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]
# Args
epsilon = self.forward_func.info.args["epsilon"]
# 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]:
# Simply using " / dy" causes the numerical instability
diff = x0 - x1
mask = F.greater_scalar(F.abs(diff), epsilon)
maskp = F.greater_scalar(diff, 0.0)
maskn = 1.0 - maskp
g_dy_ = (g_dx0 - g_dx1) * (maskp - maskn) * mask
if accum[2]:
g_dy += g_dy_
else:
g_dy.copy_from(g_dy_)
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 norm_normalization_backward(inputs, p=None, axes=None, eps=1e-12):
"""
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]
if p is None:
p = 2.0
axes = list(range(x0.ndim)) if axes is None else force_list(axes)
x_abs = F.abs(x0)
x_pow = F.pow_scalar(x_abs, p)
x_sum = F.sum(x_pow, axes, keepdims=True)
# x_norm = x_sum ** (1./p)
# Div2 backward
dx = dy * x_sum**(-1. / p)
dx_norm = -dy * x0 * x_sum**(-2. / p)
dx_norm = sum_for_arithmetics(dx_norm, x_sum)
# Norm backward
x_sign = no_grad(F.sign(x0))
dx += dx_norm * x_sum**(1. / p - 1.) * x_abs**(p - 1.) * x_sign
return dx
def norm_backward(inputs, p=None, axes=None, keep_dims=False):
"""
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]
if p is None:
p = 2.0
axes = list(range(x0.ndim)) if axes is None else force_list(axes)
x_abs = F.abs(x0)
x_pow = F.pow_scalar(x_abs, p)
x_sum = F.sum(x_pow, axes, keepdims=True)
# Add axis for mul2
if not keep_dims:
shape = list(x0.shape)
for a in axes:
shape[a] = 1
dy = dy.reshape(shape)
x_sign = no_grad(F.sign(x0))
dx = dy * x_sum**(1. / p - 1.) * x_abs**(p - 1.) * x_sign
return dx
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 warp_coordinates(self, coordinates):
theta = self.theta
theta = F.reshape(
theta, theta.shape[:1] + (1,) + theta.shape[1:], inplace=False)
if coordinates.shape[0] == self.bs:
transformed = F.batch_matmul(
F.tile(theta[:, :, :, :2],
(1, coordinates.shape[1], 1, 1)),
F.reshape(coordinates, coordinates.shape + (1,), inplace=False)) + theta[:, :, :, 2:]
else:
transformed = F.batch_matmul(
F.tile(theta[:, :, :, :2],
(1, coordinates.shape[1], 1, 1)),
F.tile(F.reshape(coordinates, coordinates.shape + (1,), inplace=False),
(self.bs / coordinates.shape[0], 1, 1, 1))) + theta[:, :, :, 2:]
transformed = F.reshape(
transformed, transformed.shape[:-1], inplace=False)
if self.tps:
control_points = self.control_points
control_params = self.control_params
distances = F.reshape(
coordinates, (coordinates.shape[0], -1, 1, 2), inplace=False) - F.reshape(control_points, (1, 1, -1, 2))
distances = F.sum(F.abs(distances), axis=distances.ndim - 1)
result = distances ** 2
result = result * F.log(distances + 1e-6)
result = result * control_params
result = F.sum(result, axis=2)
result = F.reshape(
result, (self.bs, coordinates.shape[1], 1), inplace=False)
transformed = transformed + result
return transformed
def sample_noise(inpt_size, out_size):
_f = lambda x: F.sign(x) * F.pow_scalar(F.abs(x), 0.5)
noise = _f(F.randn(shape=(inpt_size + out_size, )))
eps_w = F.batch_matmul(F.reshape(noise[:inpt_size], (1, -1)),
F.reshape(noise[inpt_size:], (1, -1)), True)
eps_b = noise[inpt_size:]
return eps_w, eps_b
def forward(self, x): N, C, H, W = x.shape log_abs = F.log(F.abs(self.scale)) logdet = H*W*F.sum(log_abs) if self.logdet: return self.scale * (x + self.loc), logdet else: return self.scale * (x + self.loc)
def forward_impl(self, inputs, outputs):
x = inputs[0].data
M = inputs[1].data
y = outputs[0].data
y.copy_from(x)
if not self.training:
return
