def build_theano_functions(self, bottom, top):
# building Theano functions
from caffe_helper.theano_util import init_theano
init_theano()
import theano as tn
import theano.tensor as T
p = np.float32(self.p_)
axis = self.axis_
if axis is None:
axis = tuple(range(1, len(bottom[0].shape)))
# blob to CudaNdArray
# Forward pass
Tensor = T.TensorType('float32', [False] * len(bottom[0].shape))
s_x = Tensor('x') # bottom data
s_dz = Tensor('dz') # top diff
s_z = s_x * ((s_x**p).sum(axis, keepdims=True)**(np.float32(-1. / p)))
# See http://goo.gl/wIVRsP for `tn.Out(x, borrow=True)`
self.f_forward = tn.function([s_x], tn.Out(s_z, borrow=True))
# Backward pass
s_l = (s_dz * s_z).sum()
s_grad = tn.grad(s_l, wrt=s_x)
self.f_backward = tn.function([s_x, s_dz], tn.Out(s_grad, borrow=True))
def setup(self, bottom, top):
from caffe_helper.theano_util import init_theano
init_theano()
import theano as tn
import theano.tensor as T
assert len(bottom) == 2
assert len(top) == 1
s_y = T.matrix('y') # y in [-inf, inf]
s_t = T.matrix('t') # t in {-1, 0, 1} where 0 is ignored
s_dloss = T.scalar('dloss')
# Forward
# s_loss = T.mean(abs(s_t) * T.log1p(T.exp(-s_y * s_t))) # unstable
s_loss = -T.sum(
abs(s_t) * (
s_y * ((s_t >= 0) - (s_y >= 0)) - T.log1p(T.exp(-abs(s_y)))))\
/ T.maximum(T.sum(abs(s_t)), 1)
# Backward
s_p = 1 / (1 + T.exp(-s_y))
s_dy = s_dloss * abs(s_t) * (s_p - (s_t >= 0)) / \
T.maximum(T.sum(abs(s_t)), 1)
def _o(s):
return tn.Out(s, borrow=True)
self.tn_forward = tn.function([s_y, s_t], s_loss)
self.tn_backward = tn.function([s_y, s_t, s_dloss], _o(s_dy))
def build_theano_functions(self, bottom, top):
# building Theano functions
from caffe_helper.theano_util import init_theano
init_theano()
import theano as tn
import theano.tensor as T
p = np.float32(self.p_)
axis = self.axis_
if axis is None:
axis = tuple(range(1, len(bottom[0].shape)))
# blob to CudaNdArray
# Forward pass
Tensor = T.TensorType('float32', [False] * len(bottom[0].shape))
s_x = Tensor('x') # bottom data
s_dz = Tensor('dz') # top diff
s_z = s_x * (
(s_x**p).sum(axis, keepdims=True)**(np.float32(-1./p)))
# See http://goo.gl/wIVRsP for `tn.Out(x, borrow=True)`
self.f_forward = tn.function([s_x], tn.Out(s_z, borrow=True))
# Backward pass
s_l = (s_dz * s_z).sum()
s_grad = tn.grad(s_l, wrt=s_x)
self.f_backward = tn.function([s_x, s_dz], tn.Out(s_grad, borrow=True))
def setup(self, bottom, top):
self.reshape(bottom, top)
from caffe_helper.theano_util import init_theano
init_theano()
import theano as tn
import theano.tensor as T
shape1 = bottom[0].shape # prediction
shape2 = bottom[1].shape # label
s_p = T.TensorType('float32', [False] * len(shape1))('p')
s_t = T.TensorType('float32', [False] * len(shape2))('t')
# Forward pass
FLTMIN = np.finfo(np.float32).tiny
s_l = -T.mean(
T.log(T.maximum(FLTMIN, s_p.flatten(2)))[
T.arange(s_t.shape[0]), T.cast(s_t, 'int32')]
)
self.f_forward = tn.function(
[s_p, s_t], tn.Out(s_l, borrow=True))
# Backward pass
s_dz = T.fscalar('dz')
sg_p = tn.grad(s_dz * s_l, wrt=s_p)
self.f_backward = tn.function(
[s_p, s_t, s_dz], tn.Out(sg_p, borrow=True))
