def test_random_shift_forward_backward(seed, inshape, shifts, border_mode,
constant_value, ctx, func_name):
from nbla_test_utils import function_tester
rng = np.random.RandomState(seed)
inputs = [rng.randn(*inshape).astype(np.float32)]
i = nn.Variable(inputs[0].shape, need_grad=True)
i.d = inputs[0]
# NNabla forward
with nn.context_scope(ctx), nn.auto_forward():
o = F.random_shift(i, shifts, border_mode, constant_value, 0, seed)
result_shifts = (0, 0, 0)
max_correl = 0
for shift_amount in itertools.product(*map(
tuple,
map(lambda x: range(*x), [(-2, 3) for _ in range(len(inshape))]))):
r = scipy_shift(inputs[0],
shift_amount,
mode=border_mode,
cval=constant_value)
correl_and_p = pearsonr(o.d.flatten(), r.flatten())
if correl_and_p[0] > max_correl:
result_shifts = shift_amount
max_correl = correl_and_p[0]
ref = scipy_shift(inputs[0],
result_shifts,
mode=border_mode,
cval=constant_value)
if shifts is None:
shifts = (0, ) * len(inputs[0].shape)
for result, shift_range in zip(result_shifts, shifts):
assert abs(result) <= shift_range
assert_allclose(o.d, ref)
assert o.parent.name == func_name
# Skipping Backward check
g = np.random.randn(*i.shape)
i.g = g
o_grad = np.random.randn(*o.shape)
o.g = o_grad
o.parent.backward([i], [o])
ref_grad = i.g.copy() - g
# Check accum=False with NaN gradient
i.g = np.float32('nan')
o.parent.backward([i], [o], [False])
assert not np.any(np.isnan(i.g))
# Check if accum option works
i.g[...] = 1
o.g = o_grad
o.parent.backward([i], [o], [False])
assert_allclose(i.g, ref_grad, atol=1e-6)
# Check if need_grad works
i.g[...] = 0
i.need_grad = False
o_grad = rng.randn(*i.shape).astype(i.data.dtype)
o.backward(o_grad)
assert np.all(i.g == 0)
def vectorizer(x, maxh=256, test=False, output_hidden=False):
"""
Building discriminator network which maps a (B, 1, 28, 28) input to
a (B, 100).
"""
# Define shortcut functions
def bn(xx):
# Batch normalization
return PF.batch_normalization(xx, batch_stat=not test)
def downsample2(xx, c):
return PF.convolution(xx,
c, (3, 3),
pad=(1, 1),
stride=(2, 2),
with_bias=False)
assert maxh / 8 > 0
with nn.parameter_scope("dis"):
# (1, 28, 28) --> (32, 16, 16)
if not test:
x_ = F.image_augmentation(x, min_scale=0.9, max_scale=1.08)
x2 = F.random_shift(x_, (2, 2))
with nn.parameter_scope("conv1"):
c1 = F.elu(
bn(
PF.convolution(x2,
maxh / 8, (3, 3),
pad=(3, 3),
stride=(2, 2),
with_bias=False)))
else:
with nn.parameter_scope("conv1"):
c1 = F.elu(
bn(
PF.convolution(x,
maxh / 8, (3, 3),
pad=(3, 3),
stride=(2, 2),
with_bias=False)))
# (32, 16, 16) --> (64, 8, 8)
with nn.parameter_scope("conv2"):
c2 = F.elu(bn(downsample2(c1, maxh / 4)))
# (64, 8, 8) --> (128, 4, 4)
with nn.parameter_scope("conv3"):
c3 = F.elu(bn(downsample2(c2, maxh / 2)))
# (128, 4, 4) --> (256, 4, 4)
with nn.parameter_scope("conv4"):
c4 = bn(
PF.convolution(c3, maxh, (3, 3), pad=(1, 1), with_bias=False))
# (256, 4, 4) --> (1,)
with nn.parameter_scope("fc1"):
#print "c4fdafafa",c4.shape
#f = PF.affine(c4, 100)
f = PF.convolution(c4, 1000, (4, 4), pad=(0, 0), with_bias=False)
if output_hidden:
return f, [c1, c2, c3, c4]
return f
def test_random_shift_forward_backward(seed, inshape, shifts, border_mode, ctx, func_name):
from nbla_test_utils import function_tester
rng = np.random.RandomState(seed)
inputs = [rng.randn(*inshape).astype(np.float32)]
i = nn.Variable(inputs[0].shape, need_grad=True)
i.d = inputs[0]
# NNabla forward
with nn.context_scope(ctx), nn.auto_forward():
o = F.random_shift(i, shifts, border_mode, 0, seed)
result_shifts = (0, 0, 0)
max_correl = 0
for shift_amount in itertools.product(*map(tuple, map(lambda x: range(*x), [(-2, 3) for _ in range(len(inshape))]))):
r = scipy_shift(inputs[0], shift_amount, mode=border_mode)
correl_and_p = pearsonr(o.d.flatten(), r.flatten())
if correl_and_p[0] > max_correl:
result_shifts = shift_amount
max_correl = correl_and_p[0]
ref = scipy_shift(inputs[0], result_shifts, mode=border_mode)
if shifts is None:
shifts = (0,) * len(inputs[0].shape)
for result, shift_range in zip(result_shifts, shifts):
assert abs(result) <= shift_range
assert np.allclose(o.d, ref)
assert o.parent.name == func_name
# Skipping Backward check
g = np.random.randn(*i.shape)
i.g = g
o_grad = np.random.randn(*o.shape)
o.g = o_grad
o.parent.backward([i], [o])
ref_grad = i.g.copy() - g
# Check accum=False with NaN gradient
i.g = np.float32('nan')
o.parent.backward([i], [o], [False])
assert not np.any(np.isnan(i.g))
# Check if accum option works
i.g[...] = 1
o.g = o_grad
o.parent.backward([i], [o], [False])
assert np.allclose(i.g, ref_grad, atol=1e-6)
# Check if need_grad works
i.g[...] = 0
i.need_grad = False
o_grad = rng.randn(*i.shape).astype(i.data.dtype)
o.backward(o_grad)
assert np.all(i.g == 0)
def augmentation(h, test, aug):
'''
Data augmentation by randomly shifting up to 2 pixels.
