Adding configs

experiments/.gitignore

@@ -1,4 +1,2 @@
-*
-
-!.gitignore
-!example*/*.toml
+*.npy
+*.pth

experiments/example_attack/config.toml

@@ -0,0 +1,38 @@
+# TODO Description
+
+# TODO
+# [train_expert]
+# output = "out/checkpoints/r32p_1xs/0/model.pth"
+# model = "r32p"
+# trainer = "sgd"
+# dataset = "cifar"
+# source_label = 9
+# target_label = 4
+# poisoner = "1xs"
+# epochs = 20
+# checkpoint_iters = 50
+
+# TODO
+[generate_labels]
+input = "out/checkpoints/r32p_1xs/{}/model_{}_{}.pth"
+opt_input = "out/checkpoints/r32p_1xs/{}/model_{}_{}_opt.pth"
+expert_model = "r32p"
+trainer = "sgd"
+dataset = "cifar"
+source_label = 9
+target_label = 4
+poisoner = "1xs"
+output_path = "experiments/example_attack/"
+lambda = 0.0
+
+[generate_labels.expert_config]
+experts = 1
+min = 0
+max = 20
+
+[generate_labels.attack_config]
+iterations = 5
+one_hot_temp = 5
+alpha = 0
+label_kwargs = {lr = 150, momentum = 0.5}
+

experiments/example_downstream/config.toml

@@ -0,0 +1,12 @@
+[downstream]
+input = "experiments/example_attack/"
+downstream_model = "r32p"
+trainer = "sgd"
+dataset = "cifar"
+source_label = 9
+target_label = 4
+poisoner = "1xs"
+output_path = "experiments/example_downstream/"
+logits = false
+alpha = 0.0
+distill_labels = false

modules/base_utils/datasets.py

@@ -177,8 +177,6 @@ def __init__(self, train: Dataset, distill: Dataset, poison_inds, transform, n_c
  self.transform = transform
  self.n_classes = n_classes
 
- print(len(self.train), len(self.distill), len(self.poison_inds))
-
  def __getitem__(self, i: int):
  seed = np.random.randint(8)
  random.seed(seed)
@@ -219,7 +217,6 @@ def __init__(
  if not (indices or eps):
  raise ValueError()
 
- print(np.unique([y for x, y in dataset]))
  if not indices:
  if label is not None:
  clean_inds = [i for i, (x, y) in enumerate(dataset)
@@ -423,7 +420,6 @@ def load_tiny_imagenet_dataset(path, train=True):
  process.wait()
  path = path + ("/train" if train else "/val/images")
  dataset = datasets.ImageFolder(path)
- print(np.unique(dataset.targets))
  return dataset
 
 def make_dataloader(
@@ -600,7 +596,7 @@ def get_matching_datasets(
  big=False
 ):
  train_transform = TRANSFORM_TRAIN_XY[dataset_flag + ('_big' if big else '')]
- test_transform = TRANSFORM_TEST_XY[dataset_flag, + ('_big' if big else '')]
+ test_transform = TRANSFORM_TEST_XY[dataset_flag + ('_big' if big else '')]
 
  train_data = load_dataset(dataset_flag, train=True)
  test_data = load_dataset(dataset_flag, train=False)

modules/base_utils/util.py

@@ -53,8 +53,7 @@ def generate_full_path(path):
 
 
 def extract_toml(experiment_name, module_name=None):
- relative_path = "experiments/" + experiment_name + "/" + experiment_name\
- + ".toml"
+ relative_path = "experiments/" + experiment_name + "/config.toml"
  full_path = generate_full_path(relative_path)
  assert os.path.exists(full_path)
 

modules/mtt_labels/run_module.py → modules/generate_labels/run_module.py

@@ -15,7 +15,7 @@
  get_mtt_attack_info, load_model,\
  either_dataloader_dataset_to_both,\
  make_pbar, clf_loss, needs_big_ims, softmax, total_mse_distance
-from modules.mtt_labels.utils import coalesce_attack_config, extract_experts,\
+from modules.generate_labels.utils import coalesce_attack_config, extract_experts,\
  extract_labels, sgd_step
 
