def __init__(
self,
ms,
level,
data,
sigma_lm,
params_flow_int,
# src=None,dst=None,transformed=None
):
# ipshell('hi')
src = data['src']
dst = data['dst']
transformed = data['transformed']
if not isinstance(src, CpuGpuArray):
raise ObsoleteError
if not isinstance(dst, CpuGpuArray):
raise ObsoleteError
if not isinstance(transformed, CpuGpuArray):
raise ObsoleteError
self.nCalls = 0
self.nCallbacks = 0
self.sigma_lm = sigma_lm
cpa_space = ms.L_cpa_space[level]
self.cpa_space = cpa_space
if src.shape[1] != cpa_space.dim_domain:
raise ValueError(src.shape, cpa_space.dim_domain)
self.mu = cpa_space.get_zeros_theta()
self.src = src
self.dst = dst
self.transformed = transformed
nPts = len(src)
self.nPts = nPts
self.err = CpuGpuArray.zeros_like(src)
self.ll = CpuGpuArray.zeros(nPts, dtype=src.dtype)
if nPts <= 1:
raise ValueError
self.err_by_der = CpuGpuArray.zeros((nPts - 1, src.shape[1]),
dtype=src.dtype)
self.ll_by_der = CpuGpuArray.zeros(nPts - 1, dtype=src.dtype)
self.params_flow_int = params_flow_int
self._pat = PAT(pa_space=cpa_space, Avees=cpa_space.get_zeros_PA())
def get_init_seg(dimy, dimx, nPixels_in_square_side, use_hex):
"""
"""
M = nPixels_in_square_side
if use_hex:
s = create_string(dimx, dimy, nPixels_in_square_side)
# print(s)
fname = os.path.join(dirname_precomputed_hex_inits, s)
try:
FilesDirs.raise_if_file_does_not_exist(fname)
print("Loading", fname)
seg = np.load(fname)
return seg
except FileDoesNotExistError:
pass
msg = """
I could not find a precomputed (image-independent)
honeycomb initilization for this image size and this values of n.
So I will compute it from scratch and we will save the result in
{}
Next time you will run the code for an image of size
nRows={}, nCols={}, with n = {},
it will be faster.
""".format(fname, dimy, dimx, nPixels_in_square_side)
print(msg)
seg = CpuGpuArray.zeros((dimy, dimx), dtype=np.int32)
# length of each side
a = np.sqrt(M**2 / (1.5 * np.sqrt(3)))
H = a
W = np.sqrt(3) * H
# XX and YY need to be float
YY, XX = np.mgrid[0:float(dimy) + 0 * 1.5 * H:1.5 * H,
0:float(dimx) + 0 * W:W]
XX[::2] += float(W) / 2
centers = np.vstack([XX.ravel(), YY.ravel()]).T.copy()
centers = CpuGpuArray(centers)
honeycomb(seg.gpu, centers.gpu, seg.size)
seg.gpu2cpu()
np.save(fname, seg.cpu)
return seg.cpu
else:
seg_cpu = np.zeros((dimy, dimx), dtype=np.int32)
yy, xx = np.mgrid[:dimy, :dimx]
xx = xx.astype(np.float)
yy = yy.astype(np.float)
dimx = float(dimx)
dimy = float(dimy)
nTimesInX = np.floor(xx / M).max() + 1
seg_cpu = np.floor(yy / M) * nTimesInX + np.floor(xx / M)
seg_cpu = seg_cpu.astype(np.int32)
return seg_cpu
def __init__(self,ms,level,data,
sigma_lm,
params_flow_int,
# src=None,dst=None,transformed=None
):
# ipshell('hi')
src=data['src']
dst=data['dst']
transformed=data['transformed']
if not isinstance(src,CpuGpuArray):
raise ObsoleteError
if not isinstance(dst,CpuGpuArray):
raise ObsoleteError
if not isinstance(transformed,CpuGpuArray):
raise ObsoleteError
self.nCalls = 0
self.nCallbacks = 0
self.sigma_lm=sigma_lm
cpa_space=ms.L_cpa_space[level]
self.cpa_space = cpa_space
if src.shape[1] != cpa_space.dim_domain:
raise ValueError(src.shape,cpa_space.dim_domain)
self.mu = cpa_space.get_zeros_theta()
self.src = src
self.dst = dst
self.transformed = transformed
nPts = len(src)
self.nPts = nPts
self.err = CpuGpuArray.zeros_like(src)
self.ll = CpuGpuArray.zeros(nPts,dtype=src.dtype)
if nPts <= 1:
raise ValueError
self.err_by_der = CpuGpuArray.zeros((nPts-1,src.shape[1]),dtype=src.dtype)
self.ll_by_der = CpuGpuArray.zeros(nPts-1,dtype=src.dtype)
self.params_flow_int=params_flow_int
self._pat = PAT(pa_space=cpa_space,
Avees=cpa_space.get_zeros_PA())
def get_init_seg(dimy,dimx,nPixels_in_square_side,use_hex):
"""
"""
M=nPixels_in_square_side
if use_hex:
s = create_string(dimx,dimy,nPixels_in_square_side)
# print s
fname = os.path.join(dirname_precomputed_hex_inits,s)
try:
FilesDirs.raise_if_file_does_not_exist(fname)
print "Loading",fname
seg = np.load(fname)
return seg
except FileDoesNotExistError:
pass
msg = """
I could not find a precomputed (image-independent)
honeycomb initilization for this image size and this values of n.
So I will compute it from scratch and we will save the result in
{}
Next time you will run the code for an image of size
nRows={}, nCols={}, with n = {},
it will be faster.
