False¶

attack_checkpoint_backward_steps: int = 1¶

attack_joint_shortcut: bool = False¶

dist_schedule: str = 'none'¶

sampling_stochastic: bool = True¶

chunks: int = 1¶

chunk: int = 0¶

merge_chunks: bool = False¶

y_target_goal_confidence: None | float = None¶

replace_with_activation: str = ''¶

leaky_relu, leaky_softplus

Type:: The activation function ReLU is replaced with

guided_iterations: object = 9999999¶

image_size: object = 128¶

l1_loss: object = 0.05¶

l2_loss: object = 0.0¶

l_perc: object = 30.0¶

l_perc_layer: object = 18¶

learn_sigma: object = True¶

merge_and_eval: object = False¶

model_path: object = 'models/ddpm-celeba.pt'¶

noise_schedule: object = 'linear'¶

num_batches: object = 1¶

num_channels: object = 128¶

num_chunks: object = 1¶

num_head_channels: object = -1¶

num_heads: object = 4¶

num_heads_upsample: object = -1¶

num_res_blocks: object = 2¶

oracle_path: object = 'models/oracle.pth'¶

output_path: object = '/path/to/results'¶

predict_xstart: object = False¶

query_label: object = 31¶

resblock_updown: object = True¶

rescale_learned_sigmas: object = False¶

rescale_timesteps: object = False¶

sampling_scale: object = 1.0¶

save_x_t: object = True¶

save_z_t: object = True¶

start_step: object = 60¶

target_label: object = -1¶

use_checkpoint: object = False¶

use_ddim: object = False¶

use_fp16: object = True¶

use_kl: object = False¶

use_logits: object = True¶

use_new_attention_order: object = False¶

use_sampling_on_x_t: object = True¶

use_scale_shift_norm: object = True¶

use_train: object = False¶

attention_resolutions: object = [32, 16, 8]¶

channel_mult: object = ''¶

class_cond: object = False¶

classifier_path: object = '/scratch/ppar/models/classifier.pth'¶

classifier_scales: object = [8, 10, 15]¶

data_dir: object = '/scratch/ppar/data/img_align_celeba/'¶

dataset: object = 'CelebA'¶

diffusion_steps: object = 500¶

dilation: object = 5¶

dropout: object = 0.0¶

masking_threshold: object = 0.15¶

method: object = 'fastdime'¶

n_samples: object = 1000¶

percentage: object = 0.5¶

scale_grads: object = False¶

self_optimized_masking: object = True¶

shortcut_label_name: object = 'Smiling'¶

task_label: object = 39¶

task_label_name: object = 'Young'¶

warmup_step: object = 30¶