def generate_mnist_experiment(train, test, output, train_output_csv_file, test_output_csv_file):
logger.info("Generating experiment...")
# Setup the files based on user inputs
train_csv_file = os.path.abspath(train)
test_csv_file = os.path.abspath(test)
if not os.path.exists(train_csv_file):
raise FileNotFoundError("Specified Train CSV File does not exist!")
if not os.path.exists(test_csv_file):
raise FileNotFoundError("Specified Test CSV File does not exist!")
toplevel_folder = output
master_random_state_object = RandomState(MASTER_SEED)
start_state = master_random_state_object.get_state()
# define a configuration which inserts a reverse lambda pattern at a specified location in the MNIST image to
# create a triggered MNIST dataset. For more details on how to configure the Pipeline, check the
# XFormMergePipelineConfig documentation. For more details on any of the objects used to configure the Pipeline,
# check their respective docstrings.
one_channel_alpha_trigger_cfg = \
tdc.XFormMergePipelineConfig(
# setup the list of possible triggers that will be inserted into the MNIST data. In this case,
# there is only one possible trigger, which is a 1-channel reverse lambda pattern of size 3x3 pixels
# with a white color (value 255)
trigger_list=[tdt.ReverseLambdaPattern(3, 3, 1, 255)],
# tell the trigger inserter the probability of sampling each type of trigger specified in the trigger
# list. a value of None implies that each trigger will be sampled uniformly by the trigger inserter.
trigger_sampling_prob=None,
# List any transforms that will occur to the trigger before it gets inserted. In this case, we do none.
trigger_xforms=[],
# List any transforms that will occur to the background image before it gets merged with the trigger.
# Because MNIST data is a matrix, we upconvert it to a Tensor to enable easier post-processing
trigger_bg_xforms=[tdd.ToTensorXForm()],
# List how we merge the trigger and the background. Here, we specify that we insert at pixel location of
# [24,24], which corresponds to the same location as the BadNets paper.
trigger_bg_merge=tdi.InsertAtLocation(np.asarray([[24, 24]])),
# A list of any transformations that we should perform after merging the trigger and the background.
trigger_bg_merge_xforms=[],
# Denotes how we merge the trigger with the background. In this case, we insert the trigger into the
# image. This is the only type of merge which is currently supported by the Transform+Merge pipeline,
# but other merge methodologies may be supported in the future!
merge_type='insert',
# Specify that 15% of the clean data will be modified. Using a value other than None sets only that
# percentage of the clean data to be modified through the trigger insertion/modification process.
per_class_trigger_frac=0.25
)
############# Create the data ############
# create the clean data
clean_dataset_rootdir = os.path.join(toplevel_folder, 'mnist_clean')
master_random_state_object.set_state(start_state)
mnist.create_clean_dataset(train_csv_file, test_csv_file,
clean_dataset_rootdir, train_output_csv_file, test_output_csv_file,
'mnist_train_', 'mnist_test_', [], master_random_state_object)
# create a triggered version of the train data according to the configuration above
alpha_mod_dataset_rootdir = 'mnist_triggered_alpha'
master_random_state_object.set_state(start_state)
tdx.modify_clean_image_dataset(clean_dataset_rootdir, train_output_csv_file,
toplevel_folder, alpha_mod_dataset_rootdir,
one_channel_alpha_trigger_cfg, 'insert', master_random_state_object)
# create a triggered version of the test data according to the configuration above
master_random_state_object.set_state(start_state)
tdx.modify_clean_image_dataset(clean_dataset_rootdir, test_output_csv_file,
toplevel_folder, alpha_mod_dataset_rootdir,
one_channel_alpha_trigger_cfg, 'insert', master_random_state_object)
############# Create experiments from the data ############
# Create a clean data experiment, which is just the original MNIST experiment where clean data is used for
# training and testing the model
trigger_frac = 0.0
trigger_behavior = tdb.WrappedAdd(1, 10)
e = tde.ClassicExperiment(toplevel_folder, trigger_behavior)
train_df = e.create_experiment(os.path.join(toplevel_folder, 'mnist_clean', 'train_mnist.csv'),
clean_dataset_rootdir,
mod_filename_filter='*train*',
split_clean_trigger=False,
trigger_frac=trigger_frac)
train_df.to_csv(os.path.join(toplevel_folder, 'mnist_clean_experiment_train.csv'), index=None)
test_clean_df, test_triggered_df = e.create_experiment(os.path.join(toplevel_folder, 'mnist_clean',
'test_mnist.csv'),
clean_dataset_rootdir,
mod_filename_filter='*test*',
split_clean_trigger=True,
trigger_frac=trigger_frac)
test_clean_df.to_csv(os.path.join(toplevel_folder, 'mnist_clean_experiment_test_clean.csv'), index=None)
test_triggered_df.to_csv(os.path.join(toplevel_folder, 'mnist_clean_experiment_test_triggered.csv'), index=None)