""".format(fname,dimy,dimx,nPixels_in_square_side)
print msg
seg = CpuGpuArray.zeros((dimy,dimx),dtype=np.int32)
# length of each side
a = np.sqrt(M ** 2 / ( 1.5 * np.sqrt(3) ))
H = a
W = np.sqrt(3)*H
# XX and YY need to be float
YY,XX = np.mgrid[0:float(dimy)+0*1.5*H:1.5*H,0:float(dimx)+0*W:W]
XX[::2]+= float(W)/2
centers = np.vstack([XX.ravel(),YY.ravel()]).T.copy()
centers = CpuGpuArray(centers)
honeycomb(seg.gpu,centers.gpu,seg.size)
seg.gpu2cpu()
np.save(fname,seg.cpu)
return seg.cpu
else:
seg_cpu = np.zeros((dimy,dimx),dtype=np.int32)
yy,xx = np.mgrid[:dimy,:dimx]
xx = xx.astype(np.float)
yy = yy.astype(np.float)
dimx = float(dimx)
dimy=float(dimy)
nTimesInX = np.floor(xx / M).max() + 1
seg_cpu = np.floor(yy / M) * nTimesInX + np.floor(xx / M)
seg_cpu = seg_cpu.astype(np.int32)
return seg_cpu
def get_cartoon(self):
"""
Replace pixels with superpixels means.
"""
img = self.img.cpu
nChannels = self.nChannels
img_disp = CpuGpuArray.zeros((img.shape[0],img.shape[1],3),dtype=np.int32)
get_cartoon(seg_gpu = self.seg.gpu, mu_i_gpu = self.superpixels.params.mu_i.gpu,
img_gpu= img_disp.gpu, nChannels=nChannels)
img_disp.gpu2cpu()
return img_disp.cpu
def get_cartoon(self):
"""
Replace pixels with superpixels means.
"""
img = self.img.cpu
nChannels = self.nChannels
img_disp = CpuGpuArray.zeros((img.shape[0], img.shape[1], 3), dtype=np.int32)
get_cartoon(seg_gpu=self.seg.gpu, mu_i_gpu=self.superpixels.params.mu_i.gpu,
img_gpu=img_disp.gpu, nChannels=nChannels)
img_disp.gpu2cpu()
return img_disp.cpu
def __init__(self,ms,level,data,
sigma_lm,
params_flow_int,
# src=None,dst=None,transformed=None
):
# ipshell('hi')
"""
Cost is level-dependent.
TODO: GPU in the LL part.
"""
src=data['src']
dst=data['dst']
transformed=data['transformed']
if not isinstance(src,CpuGpuArray):
raise ObsoleteError
if not isinstance(dst,CpuGpuArray):
raise ObsoleteError
if not isinstance(transformed,CpuGpuArray):
raise ObsoleteError
self.nCalls = 0
self.nCallbacks = 0
self.sigma_lm=sigma_lm
cpa_space=ms.L_cpa_space[level]
self.cpa_space = cpa_space
# 1/0
if src.shape[1] != cpa_space.dim_domain:
raise ValueError(src.shape,cpa_space.dim_domain)
#
# self.cpa_space = cpa_space
# self.cpa_cov_inv = msp.L_cpa_space_covs[level].cpa_cov_inv
self.mu = cpa_space.get_zeros_theta()
self.src = src
self.dst = dst
self.transformed = transformed
nPts = len(src)
self.nPts = nPts
self.err = CpuGpuArray.zeros_like(src)
self.ll = CpuGpuArray.zeros(nPts,dtype=src.dtype)
self.params_flow_int=params_flow_int
self._pat = PAT(pa_space=cpa_space,
Avees=cpa_space.get_zeros_PA())
mu = cpa_space.get_zeros_theta()
np.random.seed(0)
# theta *= 4
#
cpa_space.theta2Avees(theta=theta)
cpa_space.update_pat()
# params_flow_int.nTimeSteps *= 10
cell_idx = CpuGpuArray.zeros(len(pts),dtype=np.int32)
cpa_space.calc_cell_idx(pts,cell_idx)
# ipshell('st')
# 1/0
# 1/0
v = CpuGpuArray.zeros_like(pts)
cpa_space.calc_v(pts=pts,out=v)
print params_flow_int
print '#pts=',len(pts)
tic=time.clock()
cpa_space.calc_T_fwd(pts=pts,out=pts_transformed,
**params_flow_int)
def __init__(self, nSuperpixels, s_std, i_std, prior_count, nChannels):
"""
Initilize the parameters for the superpixels:
The means are set to zeros at this point,
and will be set later in the first M step.
The space/color covariances (and their inverse), however, are being set
to initial values here.
We use a Inverse-Wishart prior on the space covariance
Arguments:
nSuperpixels: the number of superpixels to generate
s_std: should be fixed as nPixels_on_side
i_std: control the relative importance between RGB and location.
The smaller it is, bigger the RGB effect is / more irregular the superpixels are.
prior_count: determines the weight of Inverse-Wishart prior of space covariance(ex:1,5,10)
nChannels: the number of channels of the input image (gray:1, LAB/RGB: 3)
"""
if nChannels not in (1,3):
raise NotImplementedError(nChannels)
dim_i=nChannels
dim_s=2
self.dim_i=dim_i
self.dim_s=dim_s
self.nSuperpixels=nSuperpixels
self.s_std, self.i_std, self.prior_count = s_std,i_std,prior_count
mu_s = CpuGpuArray.zeros((nSuperpixels,dim_s))
mu_i = CpuGpuArray.zeros((nSuperpixels,dim_i))
Sigma_s = CpuGpuArray.zeros(shape = (nSuperpixels,dim_s,dim_s))
J_s = CpuGpuArray.zeros_like(Sigma_s)
Sigma_i = CpuGpuArray.zeros((nSuperpixels,dim_i,dim_i))
J_i = CpuGpuArray.zeros_like(Sigma_i)
logdet_Sigma_i = CpuGpuArray.zeros((nSuperpixels,1)) # scalars
logdet_Sigma_s = CpuGpuArray.zeros((nSuperpixels,1))
# start with unnormalized counts (uniform)
counts = np.ones(nSuperpixels,dtype=np.int32)
counts = CpuGpuArray(counts)
self.params = Bunch()
self.params.mu_i = mu_i
self.params.mu_s = mu_s
self.params.Sigma_i = Sigma_i
self.params.Sigma_s = Sigma_s
self.params.prior_sigma_s_sum = Sigma_s
self.params.J_i = J_i
self.params.J_s = J_s
self.params.logdet_Sigma_i = logdet_Sigma_i
self.params.logdet_Sigma_s = logdet_Sigma_s
self.params.counts = counts