Args:
h (nnabla.Variable): Shape [B, C, H, W]
test (bool): Only valid if aug is None, and the augmentation is applied if test=False.
aug (None or bool): Whether the augmentation is applied or not.
Returns:
nnabla.Variable: Shape [B, C, H, W]
'''
if aug is None:
aug = not test
if aug:
h = F.random_shift(h, (0, 0, 2, 2), seed=0)
return h
def augment(batch, aug_list, p_aug=1.0):
if isinstance(p_aug, float):
p_aug = nn.Variable.from_numpy_array(p_aug * np.ones((1,)))
if "flip" in aug_list:
rnd = F.rand(shape=[batch.shape[0], ])
batch_aug = F.random_flip(batch, axes=(2, 3))
batch = F.where(
F.greater(F.tile(p_aug, batch.shape[0]), rnd), batch_aug, batch)
if "lrflip" in aug_list:
rnd = F.rand(shape=[batch.shape[0], ])
batch_aug = F.random_flip(batch, axes=(3,))
batch = F.where(
F.greater(F.tile(p_aug, batch.shape[0]), rnd), batch_aug, batch)
if "translation" in aug_list and batch.shape[2] >= 8:
rnd = F.rand(shape=[batch.shape[0], ])
# Currently nnabla does not support random_shift with border_mode="noise"
mask = np.ones((1, 3, batch.shape[2], batch.shape[3]))
mask[:, :, :, 0] = 0
mask[:, :, :, -1] = 0
mask[:, :, 0, :] = 0
mask[:, :, -1, :] = 0
batch_int = F.concatenate(
batch, nn.Variable().from_numpy_array(mask), axis=0)
batch_int_aug = F.random_shift(batch_int, shifts=(
batch.shape[2]//8, batch.shape[3]//8), border_mode="nearest")
batch_aug = F.slice(batch_int_aug, start=(
0, 0, 0, 0), stop=batch.shape)
mask_var = F.slice(batch_int_aug, start=(
batch.shape[0], 0, 0, 0), stop=batch_int_aug.shape)
batch_aug = batch_aug * F.broadcast(mask_var, batch_aug.shape)
batch = F.where(
F.greater(F.tile(p_aug, batch.shape[0]), rnd), batch_aug, batch)
if "color" in aug_list:
rnd = F.rand(shape=[batch.shape[0], ])
rnd_contrast = 1.0 + 0.5 * \
(2.0 * F.rand(shape=[batch.shape[0], 1, 1, 1]
) - 1.0) # from 0.5 to 1.5
rnd_brightness = 0.5 * \
(2.0 * F.rand(shape=[batch.shape[0], 1, 1, 1]
) - 1.0) # from -0.5 to 0.5
rnd_saturation = 2.0 * \
F.rand(shape=[batch.shape[0], 1, 1, 1]) # from 0.0 to 2.0
# Brightness
batch_aug = batch + rnd_brightness
# Saturation
mean_s = F.mean(batch_aug, axis=1, keepdims=True)
batch_aug = rnd_saturation * (batch_aug - mean_s) + mean_s
# Contrast
mean_c = F.mean(batch_aug, axis=(1, 2, 3), keepdims=True)
batch_aug = rnd_contrast * (batch_aug - mean_c) + mean_c
batch = F.where(
F.greater(F.tile(p_aug, batch.shape[0]), rnd), batch_aug, batch)
if "cutout" in aug_list and batch.shape[2] >= 16:
batch = F.random_erase(batch, prob=p_aug.d[0], replacements=(0.0, 0.0))
return batch
def single_image_augment(image):
image = F.image_augmentation(image, contrast=1.0, angle=0.0, flip_lr=True)
image = F.random_shift(image, shifts=(4, 4), border_mode="reflect")
return image
