def adv_train(x, y):
n_adv = int(hps.ae_frac * hps.batch_size)
if n_adv == 0:
return x
else:
adv_inputs = ae.pgd_attack(x[:n_adv], y[:n_adv], cleverhans_model, hps.p,
eps_dict[hps.p][hps.dataset], hps.pgd_n_iter)
return tf.concat([adv_inputs, x[n_adv:]], axis=0)
name='train_step')
# Separate forward pass graph for Cleverhans wrapper (for PGD attack) placed on the last GPU
logits_all_gpus = forward_pass_cleverhans(x_tf)
# Model saver
saver = tf.train.Saver()
# GPU settings
gpu_options = tf.GPUOptions(visible_device_list=str(hps.gpus)[1:-1],
per_process_gpu_memory_fraction=hps.gpu_memory)
config = tf.ConfigProto(allow_soft_placement=True, gpu_options=gpu_options)
with tf.Session(graph=graph, config=config) as sess:
with graph.as_default(), tf.device('/gpu:0'):
pgd_ae_tensor = ae.pgd_attack(x_tf, y_in, cleverhans_model, hps.p,
eps_dict[hps.p][hps.dataset],
hps.pgd_n_iter)
sess.run(tf.global_variables_initializer()) # run 'init' op
epoch_start, epoch_end = 0, hps.n_epochs
log.add('Session started with hyperparameters: {} \n'.format(hps_str))
time_start = time.time()
# for epoch in range(epoch_start, epoch_end):
for epoch in range(1, epoch_end + 1):
epoch_start_reduced_lr = 0.9
lr_actual = hps.lr / 10 if epoch >= epoch_start_reduced_lr * hps.n_epochs else hps.lr
frac_reg = min(epoch / 10.0,
1.0) # from 0 to 1 linearly over the first 10 epochs
per_process_gpu_memory_fraction=hps.gpu_memory)
config = tf.ConfigProto(allow_soft_placement=True, gpu_options=gpu_options)
# --------- Pytorch part for the Kolter-Wong model ----------
device = torch.device('cuda:' + str(hps.gpus[-1]))
torch.cuda.set_device(hps.gpus[-1])
model_torch = kolter_wong.models.select_model(hps.nn_type, hps.n_in,
hps.n_out).to(device)
for var in model_torch.parameters():
var.requires_grad = False
# ---------- end ----------
with tf.Session(graph=graph, config=config) as sess:
with graph.as_default(), tf.device('/gpu:0'):
pgd_ae_tensor = ae.pgd_attack(x_tf, y_in, cleverhans_model, hps.p, eps,
hps.pgd_n_iter)
load_model(
sess,
hps) # load the weights from hps.model_path to the current TF model
time_start = time.time()
# First on the full test set
clean_inputs_all, clean_labels_all = x_test, y_test
logits_val = sess.run(logits,
feed_dict={
x_in: clean_inputs_all,
is_train: False
})
correctly_classified = np.argmax(clean_labels_all,
axis=1) == np.argmax(logits_val, axis=1)
