def main():
parser = argparse.ArgumentParser(
description='check if CROWN verifies given alpha',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('crown_path')
parser.add_argument('model', help='.npy model generated by mat2npy.py')
parser.add_argument('inputs',
help='input files generated by model2chk.py',
nargs='+')
args = parser.parse_args()
sys.path.append(args.crown_path)
from get_bounds_ours import compute_worst_bound, get_weights_list
import save_nlayer_weights as nl
model = nl.NLayerModel(None, args.model, activation='relu')
for inp_name in args.inputs:
with open(inp_name, 'rb') as fin:
rec = pickle.load(fin)
orig_inp = rec['input']
orig_inp = np.transpose(orig_inp, (1, 2, 0))
pred = np.squeeze(model.model.predict(orig_inp[None]), 0)
assert np.argmax(pred) == rec['label']
weights, biases = get_weights_list(model)
def get_bound(inp):
gap_gx, g_x0, max_grad_norm = compute_worst_bound(weights,
biases,
rec['label'],
rec['label'],
inp,
pred,
None,
eps=rec['eps'],
untargeted=True)
return gap_gx
b0 = get_bound(orig_inp)
b1 = get_bound(orig_inp * rec['alpha'])
print('file={} label={} gap0={} gap1={}'.format(
inp_name, rec['label'], b0, b1))
args.numlayer) + "layer_relu_" + str(
nhidden) + "_best" + suffix
if not os.path.isfile(modelfile):
modelfile = args.filename
if not os.path.isfile(modelfile):
raise (RuntimeError("cannot find model file"))
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
if args.model == "mnist":
data = MNIST()
if args.cnnmodel:
model = nl.CNNModel(modelfile)
elif args.filename:
model = nl.NLayerModel(args.layers, modelfile)
else:
model = nl.NLayerModel([nhidden] * (args.numlayer - 1),
modelfile)
elif args.model == "cifar":
data = CIFAR()
if args.cnnmodel:
model = nl.CNNModel(modelfile)
elif args.filename:
model = nl.NLayerModel(args.layers,
modelfile,
image_size=32,
image_channel=3)
else:
model = nl.NLayerModel([nhidden] * (args.numlayer - 1),
modelfile,
modelfile = "models/" + args.model + "_" + str(args.numlayer) + "layer_" + activation + "_" + suffix
# if still not found, try models/mnist_3layer_relu_1024_best
if not os.path.isfile(modelfile):
modelfile = "models/" + args.model + "_" + str(args.numlayer) + "layer_" + activation + "_" + str(nhidden) + suffix + "_best"
if not os.path.isfile(modelfile):
raise(RuntimeError("cannot find model file"))
if args.LP or args.LPFULL:
# use gurobi solver
import gurobipy as grb
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
if args.model == "mnist":
data = MNIST()
model = nl.NLayerModel([nhidden] * (args.numlayer - 1), modelfile, activation=activation)
elif args.model == "cifar":
data = CIFAR()
model = nl.NLayerModel([nhidden] * (args.numlayer - 1), modelfile, image_size=32, image_channel=3, activation=activation)
else:
raise(RuntimeError("unknown model: "+args.model))
print("Evaluating", modelfile)
sys.stdout.flush()
random.seed(1215)
np.random.seed(1215)
tf.set_random_seed(1215)
# the weights and bias are saved in lists: weights and bias
# weights[i-1] gives the ith layer of weight and so on
