inception_model = BinaryLogisticRegressionWithLBFGS(
input_dim=input_dim,
weight_decay=weight_decay,
max_lbfgs_iter=max_lbfgs_iter,
num_classes=num_classes,
batch_size=batch_size,
data_sets=data_sets,
initial_learning_rate=initial_learning_rate,
keep_probs=keep_probs,
decay_epochs=decay_epochs,
mini_batch=False,
train_dir='output',
log_dir='log',
model_name='%s_inception_onlytop' % dataset_name)
inception_model.train()
inception_predicted_loss_diffs = inception_model.get_influence_on_test_loss(
[test_idx],
np.arange(len(inception_model.data_sets.train.labels)),
force_refresh=True)
x_test = X_test[test_idx, :]
y_test = Y_test[test_idx]
distances = dataset.find_distances(x_test, X_train)
flipped_idx = Y_train != y_test
rbf_margins_test = rbf_model.sess.run(rbf_model.margin,
feed_dict=rbf_model.all_test_feed_dict)
rbf_margins_train = rbf_model.sess.run(rbf_model.margin,
feed_dict=rbf_model.all_train_feed_dict)
input_dim = 2048
top_model = BinaryLogisticRegressionWithLBFGS(
input_dim=input_dim,
weight_decay=weight_decay,
max_lbfgs_iter=max_lbfgs_iter,
num_classes=num_classes,
batch_size=batch_size,
data_sets=inception_data_sets,
initial_learning_rate=initial_learning_rate,
keep_probs=keep_probs,
decay_epochs=decay_epochs,
mini_batch=False,
train_dir='output',
log_dir='log',
model_name=top_model_name)
top_model.train()
weights = top_model.sess.run(top_model.weights)
orig_weight_path = 'output/inception_weights_%s.npy' % top_model_name
np.save(orig_weight_path, weights)
with full_graph.as_default():
full_model.load_weights_from_disk(orig_weight_path,
do_save=False,
do_check=True)
full_model.reset_datasets()
### Create poisoned dataset
print('Creating poisoned dataset...')
# First pass was with step_size 0.02, 100 iterations, num_to_poison=2
# Second pass with step_size 0.01 and 200 iterations, num_to_poison=3
tf_model = BinaryLogisticRegressionWithLBFGS(
input_dim=input_dim,
weight_decay=weight_decay,
max_lbfgs_iter=max_lbfgs_iter,
num_classes=num_classes,
batch_size=batch_size,
data_sets=data_sets,
initial_learning_rate=initial_learning_rate,
keep_probs=keep_probs,
decay_epochs=decay_epochs,
mini_batch=False,
train_dir='output',
log_dir='log',
model_name='spam_logreg_lbfgs')
tf_model.train()
test_idx = 8
actual_loss_diffs, predicted_loss_diffs_cg, indices_to_remove = experiments.test_retraining(
tf_model,
test_idx,
iter_to_load=0,
force_refresh=False,
num_to_remove=500,
remove_type='maxinf',
random_seed=0)
# LiSSA
np.random.seed(17)
predicted_loss_diffs_lissa = tf_model.get_influence_on_test_loss(
[test_idx],
def run_spam(ex_to_leave_out=None, num_examples=None):
"""
If ex_to_leave_out is None, don't leave any out. Otherwise, leave out the example at the specified index.
If num_examples is None, use all the examples
"""
data_sets = load_spam(ex_to_leave_out=ex_to_leave_out,
num_examples=num_examples)
# "Spam" and "Ham"
num_classes = 2
input_dim = data_sets.train.x.shape[1]
weight_decay = 0.0001
# weight_decay = 1000 / len(lr_data_sets.train.labels)
batch_size = 100
initial_learning_rate = 0.001
keep_probs = None
decay_epochs = [1000, 10000]
max_lbfgs_iter = 1000
tf.reset_default_graph()
tf_model = BinaryLogisticRegressionWithLBFGS(
input_dim=input_dim,
weight_decay=weight_decay,
max_lbfgs_iter=max_lbfgs_iter,
num_classes=num_classes,
batch_size=batch_size,
data_sets=data_sets,
initial_learning_rate=initial_learning_rate,
keep_probs=keep_probs,
decay_epochs=decay_epochs,
mini_batch=False,
train_dir='output',
log_dir='log',
model_name='spam_logreg')
tf_model.train()
# NMV 7/26: appears to be unused right now.
# X_train = np.copy(tf_model.data_sets.train.x)
# Y_train = np.copy(tf_model.data_sets.train.labels)
# X_test = np.copy(tf_model.data_sets.test.x)
# Y_test = np.copy(tf_model.data_sets.test.labels)
# num_train_examples = Y_train.shape[0]
# num_flip_vals = 6
# num_check_vals = 6
# num_random_seeds = 40
# dims = (num_flip_vals, num_check_vals, num_random_seeds, 3)
# fixed_influence_loo_results = np.zeros(dims)
# fixed_loss_results = np.zeros(dims)
# fixed_random_results = np.zeros(dims)
#flipped_results = np.zeros((num_flip_vals, num_random_seeds, 3))
orig_results = tf_model.sess.run(
[tf_model.loss_no_reg, tf_model.accuracy_op],
feed_dict=tf_model.all_test_feed_dict)
#print('Orig loss: %.5f. Accuracy: %.3f' % (orig_results[0], orig_results[1]))
result = [tf_model, orig_results]
return result