Mb = F.max(F.abs(x), keepdims=True)
F.maximum2(M, Mb, outputs=[M])
def softsign_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]
dx0 = dy * (1 / (1 + F.abs(x0))**2)
return dx0
def binary_tanh_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]
m0 = F.less_scalar(F.abs(x0), 1.0)
m0 = no_grad(m0)
dx0 = dy * m0
return dx0
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]
# Args
delta = self.forward_func.info.args["delta"]
# 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[0] or prop_down[1] or prop_down[2]:
mask = F.less_scalar(F.abs(x0 - x1), delta)
if prop_down[0]:
if accum[0]:
g_x0 += mask * 2 * dy * (g_dx0 - g_dx1)
else:
g_x0.copy_from(mask * 2 * dy * (g_dx0 - g_dx1))
if prop_down[1]:
if accum[1]:
g_x1 += mask * 2 * dy * (g_dx1 - g_dx0)
else:
g_x1.copy_from(mask * 2 * dy * (g_dx1 - g_dx0))
if prop_down[2]:
# Simply using " / dy" causes the numerical instability
diff = x0 - x1
pmask = F.greater_scalar(diff, 0.0)
nmask = (1.0 - pmask)
omask = (1.0 - mask)
g_dx_diff = g_dx0 - g_dx1
g_dy_ = 2.0 * g_dx_diff * \
(diff * mask + delta * omask * (pmask - nmask))
if accum[2]:
g_dy += g_dy_
else:
g_dy.copy_from(g_dy_)
def secant(x0, x1, implicit_function, max_post_itr, eps=1e-16):
f0 = implicit_function(x0) # > 0
f1 = implicit_function(x1) # < 0
for i in range(max_post_itr):
nu = f0 * (x1 - x0)
de = f1 - f0
mask0 = F.greater_scalar(F.abs(de), eps)
mask1 = 1 - mask0
nu = mask0 * nu + mask1 * 0
de = mask0 * de + mask1 * 1
xm = x0 - nu / de
fm = implicit_function(xm)
mp = F.greater_equal_scalar(fm, 0)
mn = 1 - mp
x0 = mp * xm + mn * x0
f0 = mp * fm + mn * f0
x1 = mn * xm + mp * x1
f1 = mn * fm + mp * f1
return x0, x1
def invertible_conv(x, reverse, rng, scope):
r"""Invertible 1x1 Convolution Layer.
Args:
x (nn.Variable): Input variable.
reverse (bool): Whether it's a reverse direction.
rng (numpy.random.RandomState): A random generator.
scope (str): The scope.
Returns:
nn.Variable: The output variable.
"""
batch_size, c, n_groups = x.shape
with nn.parameter_scope(scope):
# initialize w by an orthonormal matrix
w_init = np.linalg.qr(rng.randn(c, c))[0][None, ...]
W_var = get_parameter_or_create("W", (1, c, c), w_init, True, True)
W = F.batch_inv(W_var) if reverse else W_var
x = F.convolution(x, F.reshape(W, (c, c, 1)), None, stride=(1, ))
if reverse:
return x
log_det = batch_size * n_groups * F.log(F.abs(F.batch_det(W)))
return x, log_det
def get_d_layer(real_layers, fake_layers):
"""
discriminator layer loss
"""
fix_range = 0.02 # hard coded, all layers are roughly scaled to this value
sum_layer_loss = 0 # adds-on for generator
layer_loss_list = []
layer_n = len(real_layers)
# hard coded, an overall average of all layers
layer_norm = [12.0, 14.0, 24.0, 100.0]
for layer_i in range(layer_n):
real_layer = real_layers[layer_i]
false_layer = fake_layers[layer_i]
layer_diff = real_layer - false_layer
layer_loss = F.mean(F.sum(F.abs(layer_diff), axis=[3])) # an l1 loss
layer_loss_list += [layer_loss]
scaled_layer_loss = fix_range * \
layer_loss / layer_norm[layer_i]
sum_layer_loss += scaled_layer_loss
return sum_layer_loss
def train(generator, discriminator, patch_gan, solver_gen, solver_dis,
weight_l1, train_iterator, val_iterator, epoch, monitor, interval):
# Create Network Graph
# for training
im, la = train_iterator.next() # for checking image shape
real = nn.Variable(im.shape) # real
x = nn.Variable(la.shape) # x
# for validation
real_val = nn.Variable(im.shape) # real
x_val = nn.Variable(la.shape) # x
# Generator
fake = generator(x, test=False)