# set those parameters related to covariance
self.initialize_params()
# intermediate arrays needed for the Gaussian parameter calculation on GPU
self.gpu_helper = Bunch()
self.gpu_helper.mu_i_helper = gpuarray.zeros((nSuperpixels,dim_i),dtype=np.int64)
self.gpu_helper.mu_s_helper = gpuarray.zeros((nSuperpixels,dim_s),dtype=np.int64)
self.gpu_helper.prior_sigma_s = self.params.prior_sigma_s_sum.gpu.copy()
self.gpu_helper.sigma_s_helper = gpuarray.zeros((nSuperpixels,3),dtype=np.int64)
self.gpu_helper.log_count_helper = gpuarray.zeros(nSuperpixels,dtype=np.double)
self.gpu_helper.non_NaN_count = gpuarray.zeros(nSuperpixels,dtype=np.int32)
self.gpu_helper.NaN_count = gpuarray.zeros(nSuperpixels,dtype=np.int32)
def __init__(self,ms,level,data,
sigma_signal,
params_flow_int,
interp_type_for_ll,
# src=None,dst=None,transformed=None
):
# ipshell('hi')
if interp_type_for_ll not in self.supported_interp_types:
msg = """
interp_type_for_ll must be in
['gpu_linear',
cv2.INTER_LINEAR,
cv2.INTER_CUBIC,
cv2.INTER_LANCZOS4]
"""
raise ValueError(msg,interp_type_for_ll)
self.interp_type_for_ll=interp_type_for_ll
src=data['src']
transformed=data['transformed']
signal=data['signal']
for obj in [src,transformed]:
if not isinstance(obj,CpuGpuArray):
raise TypeError
for obj in [signal.src,signal.dst,signal.transformed]:
if not isinstance(obj,CpuGpuArray):
raise TypeError
self.nCalls = 0
self.nCallbacks = 0
self.sigma_signal=sigma_signal
cpa_space=ms.L_cpa_space[level]
self.cpa_space = cpa_space
if src.shape[1] != cpa_space.dim_domain:
raise ValueError(src.shape,cpa_space.dim_domain)
# self.mu = cpa_simple_mean(cpa_space)
self.my = cpa_space.get_zeros_theta()
self.src = src
self.transformed = transformed
self.signal = signal
# self.dim_signal = signal.src.shape[1]
if signal.src.ndim==2:
self.dim_signal = 2
else:
raise NotImplementedError
if self.dim_signal != 2:
raise NotImplementedError(signal.src.shape)
if self.signal.src.shape != self.signal.dst.shape:
raise ValueError
if self.signal.src.shape != self.signal.transformed.shape:
raise ValueError
nPts = len(src)
self.nPts = nPts
# self.err = CpuGpuArray.zeros_like(src)
# self.signal.err = CpuGpuArray.zeros_like(src)
self.signal.err = CpuGpuArray.zeros_like(self.signal.src)
self.ll = CpuGpuArray.zeros(nPts,dtype=src.dtype)
if nPts <= 1:
raise ValueError
self.params_flow_int=params_flow_int
self._pat = PAT(pa_space=cpa_space,
Avees=cpa_space.get_zeros_PA())
def example(img=None,tess='I',eval_cell_idx=True,eval_v=True,show_downsampled_pts=True,
valid_outside=True,base=[1,1],
scale_spatial=.1,
scale_value=100,
permute_cell_idx_for_display=True,
nLevels=3,
vol_preserve=False,
zero_v_across_bdry=[0,0],
use_lims_when_plotting=True):
show_downsampled_pts = bool(show_downsampled_pts)
eval_cell_idx = bool(eval_cell_idx)
eval_v = bool(eval_cell_idx)
valid_outside = bool(valid_outside)
permute_cell_idx_for_display = bool(permute_cell_idx_for_display)
vol_preserve = bool(vol_preserve)
if img is None:
img = Img(get_std_test_img())
else:
img=Img(img)
img = img[:,:,::-1] # bgr2rgb
tw = TransformWrapper(nRows=img.shape[0],
nCols=img.shape[1],
nLevels=nLevels,
base=base,
scale_spatial=scale_spatial, # controls the prior's smoothness
scale_value=scale_value, # controls the prior's variance
tess=tess,
vol_preserve=vol_preserve,
zero_v_across_bdry=zero_v_across_bdry,
valid_outside=valid_outside)
print tw
# You probably want to do that: padding image border with zeros
border_width=1
img[:border_width]=0
img[-border_width:]=0
img[:,:border_width]=0
img[:,-border_width:]=0
# The tw.calc_T_fwd (or tw.calc_T_inv) is always done in gpu.
# After using it to compute new pts,
# you may want to use remap (to warp an image accordingly).
# If you will use tw.remap_fwd (or tw.remap_inv), which is done in gpu,
# then the image type can be either float32 or float64.
# But if you plan to use tw.tw.remap_fwd_opencv (or tw.remap_inv_opencv),
# which is done in cpu (hence slightly lower) but supports better
# interpolation methods, then the image type must be np.float32.
# img_original = CpuGpuArray(img.copy().astype(np.float32))
img_original = CpuGpuArray(img.copy().astype(np.float64))
img_wrapped_fwd= CpuGpuArray.zeros_like(img_original)
img_wrapped_bwd= CpuGpuArray.zeros_like(img_original)
seed=0
np.random.seed(seed)
ms_Avees=tw.get_zeros_PA_all_levels()
ms_theta=tw.get_zeros_theta_all_levels()
for level in range(tw.ms.nLevels):
if level==0:
tw.sample_gaussian(level,ms_Avees[level],ms_theta[level],mu=None)# zero mean
else:
tw.sample_from_the_ms_prior_coarse2fine_one_level(ms_Avees,ms_theta,
level_fine=level)
print('\nimg shape: {}\n'.format(img_original.shape))
# You don't have use these. You can use any 2d array
# that has two columns (regardless of the number of rows).
pts_src = tw.pts_src_dense
# Create buffers for the output
pts_fwd = CpuGpuArray.zeros_like(pts_src)
pts_inv = CpuGpuArray.zeros_like(pts_src)
for level in range(tw.ms.nLevels):
