def main():
import random
import numpy as np
import torch
import torch.nn as nn
print(np.random.rand(10))
print(random.random())
print(nn.Conv2d(2, 2, 3, 1, 0).weight.data)
print(torch.backends.cudnn.deterministic)
print(torch.backends.cudnn.benchmark)
print('------')
with fixed_seed(0, True):
print(np.random.rand(10))
print(random.random())
print(nn.Conv2d(2, 2, 3, 1, 0).weight.data)
print(torch.backends.cudnn.deterministic)
print(torch.backends.cudnn.benchmark)
print(np.random.rand(10))
print(random.random())
print(nn.Conv2d(2, 2, 3, 1, 0).weight.data)
print(torch.backends.cudnn.deterministic)
print(torch.backends.cudnn.benchmark)
def main():
parser = argparse.ArgumentParser(
description='Example: Uncertainty estimates in segmentation',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--data_root',
'-d',
type=str,
default='./preprocessed',
help='Directory to dataset')
parser.add_argument('--batchsize',
'-b',
type=int,
default=2,
help='Number of images in each mini-batch')
parser.add_argument('--iteration',
'-i',
type=int,
default=50000,
help='Number of sweeps over the dataset to train')
parser.add_argument('--frequency',
'-f',
type=int,
default=-1,
help='Frequency of taking a snapshot')
parser.add_argument('--gpu',
'-g',
type=str,
default='cuda:0',
help='GPU Device')
parser.add_argument('--out',
'-o',
default='logs',
help='Directory to output the result')
parser.add_argument('--resume',
'-r',
default='',
help='Resume the training from snapshot')
parser.add_argument('--valid_augment',
action='store_true',
help='Enable data augmentation during validation')
parser.add_argument('--valid_split_ratio',
type=float,
default=0.1,
help='Ratio of validation data to training data')
parser.add_argument(
'--valid_split_type',
type=str,
default='slice',
choices=['slice', 'patient'],
help='How to choice validation data from training data')
parser.add_argument('--lr', type=float, default=1e-4, help='Learning rate')
parser.add_argument('--decay',
type=float,
default=-1,
help='Weight of L2 regularization')
parser.add_argument('--mc_iteration',
type=int,
default=15,
help='Number of iteration of MCMC')
parser.add_argument('--pinfall',
type=int,
default=-1,
help='Countdown for early stopping of training.')
parser.add_argument(
'--freeze_upconv',
action='store_true',
help=
'Disables updating the up-convolutional weights. If weights are initialized with \
bilinear kernels, up-conv acts as bilinear upsampler.'
)
parser.add_argument('--test_on_test',
action='store_true',
help='Switch to the testing phase on test dataset')
parser.add_argument('--test_on_valid',
action='store_true',
help='Switch to the testing phase on valid dataset')
parser.add_argument('--seed',
type=int,
default=0,
help='Fix the random seed')
args = parser.parse_args()
print('GPU: {}'.format(args.gpu))
print('# Minibatch-size: {}'.format(args.batchsize))
print('')
# setup output directory
os.makedirs(args.out, exist_ok=True)
# NOTE: ad-hoc
# setup a normalizer
normalizer = Normalizer()
normalizer.add(Clip2D('minmax'))
normalizer.add(Subtract2D(0.5))
normalizer.add(Divide2D(1. / 255.))
# setup an augmentor
augmentor = DataAugmentor(n_dim=2)
augmentor.add(Flip2D(axis=2))
augmentor.add(
Affine2D(rotation=15.,
translate=(10., 10.),
shear=0.25,
zoom=(0.8, 1.2),
keep_aspect_ratio=True,
fill_mode=('reflect', 'reflect'),
cval=(0., 0.),
interp_order=(1, 0)))
with fixed_seed(args.seed, strict=False):
# setup a predictor
conv_param = { # NOTE: you can change layer type if you want..
'name':'conv',
'kernel_size': 3,
'stride': 1,
'padding': 2,
'padding_mode': 'reflect',
'dilation': 2,
'initialW': {'name': 'he_normal'},
'initial_bias': {'name': 'zero'},
}
upconv_param = { # NOTE: you can change layer type if you want..
