def test_battery(model, dataname):
if dataname == "cifar10":
data_source_dir = '../data/cifar10/'
classes = 10
elif dataname == "cifar100":
data_source_dir = '../data/cifar100/'
classes = 100
train_loader, valid_loader, unlabelled_loader, test_loader, num_classes = load_data_subset(0, 128, 2 ,dataname, data_source_dir, 0.0, labels_per_class = 5000)
#run_test_adversarial(True, model, test_loader, num_classes=num_classes, attack_type='pgd', attack_params={'eps':0.03, 'eps_iter': 0.01, 'iter': 200})
interpolation_tests(test_loader, model, classes)
#return
#adversarial_example_tests(test_loader, model, classes)
#interpolation_tests(test_loader, model, classes)
return
res_lst = []
shear_lst = [0.5, 1.0, 2.0, 2.5, 3.0]
for shearing in shear_lst:
deform_transform = [Shear(shearing)]
acc = deformation_tests(model, dataname, data_source_dir, deform_transform)
print "shearing", shearing, acc
#res_lst.append(acc)
print "shearing exp"
print shear_lst
print res_lst
res_lst = []
rot_lst = [20.0, 40.0, 60.0, 80.0]
for rotation in rot_lst:
deform_transform = [Rotation(rotation)]
acc = deformation_tests(model, dataname, data_source_dir, deform_transform)
print "rot", rotation, acc
#res_lst.append(acc)
print "rotation exp"
print rot_lst
print res_lst
res_lst = []
zoom_lst = [0.2, 0.4, 0.6, 0.8, 1.2, 1.4, 1.6, 1.8]
for zoom in zoom_lst:
deform_transform = [Zoom((zoom,zoom))]
acc = deformation_tests(model, dataname, data_source_dir, deform_transform)
print "zoom", zoom, acc
#res_lst.append(acc)
print "zoom exp"
print zoom_lst
print res_lst
def load_mnist(data_aug, batch_size, test_batch_size, cuda, data_target_dir):
if data_aug == 1:
hw_size = 24
transform_train = transforms.Compose([
transforms.RandomCrop(hw_size),
transforms.ToTensor(),
Rotation(15),
Zoom((0.85, 1.15)),
transforms.Normalize((0.1307, ), (0.3081, ))
])
transform_test = transforms.Compose([
transforms.CenterCrop(hw_size),
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
])
else:
hw_size = 28
transform_train = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
])
transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
])
kwargs = {'num_workers': 0, 'pin_memory': True} if cuda else {}
train_loader = torch.utils.data.DataLoader(datasets.MNIST(
data_target_dir, train=True, download=True, transform=transform_train),
batch_size=batch_size,
shuffle=True,
**kwargs)
test_loader = torch.utils.data.DataLoader(datasets.MNIST(
data_target_dir, train=False, transform=transform_test),
batch_size=test_batch_size,
shuffle=True,
**kwargs)
return train_loader, test_loader
def load_mnist_subset(data_aug,
batch_size,
test_batch_size,
cuda,
data_target_dir,
n_labels=10,
labels_per_class=100):
import numpy as np
from functools import reduce
from operator import __or__
from torch.utils.data.sampler import SubsetRandomSampler
from torchvision.datasets import MNIST
import torchvision.transforms as transforms
#from utils import onehot
if data_aug == 1:
hw_size = 24
transform_train = transforms.Compose([
transforms.RandomCrop(hw_size),
transforms.ToTensor(),
Rotation(15),
Zoom((0.85, 1.15)),
transforms.Normalize((0.1307, ), (0.3081, ))
])
transform_test = transforms.Compose([
transforms.CenterCrop(hw_size),
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
])
else:
hw_size = 28
transform_train = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(
(0.5, ), (0.5, )) #transforms.Normalize((0.1307,), (0.3081,))
])
transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(
(0.5, ), (0.5, )) #transforms.Normalize((0.1307,), (0.3081,))
])
kwargs = {'num_workers': 0, 'pin_memory': True} if cuda else {}
mnist_train = datasets.MNIST(data_target_dir,
train=True,
download=True,
transform=transform_train)
mnist_valid = datasets.MNIST(data_target_dir,
train=False,
transform=transform_test)
def get_sampler(labels, n=None):
# Only choose digits in n_labels
(indices, ) = np.where(
reduce(__or__, [labels == i for i in np.arange(n_labels)]))
# Ensure uniform distribution of labels
np.random.shuffle(indices)
indices = np.hstack([
list(filter(lambda idx: labels[idx] == i, indices))[:n]
for i in range(n_labels)
])
#print indices
indices = torch.from_numpy(indices)
sampler = SubsetRandomSampler(indices)
return sampler
#print type(mnist_train.train_labels)
# Dataloaders for MNIST
labelled = torch.utils.data.DataLoader(
mnist_train,
batch_size=batch_size,
sampler=get_sampler(mnist_train.train_labels.numpy(),
labels_per_class),
**kwargs)
unlabelled = torch.utils.data.DataLoader(
mnist_train,
batch_size=batch_size,
sampler=get_sampler(mnist_train.train_labels.numpy()),
**kwargs)
validation = torch.utils.data.DataLoader(
mnist_valid,
batch_size=test_batch_size,
sampler=get_sampler(mnist_valid.test_labels.numpy()),
**kwargs)
return labelled, unlabelled, validation