k = float(sys.argv[1]) if len(sys.argv) > 1 else 0
dataset = str(sys.argv[2]) if len(sys.argv) > 2 else 'mnist'
iparam = str(sys.argv[3]) if len(sys.argv) > 3 else None
print('k = ', k, 'dataset = ', dataset, 'params = ', iparam)
num_epochs, batch_size, verbose = 200, 100, 1
optpol = lambda epoch: optpolicy.lr_linear_to0(epoch, 1e-3)
arch = net_lenet5
net = run_experiment(dataset,
0,
batch_size,
arch,
objectives.sgvlb,
False,
optpol,
optpolicy.rw_linear,
optimizer='adam')
paramsv = build_params_from_init(net, iparam, lsinit=-10) if iparam else None
net = run_experiment(dataset,
num_epochs,
batch_size,
arch,
objectives.sgvlb,
verbose,
optpol,
optpolicy.rw_linear,
input_x, target_y, Winit = T.tensor4("input"), T.vector(
"target", dtype='int32'), init.Normal()
net = ll.InputLayer(input_shape, input_x)
net = ConvLayer(net, 20, 5, W=init.Normal())
net = MaxPool2DLayer(net, 2)
net = ConvLayer(net, 50, 5, W=init.Normal())
net = MaxPool2DLayer(net, 2)
net = ll.DenseLayer(net, 500, W=init.Normal())
net = ll.DenseLayer(net, nclass, W=init.Normal(), nonlinearity=nl.softmax)
return net, input_x, target_y, 1
num_epochs, batch_size, verbose, dataset = 200, 100, 1, 'mnist'
optp = lambda epoch: optpolicy.lr_linear(epoch, 1e-4)
arch = net_lenet5
net = run_experiment(dataset,
num_epochs,
batch_size,
arch,
objectives.sgvlb,
verbose,
optp,
optpolicy.rw_linear,
optimizer='adam',
da=True)
if not os.path.exists('./experiments/logs'):
os.mkdir('./experiments/logs')
if not os.path.exists('./experiments/logs/' + folder_name):
os.mkdir('./experiments/logs/' + folder_name)
for trainset_size in trainset_sizes:
# for alpha in alphas:
for magn_var in magn_vars:
log_fname = folder_name + '/' + filename + '-' + str(
trainset_size) + '-' + str(alpha) + '-' + str(magn_var)
run_experiment(
dataset,
num_epochs,
batch_size,
arch,
criterion,
verbose,
optpol_linear,
params=None,
optimizer='adam',
trainset_size=trainset_size,
log_fname=log_fname,
noise_type=noise_type,
alpha=alpha,
noise_magnitude=noise_magnitude,
magn_var=magn_var,
noise_ave_times=ave_times,
)
int(512 * k),
W=init.Normal(),
nonlinearity=nl.rectify)
net = BatchNormLayer(net, epsilon=1e-3)
net = ll.NonlinearityLayer(net)
net = ll.DropoutLayer(net, 0.5)
net = ll.DenseLayer(net, nclass, W=init.Normal(), nonlinearity=nl.softmax)
return net, input_x, target_y, k
k = float(sys.argv[1]) if len(sys.argv) > 1 else 1.0
dataset = str(sys.argv[2]) if len(sys.argv) > 2 else 'cifar10'
iparam = str(sys.argv[3]) if len(sys.argv) > 3 else None
averaging = int(sys.argv[4]) if len(sys.argv) > 4 else 0
print('k = ', k, 'dataset = ', dataset, 'params = ', iparam)
num_epochs, batch_size, verbose = 200, 100, 1
optpol = lambda epoch: optpolicy.lr_linear(epoch, 1e-5)
arch = lambda input_shape, s: net_vgglike(k, input_shape, s)
net = run_experiment(dataset,
num_epochs,
batch_size,
arch,
objectives.nll_l2,
verbose,
optpol,
optpolicy.rw_linear,
optimizer='adam')
folder_name = 'wn_tangent'
filename = 'wn_tang'
trainset_sizes = [100]
alphas = [0.01]
ave_times = 0
if not os.path.exists('./experiments/logs/' + folder_name):
os.mkdir('./experiments/logs/' + folder_name)
for trainset_size in trainset_sizes:
for alpha in alphas:
log_fname = folder_name + '/' + filename + '-' + str(
trainset_size) + '-' + str(alpha)
net = run_experiment(
dataset,
num_epochs,
batch_size,
arch,
objectives.nll,
verbose,
optpol_linear,
optpolicy.rw_linear,
params=None,
optimizer='adam',
train_clip=False,
trainset_size=trainset_size,
log_fname=log_fname,
alpha=alpha,
noise_ave_times=ave_times,
)