 

modules/train_expert/run_module.py

@@ -8,7 +8,6 @@
 import sys
 
 import torch
-import numpy as np
 
 from modules.base_utils.datasets import get_matching_datasets, pick_poisoner
 from modules.base_utils.util import extract_toml, load_model,\
@@ -33,6 +32,7 @@ def run(experiment_name, module_name, **kwargs):
  poisoner_flag = args["poisoner"]
  clean_label = args["source_label"]
  target_label = args["target_label"]
+ ckpt_iters = args.get("checkpoint_iters")
  train_pct = args.get("train_pct", 1.0)
  batch_size = args.get("batch_size", None)
  epochs = args.get("epochs", None)
@@ -42,7 +42,7 @@ def run(experiment_name, module_name, **kwargs):
  else args["output"].format(slurm_id)
 
  Path(output_path[:output_path.rfind('/')]).mkdir(parents=True,
- exist_ok=True)
+  exist_ok=True)
 
  # TODO: make this more extensible
  if dataset_flag == "cifar_100":
@@ -77,11 +77,8 @@ def run(experiment_name, module_name, **kwargs):
 
  print("Training...")
 
- SAVE_EPOCH = 1
- SAVE_ITER = 50
-
- def checkpoint_callback(model, opt, epoch, iteration, save_epoch, save_iter):
- if epoch % save_epoch == 0 and iteration % save_iter == 0 and iteration != 0:
+ def checkpoint_callback(model, opt, epoch, iteration, save_iter):
+ if iteration % save_iter == 0 and iteration != 0:
  index = output_path.rfind('.')
  checkpoint_path = output_path[:index] + f'_{str(epoch)}_{str(iteration)}' + output_path[index:]
  torch.save(model.state_dict(), generate_full_path(checkpoint_path))
@@ -97,7 +94,7 @@ def checkpoint_callback(model, opt, epoch, iteration, save_epoch, save_iter):
  opt=opt,
  scheduler=lr_scheduler,
  epochs=epochs,
- callback=lambda m, o, e, i: checkpoint_callback(m, o, e, i, SAVE_EPOCH, SAVE_ITER)
+ callback=lambda m, o, e, i: checkpoint_callback(m, o, e, i, ckpt_iters)
  )
 
  print("Evaluating...")
@@ -107,10 +104,6 @@ def checkpoint_callback(model, opt, epoch, iteration, save_epoch, save_iter):
  print(f"{clean_test_acc=}")
  print(f"{poison_test_acc=}")
 
- print("Saving model...")
- torch.save(model.state_dict(), generate_full_path(output_path))
-
-
 if __name__ == "__main__":
  experiment_name, module_name = sys.argv[1], sys.argv[2]
  run(experiment_name, module_name)

schemas/downstream.toml

@@ -1,6 +1,6 @@
 ### 
 # TODO
-# mtt_labels schema
+# downstream schema
 # Configured to poison and train and distill a set of model on any of the datasets.
 # Outputs the .pth of a distileld model
 ###

schemas/mtt_labels.toml → schemas/generate_labels.toml

@@ -1,16 +1,15 @@
 ### 
 # TODO
-# mtt_labels schema
+# generate_labels schema
 # Configured to poison and train and distill a set of model on any of the datasets.
 # Outputs the .pth of a distileld model
 ###
 
-[mtt_labels]
+[generate_labels]
 input = "TODO"
 opt_input = "TODO"
 output_path = "string: Path to .pth file."
 expert_model = "string: (r32p, r18, r18_tin, vgg, vgg_pretrain, vit_pretrain). For ResNets, VGG-19s, and ViTs"
-downstream_model = "string: (r32p, r18, r18_tin, vgg, vgg_pretrain, vit_pretrain). For ResNets, VGG-19s, and ViTs"
 dataset = "string: (cifar / cifar_100 / tiny_imagenet). For CIFAR-10, CIFAR-100 and Tiny Imagenet datasets"
 trainer = "string: (sgd / adam). Specifies optimizer. "
 source_label = "int: {0,1,...,9}. Specifies label to mimic"

schemas/train_expert.toml

@@ -12,6 +12,7 @@ trainer = "string: (sgd / adam). Specifies optimizer. "
 source_label = "int: {0,1,...,9}. Specifies label to mimic"
 target_label = "int: {0,1,...,9}. Specifies label to attack"
 poisoner = "string: Form: {{1,2,3,9}xp, {1,2}xs, {1,4}xl}. Integer resembles number of attacks and string represents type"
+checkpoint_iters = "TODO"
 
 [OPTIONAL]
 batch_size = "int: {0,1,...,infty}. Specifies batch size. Set to default for trainer if omitted."