def setup(self, bottom, top):
from caffe_helper.theano_util import init_theano
init_theano()
import theano as tn
import theano.tensor as T
assert len(bottom) == 2
assert len(top) == 1
s_y = T.matrix('y') # y in [-inf, inf]
s_t = T.matrix('t') # t in {-1, 0, 1} where 0 is ignored
s_dloss = T.scalar('dloss')
# Forward
# s_loss = T.mean(abs(s_t) * T.log1p(T.exp(-s_y * s_t))) # unstable
s_loss = -T.sum(
abs(s_t) * (
s_y * ((s_t >= 0) - (s_y >= 0)) - T.log1p(T.exp(-abs(s_y)))))\
/ T.maximum(T.sum(abs(s_t)), 1)
# Backward
s_p = 1 / (1 + T.exp(-s_y))
s_dy = s_dloss * abs(s_t) * (s_p - (s_t >= 0)) / \
T.maximum(T.sum(abs(s_t)), 1)
def _o(s):
return tn.Out(s, borrow=True)
self.tn_forward = tn.function([s_y, s_t], s_loss)
self.tn_backward = tn.function([s_y, s_t, s_dloss], _o(s_dy))
def setup(self, bottom, top):
from caffe_helper.theano_util import init_theano
init_theano()
import theano as tn
import theano.tensor as T
assert len(bottom) == 2
assert len(top) == 1
s_y = T.matrix('y') # y in [-inf, inf]
s_t = T.matrix('t') # t in {-1, 0, 1} where 0 is ignored
# Forward
s_loss = T.sum(abs(s_t) * T.eq((s_y >= 0), (s_t >= 0))) \
/ T.maximum(T.sum(abs(s_t)), 1)
def _o(s):
return tn.Out(s, borrow=True)
self.tn_forward = tn.function([s_y, s_t], _o(s_loss))
def setup(self, bottom, top):
from caffe_helper.theano_util import init_theano
init_theano()
import theano as tn
import theano.tensor as T
assert len(bottom) == 2
assert len(top) == 1
s_y = T.matrix('y') # y in [-inf, inf]
s_t = T.matrix('t') # t in {-1, 0, 1} where 0 is ignored
# Forward
s_loss = T.sum(abs(s_t) * T.eq((s_y >= 0), (s_t >= 0))) \
/ T.maximum(T.sum(abs(s_t)), 1)
def _o(s):
return tn.Out(s, borrow=True)
self.tn_forward = tn.function([s_y, s_t], _o(s_loss))
def setup(self, bottom, top):
self.reshape(bottom, top)
from caffe_helper.theano_util import init_theano
init_theano()
import theano as tn
import theano.tensor as T
shape1 = bottom[0].shape # prediction
shape2 = bottom[1].shape # label
s_p = T.TensorType('float32', [False] * len(shape1))('p')
s_t = T.TensorType('float32', [False] * len(shape2))('t')
# Forward pass
FLTMIN = np.finfo(np.float32).tiny
s_l = -T.mean(
T.log(T.maximum(FLTMIN, s_p.flatten(2)))[T.arange(s_t.shape[0]),
T.cast(s_t, 'int32')])
self.f_forward = tn.function([s_p, s_t], tn.Out(s_l, borrow=True))
# Backward pass
s_dz = T.fscalar('dz')
sg_p = tn.grad(s_dz * s_l, wrt=s_p)
self.f_backward = tn.function([s_p, s_t, s_dz],
tn.Out(sg_p, borrow=True))
def setup(self, bottom, top):
from caffe_helper.theano_util import init_theano
init_theano()
import theano as tn
import theano.tensor as T
from theano.tensor.signal.conv import conv2d
assert len(bottom) >= 2
assert len(bottom) <= 3
assert len(top) == 1
# parameter
self.K_ = [0.01, 0.03]
self.L_ = 1.0
param = eval(self.param_str_)
self.hsize_ = param.get('hsize', 11)
self.sigma_ = param.get('sigma', 1.5)
assert self.hsize_ % 2 == 1
hsize = self.hsize_
sigma = self.sigma_
C1 = (self.K_[0] * self.L_) ** 2
C2 = (self.K_[1] * self.L_) ** 2
# Creating gaussian filter
x = np.exp(-0.5 * ((np.arange(hsize) - int(hsize / 2)) ** 2) /