# pix2pix infers just like training mode.
fake_val = generator(x_val, test=False)
fake_val.persistent = True # Keep to visualize
# Discriminator
fake_y = discriminator(x, fake, patch_gan=patch_gan, test=False)
real_y = discriminator(x, real, patch_gan=patch_gan, test=False)
real_target = nn.Variable(fake_y.shape)
real_target.data.fill(1)
fake_target = nn.Variable(real_y.shape)
fake_target.data.zero()
loss_gen = F.mean(weight_l1 * F.abs(real - fake)) + \
F.mean(F.sigmoid_cross_entropy(fake_y, real_target))
loss_dis = F.mean(
F.sigmoid_cross_entropy(real_y, real_target) +
F.sigmoid_cross_entropy(fake_y, fake_target))
# Setting Solvers
with nn.parameter_scope('generator'):
solver_gen.set_parameters(nn.get_parameters())
with nn.parameter_scope('discriminator'):
solver_dis.set_parameters(nn.get_parameters())
# Create Monitors
monitors = {
'loss_gen':
nm.MonitorSeries("Generator loss", monitor, interval=interval),
'loss_dis':
nm.MonitorSeries("Discriminator loss", monitor, interval=interval),
'time':
nm.MonitorTimeElapsed("Training time", monitor, interval=interval),
'fake':
nm.MonitorImageTile(
"Fake images",
monitor,
interval=interval,
num_images=2,
normalize_method=lambda x: np.clip(np.divide(x, 255.0), 0.0, 1.0)),
}
i = 0
for e in range(epoch):
logger.info('Epoch = {}'.format(e))
# Training
while e == train_iterator.epoch:
# forward / backward process
real.d, x.d = train_iterator.next()
solver_dis.zero_grad()
solver_gen.zero_grad()
# Discriminator
loss_dis.forward(clear_no_need_grad=True)
loss_dis.backward(clear_buffer=True)
solver_dis.update()
# Generator
loss_gen.forward(clear_no_need_grad=True)
loss_gen.backward(clear_buffer=True)
solver_gen.update()
monitors['time'].add(i)
monitors['loss_gen'].add(i, loss_gen.d.copy())
monitors['loss_dis'].add(i, loss_dis.d.copy())
# Validation
real_val.d, x_val.d = val_iterator.next()
fake_val.forward()
pix2pix_vis = np.stack(
[label_to_image(x_val.d),
normalize_image(fake_val.d)],
axis=1).reshape((-1, ) + fake.shape[1:])
monitors['fake'].add(i, pix2pix_vis)
i += 1
# save parameters of generator
save_path = os.path.join(monitor._save_path,
'generator_model_{}.h5'.format(i))
with nn.parameter_scope('generator'):
nn.save_parameters(save_path)
return save_path
def main():
random.seed(args.seed)
np.random.seed(args.seed)
# Prepare for CUDA.
ctx = get_extension_context('cudnn', device_id=args.gpus)
nn.set_default_context(ctx)
start_full_time = time.time()
from iterator import data_iterator
# Data list for sceneflow data set
train_list = "./dataset/sceneflow_train.csv"
test_list = "./dataset/sceneflow_test.csv"
train = True
validation = True
# Set monitor path.
monitor_path = './nnmonitor' + str(datetime.now().strftime("%Y%m%d%H%M%S"))
img_left, img_right, disp_img = read_csv(train_list)
img_left_test, img_right_test, disp_img_test = read_csv(test_list)
train_samples = len(img_left)
test_samples = len(img_left_test)
train_size = int(len(img_left) / args.batchsize_train)
test_size = int(len(img_left_test) / args.batchsize_test)
# Create data iterator.
data_iterator_train = data_iterator(
train_samples, args.batchsize_train, img_left, img_right, disp_img, train=True, shuffle=True, dataset=args.dataset)
data_iterator_test = data_iterator(
test_samples, args.batchsize_test, img_left_test, img_right_test, disp_img_test, train=False, shuffle=False, dataset=args.dataset)
# Set data size
print(train_size, test_size)
# Define data shape for training.
var_left = nn.Variable(
(args.batchsize_train, 3, args.crop_height, args.crop_width))
var_right = nn.Variable(
(args.batchsize_train, 3, args.crop_height, args.crop_width))
var_disp = nn.Variable(
(args.batchsize_train, 1, args.crop_height, args.crop_width))
# Define data shape for testing.
var_left_test = nn.Variable(
(args.batchsize_test, 3, args.im_height, args.im_width))
var_right_test = nn.Variable(
(args.batchsize_test, 3, args.im_height, args.im_width))
var_disp_test = nn.Variable(
(args.batchsize_test, 1, args.im_height, args.im_width))
mask_test = nn.Variable(
(args.batchsize_test, 1, args.im_height, args.im_width))
if args.loadmodel is not None:
