def setUp(self):
C = self.n_channels // self.groups
head_ndim = 2
self.link = links.DecorrelatedBatchNormalization(self.n_channels,
groups=self.groups,
dtype=self.dtype)
self.link.cleargrads()
shape = (5, self.n_channels) + (2, ) * self.ndim
self.x = numpy.random.uniform(-1, 1, shape).astype(self.dtype)
self.gy = numpy.random.uniform(-1, 1, shape).astype(self.dtype)
if self.test:
self.mean = numpy.random.uniform(-1, 1, (C, )).astype(self.dtype)
self.projection = numpy.random.uniform(0.5, 1,
(C, C)).astype(self.dtype)
self.link.avg_mean[...] = self.mean
self.link.avg_projection[...] = self.projection
else:
spatial_axis = tuple(range(head_ndim, self.x.ndim))
x_hat = self.x.reshape((5 * self.groups, C) + self.x.shape[2:])
x_hat = x_hat.transpose((1, 0) + spatial_axis).reshape((C, -1))
self.mean = x_hat.mean(axis=1)
self.projection = _calc_projection(self.x, self.mean,
self.link.eps, self.groups)
self.check_forward_options = {'atol': 1e-4, 'rtol': 1e-3}
self.check_backward_options = {
'atol': 5e-3,
'rtol': 1e-3,
'dtype': numpy.float64
}
if self.dtype == numpy.float32:
self.check_backward_options = {'atol': 5e-2, 'rtol': 5e-2}
def setUp(self):
C = self.n_channels // self.groups
self.link = links.DecorrelatedBatchNormalization(self.n_channels,
groups=self.groups,
dtype=self.dtype)
self.link.cleargrads()
shape = (5, self.n_channels) + (2, ) * self.ndim
self.x = numpy.random.uniform(-1, 1, shape).astype(self.dtype)
self.gy = numpy.random.uniform(-1, 1, shape).astype(self.dtype)
if self.test:
self.mean = numpy.random.uniform(-1, 1, (self.groups, C)).astype(
self.dtype)
self.projection = numpy.random.uniform(
0.5, 1, (self.groups, C, C)).astype(self.dtype)
self.link.avg_mean[...] = self.mean
self.link.avg_projection[...] = self.projection
else:
self.mean = _calc_mean(self.x, self.groups)
self.projection = _calc_projection(self.x, self.mean,
self.link.eps, self.groups)
self.check_forward_options = {'atol': 1e-4, 'rtol': 1e-3}
self.check_backward_options = {'atol': 5e-3, 'rtol': 1e-3}
if self.dtype == numpy.float32:
self.check_backward_options = {'atol': 5e-2, 'rtol': 5e-2}
def create_link(self, initializers):
mean, projection = initializers
link = links.DecorrelatedBatchNormalization(self.n_channels,
groups=self.groups,
dtype=self.dtype)
link.cleargrads()
if self.test:
link.avg_mean[...] = mean
link.avg_projection[...] = projection
return link
def __init__(self, link_name, finetune, forget=False):
super(Model, self).__init__()
with self.init_scope():
if link_name == 'bn':
self.link = links.BatchNormalization(3)
elif link_name == 'brn':
self.link = links.BatchRenormalization(3)
elif link_name == 'dbn':
self.link = links.DecorrelatedBatchNormalization(
3, groups=3)
self.forget = forget
self.finetune = finetune
def __init__(self, link_name, finetune, forget=False):
super(Model, self).__init__()
with self.init_scope():
if link_name == 'bn':
self.link = links.BatchNormalization(3)
elif link_name == 'brn':
# defaults rmax=1, dmax=0 are so trivial that BRN
# becomes BN
self.link = links.BatchRenormalization(3,
rmax=2.0,
dmax=1.0)
elif link_name == 'dbn':
self.link = links.DecorrelatedBatchNormalization(
3, groups=3)
self.forget = forget
self.finetune = finetune
def check_model_compatibility(self, backend_config, save, load):
C = self.n_channels // self.groups
old_model = {
'avg_mean':
numpy.random.uniform(-1, 1, (C, )).astype(self.dtype),
'avg_projection':
numpy.random.uniform(0.5, 1, (C, C)).astype(self.dtype),
'N':
numpy.array(0)
}
save(self.temp_file_path, old_model)
model = links.DecorrelatedBatchNormalization(self.n_channels,
groups=self.groups,
dtype=self.dtype)
model.to_device(backend_config.device)
with (testing.assert_warns(UserWarning)
if self.groups != 1 else nullcontext()):
load(self.temp_file_path, model)
x = numpy.random.rand(5, self.n_channels, 2).astype(self.dtype)
x = backend_config.get_array(x)
with chainer.using_config('train', False):
model(x)
model(x)