# Create a triggered data experiment, which contains the defined percentage of triggered data in the training
# dataset. The remaining training data is clean data. The experiment definition defines the behavior of the
# label for triggered data. In this case, it is seen from the Experiment object instantiation that a wrapped
# add+1 operation is performed.
# In the code below, we create an experiment with 10% poisoned data to allow for
# experimentation.
trigger_frac = 0.2
train_df = e.create_experiment(os.path.join(toplevel_folder, 'mnist_clean', 'train_mnist.csv'),
os.path.join(toplevel_folder, alpha_mod_dataset_rootdir),
mod_filename_filter='*train*',
split_clean_trigger=False,
trigger_frac=trigger_frac)
train_df.to_csv(os.path.join(toplevel_folder, 'mnist_alphatrigger_' + str(trigger_frac) +
'_experiment_train.csv'), index=None)
test_clean_df, test_triggered_df = e.create_experiment(os.path.join(toplevel_folder,
'mnist_clean', 'test_mnist.csv'),
os.path.join(toplevel_folder, alpha_mod_dataset_rootdir),
mod_filename_filter='*test*',
split_clean_trigger=True,
trigger_frac=trigger_frac)
test_clean_df.to_csv(os.path.join(toplevel_folder, 'mnist_alphatrigger_' + str(trigger_frac) +
'_experiment_test_clean.csv'), index=None)
test_triggered_df.to_csv(os.path.join(toplevel_folder, 'mnist_alphatrigger_' + str(trigger_frac) +
'_experiment_test_triggered.csv'), index=None)
trigger_bg_merge_xforms=[],
# Denotes how we merge the trigger with the background. In this case, we insert the trigger into the
# image. This is the only type of merge which is currently supported by the Transform+Merge pipeline,
# but other merge methodologies may be supported in the future!
merge_type='insert',
# Specify that all the clean data will be modified. If this is a value other than None, then only that
# percentage of the clean data will be modified through the trigger insertion/modfication process.
per_class_trigger_frac=None
)
############# Create the data ############
# original MNIST - grayscale
clean_dataset_rootdir = os.path.join(toplevel_folder, 'mnist_clean')
master_random_state_object.set_state(start_state)
mnist.create_clean_dataset(train_csv_file, test_csv_file,
clean_dataset_rootdir, train_output_csv_file, test_output_csv_file,
'mnist_train_', 'mnist_test_', [], master_random_state_object)
# white alpha trigger w/ random rotation & location
alpha_mod_dataset_rootdir = 'mnist_triggered_alpha'
master_random_state_object.set_state(start_state)
tdx.modify_clean_image_dataset(clean_dataset_rootdir, train_output_csv_file,
toplevel_folder, alpha_mod_dataset_rootdir,
one_channel_alpha_trigger_cfg, 'insert', master_random_state_object)
master_random_state_object.set_state(start_state)
tdx.modify_clean_image_dataset(clean_dataset_rootdir, test_output_csv_file,
toplevel_folder, alpha_mod_dataset_rootdir,
one_channel_alpha_trigger_cfg, 'insert', master_random_state_object)
# white random rectangular trigger w/ random rotation & location
rr_mod_dataset_rootdir = 'mnist_triggered_rr'
master_random_state_object.set_state(start_state)
tdx.modify_clean_image_dataset(clean_dataset_rootdir, train_output_csv_file,