# Loading CNN pretrained parameters.
nn.load_parameters(args.loadmodel)
# === for Training ===
# Definition of pred
pred1, pred2, pred3 = psm_net(var_left, var_right, args.maxdisp, True)
mask_train = F.less_scalar(var_disp, args.maxdisp)
sum_mask = F.maximum_scalar(F.sum(mask_train), 1)
# Definition of loss
loss = 0.5 * (0.5 * F.sum(F.huber_loss(pred1, var_disp)*mask_train)/(sum_mask) + 0.7 * F.sum(F.huber_loss(
pred2, var_disp)*mask_train)/(sum_mask) + F.sum(F.huber_loss(pred3, var_disp)*mask_train)/(sum_mask))
# === for Testing ===
# Definition of pred
mask_test = F.less_scalar(var_disp_test, args.maxdisp)
sum_mask_test = F.maximum_scalar(F.sum(mask_test), 1)
pred_test = psm_net(var_left_test, var_right_test, args.maxdisp, False)
test_loss = F.sum(F.abs(pred_test - var_disp_test)*mask_test)/sum_mask_test
# Prepare monitors.
monitor = Monitor(monitor_path)
monitor_train = MonitorSeries('Training loss', monitor, interval=1)
monitor_test = MonitorSeries('Validation loss', monitor, interval=1)
monitor_time_train = MonitorTimeElapsed(
"Training time/epoch", monitor, interval=1)
# Create a solver (parameter updater)
solver = S.Adam(alpha=0.001, beta1=0.9, beta2=0.999)
# Set Parameters
params = nn.get_parameters()
solver.set_parameters(params)
params2 = nn.get_parameters(grad_only=False)
solver.set_parameters(params2)
for epoch in range(1, args.epochs+1):
print('This is %d-th epoch' % (epoch))
if validation:
## teting ##
total_test_loss = 0
index_test = 0
while index_test < test_size:
var_left_test.d, var_right_test.d, var_disp_test.d = data_iterator_test.next()
test_loss.forward(clear_no_need_grad=True)
total_test_loss += test_loss
print('Iter %d test loss = %.3f' % (index_test, test_loss.d))
index_test += 1
test_error = total_test_loss/test_size
print('epoch %d total 3-px error in val = %.3f' %
(epoch, test_error.d))
# Pass validation loss to a monitor.
monitor_test.add(epoch, test_error)
if train:
## training ##
total_train_loss = 0
index = 0
while index < train_size:
# Get mini batch
# Preprocess
var_left.d, var_right.d, var_disp.d = data_iterator_train.next()
loss.forward(clear_no_need_grad=True)
# Initialize gradients
solver.zero_grad()
# Backward execution
loss.backward(clear_buffer=True)
# Update parameters by computed gradients
solver.update()
print('Iter %d training loss = %.3f' %
(index, loss.d))
total_train_loss += loss.d
index += 1
train_error = total_train_loss/train_size
monitor_time_train.add(epoch)
print('epoch %d total training loss = %.3f' % (epoch, train_error))
# Pass training loss to a monitor.
monitor_train.add(epoch, train_error)
print('full training time = %.2f HR' %
((time.time() - start_full_time)/3600))
# Save Parameter
out_param_file = os.path.join(
args.savemodel, 'psmnet_trained_param_' + str(epoch) + '.h5')
nn.save_parameters(out_param_file)
def sigma_regularization(ctx, log_var, one):
with nn.context_scope(ctx):
h = F.exp(log_var)
h = F.pow_scalar(h, 0.5)
r = F.mean(F.abs(h - one))
return r
def sr_loss(ctx, pred0, pred1):
with nn.context_scope(ctx):
loss_sr = F.mean(F.abs(pred0 - pred1))
return loss_sr
def reconstruction_loss(imgA, imgB):
return F.mean(F.abs(imgA - imgB))
def sigma_regularization(ctx, log_var, one):
with nn.context_scope(ctx):
h = F.exp(log_var)
h = F.pow_scalar(h, 0.5)
r = F.mean(F.abs(h - one))
return r
def sr_loss(ctx, pred0, pred1):
with nn.context_scope(ctx):
loss_sr = F.mean(F.abs(pred0 - pred1))
return loss_sr
def parametric_pow2_quantize(x,
sign=True,
with_zero=True,
n_init=8,
n_min=1,
n_max=16,
m_init=1,
m_min=-8,
m_max=8,
fix_parameters=False):
"""Parametric version of `pow2_quantize` where the
bitwidth `n` and dynamic range `m` are learnable parameters.
Args:
x(~nnabla.Variable): N-D array as input
sign (bool): keep sign information during quantization.
with_zero (bool): quantize small weights to zero.
n_init (:obj:`nnabla.initializer.BaseInitializer` or :obj:`numpy.ndarray`): Initializer for bitwidth parameter.
n_min (int): lower bound for bitwidth.
n_max (int): upper bound for bitwidth.
m_init (:obj:`nnabla.initializer.BaseInitializer` or :obj:`numpy.ndarray`): Initializer for dynamic range.
m_min (float): lower bound for dynamic range.
m_max (float): upper bound for dynamic range.
fix_parameters (bool): When set to `True`, the negative slope values
will not be updated.