#######################################################################
# instead of the tw.sample_from_the_ms_prior() above,
# you may want to use one of the following.
# 1)
# tw.sample_gaussian(level,ms_Avees[level],ms_theta[level],mu=None)# zero mean
# 2)
# tw.sample_gaussian(level,ms_Avees[level],ms_theta[level],mu=some_user_specified_mu)
# The following should be used only for level>0 :
# 3)
# tw.sample_normal_in_one_level_using_the_coarser_as_mean(Avees_coarse=ms_Avees[level-1],
# Avees_fine=ms_Avees[level],
# theta_fine=ms_theta[level],
# level_fine=level)
#
#######################################################################
# You can also change the values this way:
# cpa_space = tw.ms.L_cpa_space[level]
# theta = cpa_space.get_zeros_theta()
# theta[:] = some values
# Avees = cpa_space.get_zeros_PA()
# cpa_space.theta2Avees(theta,Avees)
# cpa_space.update_pat(Avees)
# This step is important and must be done
# before are trying to "use" the new values of
# the (vectorized) A's.
tw.update_pat_from_Avees(ms_Avees[level],level)
if eval_v:
# Evaluating the velocity field.
# You don't have to do it in unless you want to visualize v.
# (when evaluting the treansformation, v will be internally
# evaluated anyway -- but its result won't be stored)
tw.calc_v(level=level)
# optional, if you want to time it
timer_gpu_T_fwd = GpuTimer()
# Simply calling
# tic = time.clock()
# and then
# tic = time.clock()
# won't work.
# In fact, most likely you will get that toc-tic is zero.
# You need to use the GpuTimer object. When you do that,
# one side effect is that suddenly the toc-tic from above will
# give you a more realistic result.
tic = time.clock()
timer_gpu_T_fwd.tic()
tw.calc_T_fwd(pts_src,pts_fwd,level=level)
timer_gpu_T_fwd.toc()
toc = time.clock()
print 'Time, in sec, for computing T_fwd:'
print timer_gpu_T_fwd.secs
print toc-tic # likely to be 0, unless you also used the GpuTimer.
# You can also time the inv of course. Results will be similar.
tw.calc_T_inv(pts_src,pts_inv,level=level)
if eval_cell_idx:
# cell_idx is computed here just for display.
cell_idx = CpuGpuArray.zeros(len(pts_src),dtype=np.int32)
tw.calc_cell_idx(pts_src,cell_idx,level,
permute_for_disp=permute_cell_idx_for_display)
# If may also want ro to time the remap.
# However, the remap is usually very fast (e.g, about 2 milisec).
# timer_gpu_remap_fwd = GpuTimer()
# tic = time.clock()
# timer_gpu_remap_fwd.tic()
# tw.remap_fwd(pts_inv=pts_inv,img=img_original,img_wrapped_fwd=img_wrapped_fwd)
tw.remap_fwd(pts_inv=pts_inv,img=img_original,img_wrapped_fwd=img_wrapped_fwd)
# timer_gpu_remap_fwd.toc()
# toc = time.clock()
# If the img type is np.float32, you may also use
# tw.remap_fwd_opencv instead of tw.remap_fw. The differences between
# the two methods are explained above
tw.remap_inv(pts_fwd=pts_fwd,img=img_original,img_wrapped_inv=img_wrapped_bwd)
# For display purposes, do gpu2cpu transfer
print ("For display purposes, do gpu2cpu transfer")
if eval_cell_idx:
cell_idx.gpu2cpu()
if eval_v:
tw.v_dense.gpu2cpu()
pts_fwd.gpu2cpu()
pts_inv.gpu2cpu()
img_wrapped_fwd.gpu2cpu()
img_wrapped_bwd.gpu2cpu()
figsize = (12,12)
plt.figure(figsize=figsize)
if eval_v:
plt.subplot(332)
tw.imshow_vx()
plt.title('vx')
plt.subplot(333)
tw.imshow_vy()
plt.title('vy')
if eval_cell_idx:
plt.subplot(331)
cell_idx_disp = cell_idx.cpu.reshape(img.shape[0],-1)
plt.imshow(cell_idx_disp)
plt.title('tess (type {})'.format(tess))
if show_downsampled_pts:
ds=20
pts_src_grid = pts_src.cpu.reshape(tw.nRows,-1,2)
pts_src_ds=pts_src_grid[::ds,::ds].reshape(-1,2)
pts_fwd_grid = pts_fwd.cpu.reshape(tw.nRows,-1,2)
pts_fwd_ds=pts_fwd_grid[::ds,::ds].reshape(-1,2)
pts_inv_grid = pts_inv.cpu.reshape(tw.nRows,-1,2)
pts_inv_ds=pts_inv_grid[::ds,::ds].reshape(-1,2)
use_lims=use_lims_when_plotting
# return tw
plt.subplot(334)
plt.plot(pts_src_ds[:,0],pts_src_ds[:,1],'r.')
plt.title('pts ds')
tw.config_plt()
plt.subplot(335)
plt.plot(pts_fwd_ds[:,0],pts_fwd_ds[:,1],'g.')
plt.title('fwd(pts)')
tw.config_plt(axis_on_or_off='on',use_lims=use_lims)
plt.subplot(336)
plt.plot(pts_inv_ds[:,0],pts_inv_ds[:,1],'b.')