'name':'deconv',
'kernel_size': 3,
'stride': 2,
'padding': 0,
'initialW': {'name': 'bilinear'},
'initial_bias': {'name': 'zero'},
}
norm_param = {'name': 'batch'}
predictor = BayesianUNet(ndim=2,
in_channels=3,
out_channels=2,
nlayer=4,
nfilter=32,
conv_param=conv_param,
upconv_param=upconv_param,
norm_param=norm_param)
if args.freeze_upconv:
predictor.freeze_layers(name='upconv',
recursive=True,
verbose=True)
# setup dataset
train, valid, test = get_dataset(args.data_root, args.valid_split_type,
args.valid_split_ratio,
args.valid_augment, normalizer,
augmentor)
# run
if args.test_on_test:
test_phase(predictor, test, args)
elif args.test_on_valid:
test_phase(predictor, valid, args)
else:
save_args(args, args.out)
predictor.save_args(os.path.join(args.out, 'model.json'))
normalizer.summary(os.path.join(args.out, 'norm.json'))
augmentor.summary(os.path.join(args.out, 'augment.json'))
train_phase(predictor, train, valid, args)
def main():
parser = argparse.ArgumentParser(
description='Example: Uncertainty estimates in regression')
parser.add_argument('--batchsize',
'-b',
type=int,
default=50,
help='Number of images in each mini-batch')
parser.add_argument('--epoch',
'-e',
type=int,
default=300,
help='Number of sweeps over the dataset to train')
parser.add_argument('--frequency',
'-f',
type=int,
default=-1,
help='Frequency of taking a snapshot')
parser.add_argument('--gpu',
'-g',
type=str,
default='cuda:0',
help='GPU Device')
parser.add_argument('--out',
'-o',
default='logs',
help='Directory to output the log files')
parser.add_argument('--resume',
'-r',
default='',
help='Resume the training from snapshot')
parser.add_argument('--unit',
'-u',
type=int,
default=20,
help='Number of units')
parser.add_argument('--noplot',
dest='plot',
action='store_false',
help='Disable PlotReport extension')
parser.add_argument('--test_on_test',
action='store_true',
help='Switch to the testing phase on test dataset')
parser.add_argument('--test_on_valid',
action='store_true',
help='Switch to the testing phase on valid dataset')
parser.add_argument('--mc_iteration',
type=int,
default=50,
help='Number of iteration of MCMC')
parser.add_argument('--decay',
type=float,
default=-1,
help='Weight of L2 regularization')
parser.add_argument('--seed',
type=int,
default=0,
help='Fix the random seed')
args = parser.parse_args()
os.makedirs(args.out, exist_ok=True)
with fixed_seed(args.seed, strict=False):
# setup a predictor
predictor = BayesianMLP(n_in=1,
n_units=args.unit,
n_out=1,
drop_ratio=0.1)
# setup dataset
train = Dataset(x_lim=(-5, 5), n_samples=1000)
valid = Dataset(x_lim=(-5, 5), n_samples=1000)
test = Dataset(x_lim=(-10, 10), n_samples=500)
# run
if args.test_on_test:
test_phase(predictor, test, args)
elif args.test_on_valid:
test_phase(predictor, valid, args)
else:
train_phase(predictor, train, valid, args)
def main():
parser = argparse.ArgumentParser(description='Example: Uncertainty estimates with adversarial training in image synthesis',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--data_root', '-d', type=str, default='./preprocessed',
help='Directory to dataset')
parser.add_argument('--batchsize', '-b', type=int, default=5,
help='Number of images in each mini-batch')
parser.add_argument('--iteration', '-i', type=int, default=200000,
help='Number of sweeps over the dataset to train')
parser.add_argument('--frequency', '-f', type=int, default=-1,
help='Frequency of taking a snapshot')
parser.add_argument('--gpu', '-g', type=str, default='cuda:0',
help='GPU Device')
parser.add_argument('--out', '-o', default='logs',
help='Directory to output the result')
parser.add_argument('--resume', '-r', default='',
help='Resume the training from snapshot')
parser.add_argument('--valid_augment', action='store_true',
help='Enable data augmentation during validation')
parser.add_argument('--valid_split_ratio', type=float, default=0.1,
help='Ratio of validation data to training data')
parser.add_argument('--lr', type=float, default=4e-4,
help='Learning rate')
parser.add_argument('--alpha', type=float, default=50.,
help='Weight of conditional loss')
parser.add_argument('--beta', type=float, default=0.5,
help='Exponential decay rate of the first order moment in Adam')
parser.add_argument('--decay', type=float, default=-1,
help='Weight of L2 regularization')
parser.add_argument('--mc_iteration', type=int, default=15,
help='Number of iteration of MCMC')
parser.add_argument('--pinfall', type=int, default=-1,
help='Countdown for early stopping of training.')
parser.add_argument('--freeze_upconv', action='store_true',
help='Disables updating the up-convolutional weights. If weights are initialized with \
bilinear kernels, up-conv acts as bilinear upsampler.')
parser.add_argument('--test_on_test', action='store_true',
help='Switch to the testing phase on test dataset')
parser.add_argument('--test_on_valid', action='store_true',
help='Switch to the testing phase on valid dataset')
parser.add_argument('--seed', type=int, default=0,
help='Fix the random seed')
args = parser.parse_args()
print('GPU: {}'.format(args.gpu))
print('# Minibatch-size: {}'.format(args.batchsize))
print('')
# setup output directory
os.makedirs(args.out, exist_ok=True)
# NOTE: ad-hoc
normalizer = get_normalizer()
augmentor = get_augmentor()
# setup a generator
with fixed_seed(args.seed, strict=False):
generator = build_generator()
if args.freeze_upconv:
generator.freeze_layers(name='upconv',
recursive=True,
verbose=True)
# setup dataset
train, valid, test = get_dataset(args.data_root,
args.valid_split_ratio,
args.valid_augment,
normalizer, augmentor)
# run
if args.test_on_test:
raise RuntimeError('This example is under construction. Please tune the hyperparameters first..')
test_phase(generator, test, args)
elif args.test_on_valid:
test_phase(generator, valid, args)
else:
save_args(args, args.out)
generator.save_args(os.path.join(args.out, 'model.json'))
normalizer.summary(os.path.join(args.out, 'norm.json'))
augmentor.summary(os.path.join(args.out, 'augment.json'))
train_phase(generator, train, valid, args)