Returns:
~nnabla.Variable: N-D array.
"""
def clip_scalar(v, min_value, max_value):
return F.minimum_scalar(F.maximum_scalar(v, min_value), max_value)
def broadcast_scalar(v, shape):
return F.broadcast(F.reshape(v, (1, ) * len(shape), inplace=False),
shape=shape)
def quantize_pow2(v):
return 2**F.round(F.log(F.abs(v)) / np.log(2.))
n = get_parameter_or_create("n", (),
ConstantInitializer(n_init),
need_grad=True,
as_need_grad=not fix_parameters)
m = get_parameter_or_create("m", (),
ConstantInitializer(m_init),
need_grad=True,
as_need_grad=not fix_parameters)
# ensure that bitwidth is in specified range and an integer
n_q = F.round(clip_scalar(n, n_min, n_max))
if sign:
n_q = n_q - 1
if with_zero:
n_q = n_q - 1
# ensure that dynamic range is in specified range and an integer
m_q = F.round(clip_scalar(m, m_min, m_max))
# compute min/max value that we can represent
x_max = 2**m_q
x_min = 2**(m_q - (2**n_q) + 1)
# broadcast variables to correct size
x_min = broadcast_scalar(x_min, shape=x.shape)
x_max = broadcast_scalar(x_max, shape=x.shape)
# if unsigned, then quantize all negative values to zero
if not sign:
x = F.relu(x)
# compute absolute value/sign of input
ax = F.abs(x)
sx = F.sign(x)
if with_zero:
# prune smallest elements (in magnitude) to zero if they are smaller
# than `x_min / \sqrt(2)`
x_threshold = x_min / np.sqrt(2)
idx1 = F.greater_equal(ax, x_threshold) * F.less(ax, x_min)
idx2 = F.greater_equal(ax, x_min) * F.less(ax, x_max)
idx3 = F.greater_equal(ax, x_max)
else:
idx1 = F.less(ax, x_min)
idx2 = F.greater_equal(ax, x_min) * F.less(ax, x_max)
idx3 = F.greater_equal(ax, x_max)
# do not backpropagate gradient through indices
idx1.need_grad = False
idx2.need_grad = False
idx3.need_grad = False
# do not backpropagate gradient through sign
sx.need_grad = False
# take care of values outside of dynamic range
return sx * (x_min * idx1 + quantize_pow2(ax) * idx2 + x_max * idx3)
def parametric_pow2_quantize_xmin_xmax(x,
sign=True,
with_zero=True,
xmin_init=2**-7,
xmin_min=2**-15,
xmin_max=256,
xmax_init=2**0,
xmax_min=2**-8,
xmax_max=256,
fix_parameters=False):
"""Parametric version of `pow2_quantize` where the
min value `xmin` and max value `xmax` are learnable parameters.
Returns:
~nnabla.Variable: N-D array.
"""
def clip_scalar(v, min_value, max_value):
return F.minimum_scalar(F.maximum_scalar(v, min_value), max_value)
def broadcast_scalar(v, shape):
return F.broadcast(F.reshape(v, (1, ) * len(shape), inplace=False),
shape=shape)
def quantize_pow2(v):
return 2.**F.round(F.log(F.abs(v)) / np.log(2.))
xmin = get_parameter_or_create("xmin", (),
ConstantInitializer(xmin_init),
need_grad=True,
as_need_grad=not fix_parameters)
xmax = get_parameter_or_create("xmax", (),
ConstantInitializer(xmax_init),
need_grad=True,
as_need_grad=not fix_parameters)
# ensure that minimum dynamic range is in specified range and a power-of-two
xmin = quantize_pow2(clip_scalar(xmin, xmin_min, xmin_max))
# ensure that minimum dynamic range is in specified range and a power-of-two
xmax = quantize_pow2(clip_scalar(xmax, xmax_min, xmax_max))
# broadcast variables to correct size
xmin = broadcast_scalar(xmin, shape=x.shape)
xmax = broadcast_scalar(xmax, shape=x.shape)
# if unsigned, then quantize all negative values to zero
if not sign:
x = F.relu(x)
# compute absolute value/sign of input
ax = F.abs(x)
sx = F.sign(x)
if with_zero:
# prune smallest elements (in magnitude) to zero if they are smaller
# than `x_min / \sqrt(2)`
x_threshold = xmin / np.sqrt(2)
idx1 = F.greater_equal(ax, x_threshold) * F.less(ax, xmin)
idx2 = F.greater_equal(ax, xmin) * F.less(ax, xmax)
idx3 = F.greater_equal(ax, xmax)
else:
idx1 = F.less(ax, xmin)
idx2 = F.greater_equal(ax, xmin) * F.less(ax, xmax)
idx3 = F.greater_equal(ax, xmax)
# do not backpropagate gradient through indices
idx1.need_grad = False
idx2.need_grad = False
idx3.need_grad = False
# do not backpropagate gradient through sign
sx.need_grad = False
# take care of values outside of dynamic range
return sx * (xmin * idx1 + quantize_pow2(ax) * idx2 + xmax * idx3)
def quantize_pow2(v):
return 2.**F.round(F.log(F.abs(v)) / np.log(2.))