(sigma ** 2))
filt = x.reshape(-1, 1) * x.reshape(1, -1)
filt /= filt.sum()
# Build a Theano function which computes SSIM and its gradients wrt two
# images
simg1_in = T.ftensor3()
simg2_in = T.ftensor3()
if len(bottom) > 2:
smask = T.ftensor3()
sk = T.sum(simg1_in * simg2_in * smask) \
/ T.sum(simg1_in * simg1_in * smask)
simg1 = sk * simg1_in * smask
simg2 = simg2_in * smask
else:
sk = T.sum(simg1_in * simg2_in) \
/ T.sum(simg1_in * simg1_in)
simg1 = sk * simg1_in
simg2 = simg2_in
sfilt = tn.shared(filt.astype(np.float32))
smu1 = conv2d(simg1, sfilt)
smu2 = conv2d(simg2, sfilt)
smu1_sq = smu1 * smu1
smu2_sq = smu2 * smu2
smu1_mu2 = smu1 * smu2
ssig1_sq = conv2d(simg1 * simg1, sfilt) - smu1_sq
ssig2_sq = conv2d(simg2 * simg2, sfilt) - smu2_sq
ssig12 = conv2d(simg1 * simg2, sfilt) - smu1_mu2
sssim = (
((2 * smu1_mu2 + C1) * (2 * ssig12 + C2))
/ ((smu1_sq + smu2_sq + C1) * (ssig1_sq + ssig2_sq + C2))
).mean()
sdssim = (1 - sssim) / 2
gimg1, gimg2 = tn.grad(sdssim, [simg1_in, simg2_in])
if len(bottom) > 2:
self.fdssim_with_grad = tn.function(
[simg1_in, simg2_in, smask], [sdssim, gimg1, gimg2])
else:
self.fdssim_with_grad = tn.function(
[simg1_in, simg2_in], [sdssim, gimg1, gimg2])
def setup(self, bottom, top):
from caffe_helper.theano_util import init_theano
init_theano()
import theano as tn
import theano.tensor as T
from theano.tensor.signal.conv import conv2d
assert len(bottom) >= 2
assert len(bottom) <= 3
assert len(top) == 1
# parameter
self.K_ = [0.01, 0.03]
self.L_ = 1.0
param = eval(self.param_str_)
self.hsize_ = param.get('hsize', 11)
self.sigma_ = param.get('sigma', 1.5)
assert self.hsize_ % 2 == 1
hsize = self.hsize_
sigma = self.sigma_
C1 = (self.K_[0] * self.L_)**2
C2 = (self.K_[1] * self.L_)**2
# Creating gaussian filter
x = np.exp(-0.5 * ((np.arange(hsize) - int(hsize / 2))**2) /
(sigma**2))
filt = x.reshape(-1, 1) * x.reshape(1, -1)
filt /= filt.sum()
# Build a Theano function which computes SSIM and its gradients wrt two
# images
simg1_in = T.ftensor3()
simg2_in = T.ftensor3()
if len(bottom) > 2:
smask = T.ftensor3()
sk = T.sum(simg1_in * simg2_in * smask) \
/ T.sum(simg1_in * simg1_in * smask)
simg1 = sk * simg1_in * smask
simg2 = simg2_in * smask
else:
sk = T.sum(simg1_in * simg2_in) \
/ T.sum(simg1_in * simg1_in)
simg1 = sk * simg1_in
simg2 = simg2_in
sfilt = tn.shared(filt.astype(np.float32))
smu1 = conv2d(simg1, sfilt)
smu2 = conv2d(simg2, sfilt)
smu1_sq = smu1 * smu1
smu2_sq = smu2 * smu2
smu1_mu2 = smu1 * smu2
ssig1_sq = conv2d(simg1 * simg1, sfilt) - smu1_sq
ssig2_sq = conv2d(simg2 * simg2, sfilt) - smu2_sq
ssig12 = conv2d(simg1 * simg2, sfilt) - smu1_mu2
sssim = (((2 * smu1_mu2 + C1) * (2 * ssig12 + C2)) /
((smu1_sq + smu2_sq + C1) *
(ssig1_sq + ssig2_sq + C2))).mean()
sdssim = (1 - sssim) / 2
gimg1, gimg2 = tn.grad(sdssim, [simg1_in, simg2_in])
if len(bottom) > 2:
self.fdssim_with_grad = tn.function([simg1_in, simg2_in, smask],
[sdssim, gimg1, gimg2])
else:
self.fdssim_with_grad = tn.function([simg1_in, simg2_in],
[sdssim, gimg1, gimg2])
