print("Initializing model...") top = 70 epss = [ int(i**2 + 0.5) / 100 for i in np.arange(0, int(top**0.5 + 1) + 0.5, 0.5) ] print(epss) accuracieses = [] for parseval in [False, True]: aggregation = 'convex' if parseval else 'sum' resnet_ctor = ParsevalResNet if parseval else ResNet from standard_resnets import get_wrn model = standard_resnets.get_wrn(zaggydepth, k, ds_test.image_shape, ds_test.class_count, aggregation=aggregation, resnet_ctor=resnet_ctor) saved_path = dirs.SAVED_MODELS if parseval: saved_path += '/wrn-28-10-p-t--2018-01-24-21-18/ResNet' # Parseval else: saved_path += '/wrn-28-10-t--2018-01-23-19-13/ResNet' # vanilla model.load_state(saved_path) cost, ev = model.test(ds_test) accuracies = [ev['accuracy']] for eps in epss[1:]: print("Creating adversarial examples...") clip_max = (255 - np.max(Cifar10Loader.mean)) / np.max(
import sys import datetime from models import ResidualBlockProperties, ResNet from data_utils import Cifar10Loader import standard_resnets from training import train import dirs dimargs = sys.argv[1:] if len(dimargs) not in [0, 2]: print("usage: train-wrn.py [<Zagoruyko-depth> <widening-factor>]") zaggydepth, k = (16, 4) if len(dimargs) == 0 else map(int, dimargs) print("Loading and preparing data...") ds_train, ds_val = Cifar10Loader.load_train_val() print("Initializing model...") model = standard_resnets.get_wrn( zaggydepth, k, ds_train.image_shape, ds_train.class_count) print("Starting training and validation loop...") train(model, ds_train, ds_val, epoch_count=200) print("Saving model...") model.save_state(dirs.SAVED_MODELS + '/wrn-%d-%d--' % (zaggydepth, k) + datetime.datetime.now().strftime("%Y-%m-%d-%H-%M"))
import datetime from models import ResidualBlockProperties, ResNetN from data_utils import Cifar10Loader, Dataset import standard_resnets from training import train import dirs dimargs = sys.argv[1:] if len(dimargs) not in [0, 2]: print("usage: train-wrn.py [<Zagoruyko-depth> <widening-factor>]") zaggydepth, k = (16, 4) if len(dimargs) == 0 else map(int, dimargs) print("Loading and preparing data...") ds_train, ds_val = Cifar10Loader.load_train_val() print("Initializing model...") model = standard_resnets.get_wrn(zaggydepth, k, ds_train.image_shape, ds_train.class_count, aggregation='convex', resnet_ctor=ResNetN) print("Starting training and validation loop...") train(model, ds_train, ds_val, epoch_count=200) print("Saving model...") model.save_state(dirs.SAVED_MODELS + '/wrn-%d-%d-nk--' % (zaggydepth, k) + datetime.datetime.now().strftime("%Y-%m-%d-%H-%M"))