def bidirectional_sphere_trace(self, camloc, raydir, t_start, t_finish):
t_f = F.identity(t_start)
x_f = camloc + t_f * raydir
s_f = self.sdf(x_f)
mask_hit_eps_f = 0 * F.identity(t_f)
t_b = F.identity(t_finish)
x_b = camloc + t_b * raydir
s_b = self.sdf(x_b)
mask_hit_eps_b = 0 * F.identity(t_b)
for i in range(self.sphere_trace_itr - 1):
# Forward direction
mask_hit_eps_f_i = F.less_equal_scalar(F.abs(s_f), self.eps)
mask_hit_eps_f += (1 - mask_hit_eps_f) * mask_hit_eps_f_i
t_f += (1 - mask_hit_eps_f) * s_f
x_f = camloc + t_f * raydir
s_f_prev = F.identity(s_f)
s_f = self.sdf(x_f)
mask_pos_f_prev = (1 - mask_hit_eps_f) * \
F.greater_scalar(s_f_prev, 0)
mask_neg_f = (1 - mask_hit_eps_f) * F.less_scalar(s_f, 0)
mask_revert_f = mask_pos_f_prev * mask_neg_f
t_f -= mask_revert_f * s_f_prev
s_f = mask_revert_f * s_f_prev + (1 - mask_revert_f) * s_f
# Backward direction
mask_hit_eps_b_i = F.less_equal_scalar(F.abs(s_b), self.eps)
mask_hit_eps_b += (1 - mask_hit_eps_b) * mask_hit_eps_b_i
t_b -= (1 - mask_hit_eps_b) * s_b
x_b = camloc + t_b * raydir
s_b_prev = F.identity(s_b)
s_b = self.sdf(x_b)
mask_pos_b_prev = (1 - mask_hit_eps_b) * \
F.greater_scalar(s_b_prev, 0)
mask_neg_b = (1 - mask_hit_eps_b) * F.less_scalar(s_b, 0)
mask_revert_b = mask_pos_b_prev * mask_neg_b
t_b += mask_revert_b * s_b_prev
s_b = mask_revert_b * s_b_prev + (1 - mask_revert_b) * s_b
## print("s_f neg", np.sum(s_f.data < 0))
## print("s_b neg", np.sum(s_b.data < 0))
# Fine grained start/finish points
t_f0 = t_f
t_f1 = t_f + mask_revert_f * s_f_prev
x_hit_st0 = camloc + t_f0 * raydir
## x0, x1 = self.post_method(x_hit_st0, camloc + t_f1 * raydir)
## t_f0 = F.norm((x0 - camloc), axis=(x0.ndim - 1), keepdims=True)
## t_f1 = F.norm((x1 - camloc), axis=(x1.ndim - 1), keepdims=True)
mask_hit_f1b = mask_revert_f * F.less(t_f1, t_b)
t_b = t_f1 * mask_hit_f1b + t_b * (1 - mask_hit_f1b)
# Reverse the opposite case
mask_fb = F.less(t_f, t_b)
t_f = t_f * mask_fb + t_start * (1 - mask_fb)
t_b = t_b * mask_fb + t_finish * (1 - mask_fb)
return x_hit_st0, t_f, t_b, mask_hit_eps_f
def get_tecogan_model(conf, r_inputs, r_targets, scope_name, tecogan=True):
"""
Create computation graph and variables for TecoGAN.