plt.title('inv(pts)')
tw.config_plt(axis_on_or_off='on',use_lims=use_lims)
plt.subplot(337)
plt.imshow(img_original.cpu.astype(np.uint8))
plt.title('img')
# plt.axis('off')
plt.subplot(338)
plt.imshow(img_wrapped_fwd.cpu.astype(np.uint8))
# plt.axis('off')
plt.title('fwd(img)')
plt.subplot(339)
plt.imshow(img_wrapped_bwd.cpu.astype(np.uint8))
# plt.axis('off')
plt.title('inv(img)')
return tw
msg="""
The code below is for landmarks,
not signals"""
raise NotImplementedError(msg)
yy,xx = np.mgrid[-2:2:1,-2:2:1]
x = np.vstack([xx.ravel(),yy.ravel()]).T
del xx,yy
x = CpuGpuArray(x.copy().astype(np.float))
print x
y = np.random.standard_normal(x.shape)
y = CpuGpuArray(y)
err = CpuGpuArray.zeros_like(y)
nPts = len(err)
ll = CpuGpuArray.zeros(nPts)
calc_signal_err_per_sample(x.gpu,y.gpu,err.gpu)
sigma=1.0
calc_ll_per_sample(ll.gpu,err.gpu,sigma)
err.gpu2cpu()
ll.gpu2cpu()
print np.allclose( ll.cpu,
-0.5*(err.cpu[:,0]**2+err.cpu[:,1]**2)/(sigma**2))
# np.random.seed(0)
theta = np.random.multivariate_normal(mean=mu,cov=cpa_covs.cpa_cov)
theta *=100
##
# theta.fill(0)
# theta[8]=1
# theta[10]=1
cpa_space.theta2Avees(theta=theta)
cpa_space.update_pat()
# 1/0
# params_flow_int.nTimeSteps *= 10
cell_idx = CpuGpuArray.zeros(len(pts),dtype=np.int32)
cpa_space.calc_cell_idx(pts,cell_idx)
cell_idx.gpu2cpu()
print cell_idx
img=cell_idx.cpu.reshape(cpa_space.x_dense_grid.shape[1:])
# img = pts.cpu[:,0].reshape(cpa_space.x_dense_grid.shape[1:])
plt.figure(1)
of.plt.set_figure_size_and_location(0,0,800,800)
plt.clf()
plt.subplot(131)
plt.imshow(img[:,:,0,0],interpolation="None");plt.colorbar()
v = CpuGpuArray.zeros_like(pts)
def __init__(
self,
ms,
level,
data,
sigma_signal,
params_flow_int,
interp_type_for_ll,
# src=None,dst=None,transformed=None
):
# ipshell('hi')
if interp_type_for_ll not in self.supported_interp_types:
msg = """
interp_type_for_ll must be in
['gpu_linear',
cv2.INTER_LINEAR,
cv2.INTER_CUBIC,
cv2.INTER_LANCZOS4]
"""
raise ValueError(msg, interp_type_for_ll)
self.interp_type_for_ll = interp_type_for_ll
src = data['src']
transformed = data['transformed']
signal = data['signal']
for obj in [src, transformed]:
if not isinstance(obj, CpuGpuArray):
raise TypeError
for obj in [signal.src, signal.dst, signal.transformed]:
if not isinstance(obj, CpuGpuArray):
raise TypeError
self.nCalls = 0
self.nCallbacks = 0
self.sigma_signal = sigma_signal
cpa_space = ms.L_cpa_space[level]
self.cpa_space = cpa_space
if src.shape[1] != cpa_space.dim_domain:
raise ValueError(src.shape, cpa_space.dim_domain)
# self.mu = cpa_simple_mean(cpa_space)
self.my = cpa_space.get_zeros_theta()
self.src = src
self.transformed = transformed
self.signal = signal
# self.dim_signal = signal.src.shape[1]
if signal.src.ndim == 2:
self.dim_signal = 2
else:
raise NotImplementedError
if self.dim_signal != 2:
raise NotImplementedError(signal.src.shape)
if self.signal.src.shape != self.signal.dst.shape:
raise ValueError
if self.signal.src.shape != self.signal.transformed.shape:
raise ValueError
nPts = len(src)
self.nPts = nPts
# self.err = CpuGpuArray.zeros_like(src)
# self.signal.err = CpuGpuArray.zeros_like(src)
self.signal.err = CpuGpuArray.zeros_like(self.signal.src)
self.ll = CpuGpuArray.zeros(nPts, dtype=src.dtype)
if nPts <= 1:
raise ValueError
self.params_flow_int = params_flow_int
self._pat = PAT(pa_space=cpa_space, Avees=cpa_space.get_zeros_PA())
def __init__(self,XMINS,XMAXS,nCs,
zero_v_across_bdry,
vol_preserve,
warp_around=None,
conformal=False,
zero_vals=None,
cpa_calcs=None,
tess='II',
valid_outside=None,
only_local=False,
cont_constraints_are_separable=False):
if conformal:
raise ValueError("This was a bad idea")
if not self.has_GPU:
raise ValueError("Where is my gpu?")
if conformal:
raise ValueError
if tess not in ['I','II']:
raise ValueError(tess)
if tess == 'I' and self.dim_domain == 1:
raise ValueError
if tess == 'I' and self.dim_domain not in (2,3):
raise NotImplementedError
if only_local and tess != 'I':
raise NotImplementedError
if zero_vals is None:
raise ValueError
if cpa_calcs is None:
raise ValueError("You must pass this argument")
self._calcs = cpa_calcs
if len(nCs) != self.dim_domain:
raise ValueError('len(nCs) = {0} =/= {1} = dim_domain'.format(len(nCs),self.dim_domain))
if warp_around is None:
# warp_around = [False] * self.dim_domain
raise ValueError("You must pass this argument")
try:# Test if iterable
zero_vals.__iter__
except AttributeError:
raise ValueError(zero_vals)
try: # Test if iterable
nCs.__iter__
except AttributeError:
raise ValueError(nCs)
try: # Test if iterable
zero_v_across_bdry.__iter__
except:
raise
try: # Test if iterable
warp_around.__iter__
except:
raise
if len(warp_around) != self.dim_domain:
raise ValueError(len(warp_around) , self.dim_domain)
if len(zero_v_across_bdry) != self.dim_domain:
raise ValueError(len(zero_v_across_bdry) , self.dim_domain)
if tess=='I':
if self.dim_domain==2:
if any(zero_v_across_bdry) and valid_outside:
raise ValueError("dim_domain==2","tess='I'",
"zero_v_across_bdry={}".format(zero_v_across_bdry),
"valid_outside={}".format(valid_outside),
"These choices are inconsistent with each other")
if not all(zero_v_across_bdry) and not valid_outside:
raise ValueError("dim_domain>1","tess='I'",
"zero_v_across_bdry={}".format(zero_v_across_bdry),
"valid_outside={}".format(valid_outside),
"These choices are inconsistent with each other")
elif self.dim_domain==3:
if valid_outside:
raise NotImplementedError
elif not all(zero_v_across_bdry):
