def huber_loss_backward(inputs, delta=1.0):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
dy = inputs[0]
x0 = inputs[1]
x1 = inputs[2]
d = x0 - x1
m0 = F.less_scalar(F.abs(d), delta)
m1 = 1 - m0
mg = F.greater(x0, x1)
ml = 1 - mg
m0 = no_grad(m0)
m1 = no_grad(m1)
mg = no_grad(mg)
ml = no_grad(ml)
t0 = 2 * d * m0
t1 = 2 * delta * m1 * mg
t2 = -2 * delta * m1 * ml
dx0 = dy * (t0 + t1 + t2)
dx1 = -dx0
return dx0, dx1
def backward_impl(self, inputs, outputs, prop_down, accum):
# inputs: [inputs_fwd_graph] + [inputs_bwd_graph] or
# [inputs_fwd_graph] + [outputs_fwd_graph] + [inputs_bwd_graph]
# Inputs
x0 = inputs[0].data
x1 = inputs[1].data
dy = inputs[2].data
# Outputs
dx0 = outputs[0].data
dx1 = outputs[1].data
# Grads of inputs
g_x0 = inputs[0].grad
g_x1 = inputs[1].grad
g_dy = inputs[2].grad
# Grads of outputs
g_dx0 = outputs[0].grad
g_dx1 = outputs[1].grad
# Computation
if prop_down[2]:
mask = F.greater(x0, x1)
if accum[2]:
g_dy += g_dx0 * mask + g_dx1 * (1.0 - mask)
else:
g_dy.copy_from(g_dx0 * mask + g_dx1 * (1.0 - mask))
def epsilon_insensitive_loss_backward(inputs, epsilon):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
dy = inputs[0]
x0 = inputs[1]
x1 = inputs[2]
d = x0 - x1
m0 = F.greater_scalar(F.abs(d), epsilon)
m1 = 1 - m0
mg = F.greater(x0, x1)
ml = 1 - mg
m0 = no_grad(m0)
mg = no_grad(mg)
ml = no_grad(ml)
t0 = m0 * mg
t1 = -m0 * ml
dx0 = dy * (t0 + t1)
dx1 = -dx0
return dx0, dx1
def maximum2_backward(inputs):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
dy = inputs[0]
x0 = inputs[1]
x1 = inputs[2]
m0 = F.greater(x0, x1)
m1 = 1 - m0
m0 = no_grad(m0)
m1 = no_grad(m1)
dx0 = dy * m0
dx1 = dy * m1
return dx0, dx1
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