"""
# r_inputs, r_targets : shape (batch, conf.train.rnn_n, h, w, c)
rnn_length = conf.train.rnn_n
if tecogan:
r_inputs, r_targets = get_tecogan_inputs(r_inputs, r_targets)
rnn_length = rnn_length * 2 - 1
# get the consecutive frame sequences from the input sequence
frame_t_pre, frame_t = r_inputs[:, 0:-1, :, :, :], r_inputs[:, 1:, :, :, :]
# Get flow estimations
fnet_output = get_fnet_output(conf, rnn_length, frame_t_pre, frame_t,
scope_name)
# Get the generated HR output frames
gen_outputs = get_generator_output(conf, rnn_length, r_inputs,
fnet_output.flow_hr, scope_name)
s_gen_output = F.reshape(
gen_outputs, (conf.train.batch_size * rnn_length,
conf.train.crop_size * 4, conf.train.crop_size * 4, 3),
inplace=False)
s_targets = F.reshape(
r_targets, (conf.train.batch_size * rnn_length,
conf.train.crop_size * 4, conf.train.crop_size * 4, 3),
inplace=False)
# Content loss (l2 loss)
content_loss = F.mean(
F.sum(F.squared_error(s_gen_output, s_targets), axis=[3]))
# Warp loss (l2 loss)
warp_loss = get_warp_loss(conf, rnn_length, frame_t, frame_t_pre,
fnet_output.flow_lr)
if tecogan:
d_data = get_d_data(conf, fnet_output.flow_hr, gen_outputs, r_targets,
rnn_length)
# Build the tempo discriminator for the real part and fake part
t_d = get_t_d(conf, r_inputs, d_data)
# Discriminator layer loss:
d_layer_loss = get_d_layer(t_d.real_layers, t_d.fake_layers)
# vgg loss (cosine similarity)
loss_vgg = get_vgg_loss(s_gen_output, s_targets)
# ping pong loss (an l1 loss)
gen_out_first = gen_outputs[:, 0:conf.train.rnn_n - 1, :, :, :]
gen_out_last_rev = gen_outputs[:, -1:-conf.train.rnn_n:-1, :, :, :]
pp_loss = F.mean(F.abs(gen_out_first - gen_out_last_rev))
# adversarial loss
t_adversarial_loss = F.mean(-F.log(t_d.tdiscrim_fake_output +
conf.train.eps))
# Overall generator loss
gen_loss = content_loss + pp_loss * conf.gan.pp_scaling + conf.gan.ratio * \
t_adversarial_loss + conf.gan.vgg_scaling * loss_vgg + \
conf.gan.dt_ratio_0 * d_layer_loss
# Discriminator loss
t_discrim_fake_loss = F.log(1 - t_d.tdiscrim_fake_output +
conf.train.eps)
t_discrim_real_loss = F.log(t_d.tdiscrim_real_output + conf.train.eps)
t_discrim_loss = F.mean(-(t_discrim_fake_loss + t_discrim_real_loss))
fnet_loss = gen_loss + warp_loss
set_persistent_all(r_targets, r_inputs, loss_vgg, gen_out_first,
gen_out_last_rev, pp_loss, d_layer_loss,
content_loss, warp_loss, gen_loss,
t_adversarial_loss, t_discrim_loss,
t_discrim_real_loss, d_data.t_vel,
d_data.t_gen_output, s_gen_output, s_targets)
Network = collections.namedtuple(
'Network', 'content_loss, warp_loss, fnet_loss, vgg_loss,'
'gen_loss, pp_loss, sum_layer_loss,t_adversarial_loss,'
't_discrim_loss,t_gen_output,t_discrim_real_loss')
return Network(content_loss=content_loss,
warp_loss=warp_loss,
fnet_loss=fnet_loss,
vgg_loss=loss_vgg,
gen_loss=gen_loss,
pp_loss=pp_loss,
sum_layer_loss=d_layer_loss,
t_adversarial_loss=t_adversarial_loss,
t_discrim_loss=t_discrim_loss,
t_gen_output=d_data.t_gen_output,
t_discrim_real_loss=t_discrim_real_loss)
gen_loss = content_loss
fnet_loss = gen_loss + warp_loss
set_persistent_all(content_loss, s_gen_output, warp_loss, gen_loss,
fnet_loss)
Network = collections.namedtuple(
'Network', 'content_loss, warp_loss, fnet_loss, gen_loss')
return Network(
content_loss=content_loss,
warp_loss=warp_loss,
fnet_loss=fnet_loss,
gen_loss=gen_loss,
)
def train():
parser = argparse.ArgumentParser()
parser.add_argument("--num-train-examples", type=int, default=1600)
parser.add_argument("--num-valid-examples", type=int, default=100)
parser.add_argument("--accum-grad", type=int, default=32)
parser.add_argument("--max-iter", type=int, default=6400)
parser.add_argument("--valid-interval", type=int, default=100)
parser.add_argument("--context", type=str, default="cpu")
parser.add_argument("--device-id", type=int, default=0)
args = parser.parse_args()
from nnabla.ext_utils import get_extension_context
extension_module = args.context