raise ValueError("dim_domain==3","tess='I'",
"zero_v_across_bdry={}".format(zero_v_across_bdry),
"These choices are inconsistent with each other")
else:
raise NotImplementedError
self.XMINS = np.asarray(XMINS,dtype=my_dtype)
self.XMAXS = np.asarray(XMAXS,dtype=my_dtype)
if (self.XMINS>=self.XMAXS).any():
raise ValueError(XMINS,XMAXS)
self.warp_around = warp_around
self.tess=tess
if tess == 'II':
nC = reduce(np.dot,nCs) # of cells
elif tess == 'I':
if self.dim_domain == 2:
nC = reduce(np.dot,nCs) * 4
elif self.dim_domain == 3:
nC = reduce(np.dot,nCs) * 5
else:
raise NotImplementedError
else:
raise ValueError(tess)
self.nCs = np.asarray(nCs)
self.nC=nC
if self.dim_domain !=1:
if self.dim_domain in (2,3):
self.expm_eff = ExpmEff(nC)
else:
self.expm_eff = ExpmEff(nC,use_parallel=1)
nHomoCoo=self.nHomoCoo
self._signed_sqAs_times_dt= np.empty((nC,nHomoCoo,nHomoCoo),
dtype=self.my_dtype)
# In each matrix, fill last row with zeros
# self._signed_sqAs_times_dt[:,-1].fill(0)
# self._sqAs_vectorized = np.zeros((nC,nHomoCoo*nHomoCoo),
# dtype=self.my_dtype)
# self._Tlocals_vectorized = np.empty((nC,nHomoCoo*nHomoCoo),dtype=self.my_dtype)
self._As_vectorized = CpuGpuArray.zeros((nC,self.lengthAvee),dtype=self.my_dtype)
self._signed_As_vectorized = CpuGpuArray.zeros((nC,self.lengthAvee),dtype=self.my_dtype)
self._signed_As_times_dt_vectorized = CpuGpuArray.zeros((nC,self.lengthAvee),dtype=self.my_dtype)
self._Tlocals_vectorized = CpuGpuArray.zeros((nC,self.lengthAvee),dtype=self.my_dtype)
if self.has_GPU:
self.sharedmemory = decide_sharedmemory(self.dim_domain,
self.dim_range,
self.nC)
self._gpu_calcs = GpuCalcs(nC,my_dtype,
dim_domain=self.dim_domain,
dim_range=self.dim_range,
tess=self.tess,
sharedmemory=self.sharedmemory)
else:
raise NotImplementedError
self.only_local=only_local
self.zero_v_across_bdry=zero_v_across_bdry
self.vol_preserve=vol_preserve
self.subspace_string=self.create_subspace_string(self.XMINS,
self.XMAXS,
nCs,
zero_v_across_bdry,
vol_preserve,
warp_around,
conformal,
zero_vals,
valid_outside=valid_outside,
cont_constraints_are_separable=cont_constraints_are_separable)
self.directory = os.path.join(dirnames.cpa,'{0}d'.format(self.dim_domain),
self.subspace_string)
FilesDirs.mkdirs_if_needed(self.directory)
if self.only_local:
self.filename_subspace = os.path.join(self.directory,'local.pkl')
else:
self.filename_subspace = os.path.join(self.directory,'subspace.pkl')
def example(tess='I',
base=[2, 2, 2],
nLevels=1,
zero_v_across_bdry=[True] * 3,
vol_preserve=False,
nRows=100,
nCols=100,
nSlices=100,
use_mayavi=False,
eval_v=False,
eval_cell_idx=False):
tw = TransformWrapper(nRows=nRows,
nCols=nCols,
nSlices=nSlices,
nLevels=nLevels,
base=base,
zero_v_across_bdry=zero_v_across_bdry,
tess=tess,
valid_outside=False,
only_local=False,
vol_preserve=vol_preserve)
print_iterable(tw.ms.L_cpa_space)
print tw
# create some fake 3D image.
img = np.zeros((nCols, nRows, nSlices), dtype=np.float64)
# img[:]=np.random.random_integers(0,255,img.shape)
# Fill the image with the x coordinates as fake values
img[:] = tw.pts_src_dense.cpu[:, 0].reshape(img.shape)
img0 = CpuGpuArray(img.copy().astype(np.float64))
img_wrapped_fwd = CpuGpuArray.zeros_like(img0)
img_wrapped_inv = CpuGpuArray.zeros_like(img0)
seed = 0
np.random.seed(seed)
ms_Avees = tw.get_zeros_PA_all_levels()
ms_theta = tw.get_zeros_theta_all_levels()
if tess == 'II':
for level in range(tw.ms.nLevels):
cpa_space = tw.ms.L_cpa_space[level]
Avees = ms_Avees[level]
# 1/0
if level == 0:
tw.sample_gaussian(level,
ms_Avees[level],
ms_theta[level],
mu=None) # zero mean
# ms_theta[level].fill(0)
# ms_theta[level][-4]=10
cpa_space.theta2Avees(theta=ms_theta[level], Avees=Avees)
else:
tw.sample_from_the_ms_prior_coarse2fine_one_level(
ms_Avees, ms_theta, level_fine=level)
else:
# For tess='I' in 3D, I have yet to implement the coarse-to-fine sampling.
for level in range(tw.ms.nLevels):
cpa_space = tw.ms.L_cpa_space[level]
velTess = cpa_space.zeros_velTess()
ms_Avees[level].fill(0)
Avees = ms_Avees[level]
tw.sample_gaussian_velTess(level, Avees, velTess, mu=None)
print 'img shape:', img0.shape
# You don't have use these. You can use any 2d array
# that has 3 columns (regardless of the number of rows).
pts_src = tw.pts_src_dense
pts_src = CpuGpuArray(pts_src.cpu[::1].copy())
# Create a buffer for the output
pts_fwd = CpuGpuArray.zeros_like(pts_src)
pts_inv = CpuGpuArray.zeros_like(pts_src)
for level in range(tw.ms.nLevels):
tw.update_pat_from_Avees(ms_Avees[level], level)
if eval_v:
# Evaluating the velocity field.
# You don't have to do it in unless you want to visualize v.
# (when evaluting the treansformation, v will be internally
# evaluated anyway -- but its result won't be stored)
tw.calc_v(level=level)
print 'level', level
print
print 'number of points:', len(pts_src)
print 'number of cells:', tw.ms.L_cpa_space[level].nC
# optional, if you want to time it
timer_gpu_T_fwd = GpuTimer()
# Simply calling
# tic = time.clock()
# and then
# tic = time.clock()
# won't work.
# In fact, most likely you will get that toc-tic is zero.
# You need to use the GpuTimer object. When you do that,
# one side effect is that suddenly the toc-tic from above will
# give you a more realistic result.
tic = time.clock()
timer_gpu_T_fwd.tic()
tw.calc_T_fwd(pts_src, pts_fwd, level=level)
timer_gpu_T_fwd.toc()
toc = time.clock()
print 'Time, in sec, for computing T_fwd:'
print timer_gpu_T_fwd.secs
print toc - tic # likely to be 0, unless you also used the GpuTimer.