ctx = get_extension_context(extension_module, device_id=args.device_id)
nn.set_default_context(ctx)
# prepare dataset
tdataset = []
for i in range(args.num_train_examples):
V, E = random_graph(rng)
deg = degrees(V, E)
tdataset.append(([V], [utils.from_adjacency_list(E)], [deg]))
vdataset = []
for i in range(args.num_valid_examples):
V, E = random_graph(rng)
deg = degrees(V, E)
vdataset.append(([V], [utils.from_adjacency_list(E)], [deg]))
# prepare data iterator
tdata = data_iterator(SimpleDataSource2(tdataset, shuffle=True), 1, False,
False, False)
vdata = data_iterator(SimpleDataSource2(vdataset, shuffle=False), 1, False,
False, False)
# prepare monitors
monitor = M.Monitor("./degree")
tloss = M.MonitorSeries("Training Loss", monitor, interval=10)
verror = M.MonitorSeries("Validation Error", monitor, interval=10)
# prepare solver
solver = S.Adam()
# training loop
for i in range(args.max_iter):
l = 0
for b in range(args.accum_grad):
# read data
V, E, degree = tdata.next()
V = V[0][0]
E = E[0][0]
degree = degree[0][0]
# predict
output = predict(V, E)
# initialize solver
if i == 0 and b == 0:
solver.set_parameters(nn.get_parameters())
# calculate loss
label = nn.Variable(degree.shape)
label.data.data = degree
label = F.reshape(label, (len(V), 1))
loss = F.mean(F.squared_error(output, label))
# training
loss.forward(clear_no_need_grad=True)
loss.backward(clear_buffer=True)
l += loss.data.data
solver.update()
tloss.add(i, l / args.accum_grad)
l = 0
if i % args.valid_interval == 0:
# validation
# read data
e = 0
n = 0
for b in range(vdata.size):
V, E, degree = vdata.next()
V = V[0][0]
E = E[0][0]
degree = degree[0][0]
output = predict(V, E)
label = nn.Variable(degree.shape)
label.data.data = degree
label = F.reshape(label, (len(V), 1))
error = F.sum(F.less_scalar(F.abs(F.sub2(output, label)), 0.5))
error.forward()
e += error.data.data
n += len(V)
verror.add(i, e / n)
def calculate_alpha(
parameter_list,
X,
Y,
Y_label,
feature_valid,
solver,
output_valid,
pred,
loss,
phi,
l2,
):
min_loss = 10000.0
feature_valid.d = X
output_valid.d = Y
for epoch in range(args.epoch):
phi_loss = 0
loss.forward()
solver.zero_grad()
loss.backward()
phi_loss = phi.d / len(X)
temp_W = parameter_list
grad_loss = F.add_n(
*[F.mean(F.abs(p.grad)) for p in nn.get_parameters().values()])
grad_norm = F.add_n(
*[F.norm(p.grad) for p in nn.get_parameters().values()])
if grad_loss.data < min_loss:
if epoch == 0:
init_grad = grad_loss.data
min_loss = grad_loss.data
best_W = temp_W
if min_loss < init_grad / 200:
print("stopping criteria reached in epoch :{}".format(epoch))
break
parameter_list = backtracking_line_search(grad_norm, X, Y, loss,
len(X), loss.d, l2)
if epoch % 100 == 0:
print("Epoch:{:4d}\tloss:{}\tphi_loss:{}\tl2(lmbd):{}\tgrad:{}".
format(epoch, loss.d, phi_loss, args.lmbd * l2.d,
grad_loss.data))
for weight, param in zip(nn.get_parameters().values(), best_W):
weight.data.copy_from(param.data)
softmax_value = F.softmax(pred)
softmax_value.forward()
# derivative of softmax cross entropy
weight_matrix = softmax_value.d - Y
weight_matrix = np.divide(weight_matrix, (-2.0 * args.lmbd * len(Y)))
np.save(os.path.join(data_dir, "weight_matrix.npy"), weight_matrix)
# computer alpha
alpha = []
for ind, label in enumerate(Y_label.reshape(-1)):
alpha.append(float(weight_matrix[ind, int(label)]))
alpha = np.abs(np.array(alpha))
np.save(os.path.join(data_dir, "alpha_vgg_nnabla_score.npy"), alpha)
# calculate correlation
w = np.matmul(X.T, weight_matrix)
temp = np.matmul(X, w)
softmax_value = F.softmax(nn.Variable.from_numpy_array(temp))
softmax_value.forward()
y_p = softmax_value.d
print(
"L1 difference between ground truth prediction and prediction by representer theorem decomposition"
)
print(np.mean(np.abs(Y - y_p)))
from scipy.stats.stats import pearsonr
print(
"pearson correlation between ground truth prediction and prediciton by representer theorem"
)
corr, _ = pearsonr(Y.reshape(-1), y_p.reshape(-1))
print(corr)