# You can also time the inv of course. Results will be similar.
tw.calc_T_inv(pts_src, pts_inv, level=level)
if eval_cell_idx:
# cell_idx is computed here just for display.
cell_idx = CpuGpuArray.zeros(len(pts_src), dtype=np.int32)
tw.calc_cell_idx(pts_src, cell_idx, level)
tw.remap_fwd(pts_inv, img0, img_wrapped_fwd)
tw.remap_inv(pts_fwd, img0, img_wrapped_inv)
# For display purposes, do gpu2cpu transfer
print "For display purposes, do gpu2cpu transfer"
if eval_cell_idx:
cell_idx.gpu2cpu()
if eval_v:
tw.v_dense.gpu2cpu()
pts_fwd.gpu2cpu()
pts_inv.gpu2cpu()
img_wrapped_fwd.gpu2cpu()
img_wrapped_inv.gpu2cpu()
if use_mayavi:
ds = 1 # downsampling factor
i = 17
pts_src_grid = pts_src.cpu.reshape(tw.nRows, tw.nCols, -1, 3)
pts_src_ds = pts_src_grid[::ds, ::ds, i].reshape(-1, 3)
pts_fwd_grid = pts_fwd.cpu.reshape(tw.nRows, tw.nCols, -1, 3)
pts_fwd_ds = pts_fwd_grid[::ds, ::ds, i].reshape(-1, 3)
pts_inv_grid = pts_inv.cpu.reshape(tw.nRows, tw.nCols, -1, 3)
pts_inv_ds = pts_inv_grid[::ds, ::ds, i].reshape(-1, 3)
from of.my_mayavi import *
mayavi_mlab_close_all()
mayavi_mlab_figure_bgwhite('src')
x, y, z = pts_src_ds.T
mayavi_mlab_plot3d(x, y, z)
mayavi_mlab_figure_bgwhite('fwd')
x, y, z = pts_fwd_ds.T
mayavi_mlab_plot3d(x, y, z)
figsize = (12, 12)
plt.figure(figsize=figsize)
i = 17 # some slice
plt.subplot(131)
plt.imshow(img0.cpu[:, :, i].astype(np.uint8), interpolation="Nearest")
plt.title('slice from img')
plt.subplot(132)
plt.imshow(img_wrapped_fwd.cpu[:, :, i].astype(np.uint8),
interpolation="Nearest")
plt.axis('off')
plt.title('slice from fwd(img)')
plt.subplot(133)
plt.imshow(img_wrapped_inv.cpu[:, :, i].astype(np.uint8),
interpolation="Nearest")
plt.axis('off')
plt.title('slice from inv(img)')
if 0: # debug
cpa_space = tw.ms.L_cpa_space[level]
if eval_v:
vx = tw.v_dense.cpu[:, 0].reshape(
cpa_space.x_dense_grid_img.shape[1:])
vy = tw.v_dense.cpu[:, 1].reshape(
cpa_space.x_dense_grid_img.shape[1:])
vz = tw.v_dense.cpu[:, 2].reshape(
cpa_space.x_dense_grid_img.shape[1:])
plt.figure()
plt.imshow(vz[:, :, 17], interpolation="Nearest")
plt.colorbar()
plt.title('vz in some slice')
return tw
def example(tess='I',base=[2,2,2],nLevels=1,
zero_v_across_bdry=[True]*3,
vol_preserve=False,
nRows=100, nCols=100,nSlices=100,
use_mayavi=False,
eval_v=False,
eval_cell_idx=False):
tw = TransformWrapper(nRows=nRows,
nCols=nCols,
nSlices=nSlices,
nLevels=nLevels,
base=base,
zero_v_across_bdry=zero_v_across_bdry,
tess=tess,
valid_outside=False,
only_local=False,
vol_preserve=vol_preserve)
print_iterable(tw.ms.L_cpa_space)
print tw
# create some fake 3D image.
img = np.zeros((nCols,nRows,nSlices),dtype=np.float64)
# img[:]=np.random.random_integers(0,255,img.shape)
# Fill the image with the x coordinates as fake values
img[:]=tw.pts_src_dense.cpu[:,0].reshape(img.shape)
img0 = CpuGpuArray(img.copy().astype(np.float64))
img_wrapped_fwd= CpuGpuArray.zeros_like(img0)
img_wrapped_inv= CpuGpuArray.zeros_like(img0)
seed=0
np.random.seed(seed)
ms_Avees=tw.get_zeros_PA_all_levels()
ms_theta=tw.get_zeros_theta_all_levels()
if tess == 'II' :
for level in range(tw.ms.nLevels):
cpa_space = tw.ms.L_cpa_space[level]
Avees = ms_Avees[level]
# 1/0
if level==0:
tw.sample_gaussian(level,ms_Avees[level],ms_theta[level],mu=None)# zero mean
# ms_theta[level].fill(0)
# ms_theta[level][-4]=10
cpa_space.theta2Avees(theta=ms_theta[level],Avees=Avees)
else:
tw.sample_from_the_ms_prior_coarse2fine_one_level(ms_Avees,ms_theta,
level_fine=level)
else:
# For tess='I' in 3D, I have yet to implement the coarse-to-fine sampling.
for level in range(tw.ms.nLevels):
cpa_space = tw.ms.L_cpa_space[level]
velTess = cpa_space.zeros_velTess()
ms_Avees[level].fill(0)
Avees = ms_Avees[level]
tw.sample_gaussian_velTess(level,Avees,velTess,mu=None)
print 'img shape:',img0.shape
# You don't have use these. You can use any 2d array
# that has 3 columns (regardless of the number of rows).
pts_src = tw.pts_src_dense
pts_src=CpuGpuArray(pts_src.cpu[::1].copy())
# Create a buffer for the output
pts_fwd = CpuGpuArray.zeros_like(pts_src)
pts_inv = CpuGpuArray.zeros_like(pts_src)
for level in range(tw.ms.nLevels):
tw.update_pat_from_Avees(ms_Avees[level],level)
if eval_v:
# Evaluating the velocity field.
# You don't have to do it in unless you want to visualize v.
# (when evaluting the treansformation, v will be internally
# evaluated anyway -- but its result won't be stored)
tw.calc_v(level=level)
print 'level',level
print
print 'number of points:',len(pts_src)
print 'number of cells:',tw.ms.L_cpa_space[level].nC
# optional, if you want to time it
timer_gpu_T_fwd = GpuTimer()
# Simply calling
# tic = time.clock()
# and then
# tic = time.clock()
# won't work.
# In fact, most likely you will get that toc-tic is zero.
# You need to use the GpuTimer object. When you do that,
# one side effect is that suddenly the toc-tic from above will
# give you a more realistic result.
tic = time.clock()
timer_gpu_T_fwd.tic()
tw.calc_T_fwd(pts_src,pts_fwd,level=level)
timer_gpu_T_fwd.toc()
toc = time.clock()
print 'Time, in sec, for computing T_fwd:'
print timer_gpu_T_fwd.secs
print toc-tic # likely to be 0, unless you also used the GpuTimer.
# You can also time the inv of course. Results will be similar.
tw.calc_T_inv(pts_src,pts_inv,level=level)
if eval_cell_idx:
# cell_idx is computed here just for display.
cell_idx = CpuGpuArray.zeros(len(pts_src),dtype=np.int32)
tw.calc_cell_idx(pts_src,cell_idx,level)
tw.remap_fwd(pts_inv,img0,img_wrapped_fwd)
tw.remap_inv(pts_fwd,img0,img_wrapped_inv)
# For display purposes, do gpu2cpu transfer
print "For display purposes, do gpu2cpu transfer"
if eval_cell_idx:
cell_idx.gpu2cpu()
if eval_v:
tw.v_dense.gpu2cpu()
pts_fwd.gpu2cpu()
pts_inv.gpu2cpu()
img_wrapped_fwd.gpu2cpu()
img_wrapped_inv.gpu2cpu()
if use_mayavi:
ds=1 # downsampling factor
i= 17
pts_src_grid = pts_src.cpu.reshape(tw.nRows,tw.nCols,-1,3)
pts_src_ds=pts_src_grid[::ds,::ds,i].reshape(-1,3)
pts_fwd_grid = pts_fwd.cpu.reshape(tw.nRows,tw.nCols,-1,3)
pts_fwd_ds=pts_fwd_grid[::ds,::ds,i].reshape(-1,3)
pts_inv_grid = pts_inv.cpu.reshape(tw.nRows,tw.nCols,-1,3)
pts_inv_ds=pts_inv_grid[::ds,::ds,i].reshape(-1,3)
from of.my_mayavi import *
mayavi_mlab_close_all()
mayavi_mlab_figure_bgwhite('src')
x,y,z=pts_src_ds.T
mayavi_mlab_plot3d(x,y,z)
mayavi_mlab_figure_bgwhite('fwd')
x,y,z=pts_fwd_ds.T
mayavi_mlab_plot3d(x,y,z)
figsize = (12,12)
plt.figure(figsize=figsize)
i= 17 # some slice
plt.subplot(131)
plt.imshow(img0.cpu[:,:,i].astype(np.uint8),interpolation="Nearest")
plt.title('slice from img')
plt.subplot(132)
plt.imshow(img_wrapped_fwd.cpu[:,:,i].astype(np.uint8),interpolation="Nearest")
plt.axis('off')
plt.title('slice from fwd(img)')
plt.subplot(133)
plt.imshow(img_wrapped_inv.cpu[:,:,i].astype(np.uint8),interpolation="Nearest")
plt.axis('off')
plt.title('slice from inv(img)')
if 0: # debug
cpa_space=tw.ms.L_cpa_space[level]
if eval_v:
vx=tw.v_dense.cpu[:,0].reshape(cpa_space.x_dense_grid_img.shape[1:])
vy=tw.v_dense.cpu[:,1].reshape(cpa_space.x_dense_grid_img.shape[1:])
vz=tw.v_dense.cpu[:,2].reshape(cpa_space.x_dense_grid_img.shape[1:])
plt.figure()
plt.imshow(vz[:,:,17],interpolation="Nearest");plt.colorbar()
plt.title('vz in some slice')
return tw
from pycuda import autoinit
from of.gpu import CpuGpuArray
import numpy as np
msg = """
The code below is for landmarks,
not signals"""
raise NotImplementedError(msg)
yy, xx = np.mgrid[-2:2:1, -2:2:1]
x = np.vstack([xx.ravel(), yy.ravel()]).T
del xx, yy
x = CpuGpuArray(x.copy().astype(np.float))
print x
y = np.random.standard_normal(x.shape)
y = CpuGpuArray(y)
err = CpuGpuArray.zeros_like(y)
nPts = len(err)
ll = CpuGpuArray.zeros(nPts)
calc_signal_err_per_sample(x.gpu, y.gpu, err.gpu)
sigma = 1.0
calc_ll_per_sample(ll.gpu, err.gpu, sigma)
err.gpu2cpu()
ll.gpu2cpu()
print np.allclose(
ll.cpu, -0.5 * (err.cpu[:, 0]**2 + err.cpu[:, 1]**2) / (sigma**2))
def __init__(self, nSuperpixels, s_std, i_std, prior_count, nChannels):
"""
Initilize the parameters for the superpixels:
The means are set to zeros at this point,
and will be set later in the first M step.
The space/color covariances (and their inverse), however, are being set
to initial values here.
We use a Inverse-Wishart prior on the space covariance
Arguments:
nSuperpixels: the number of superpixels to generate
s_std: should be fixed as nPixels_on_side
i_std: control the relative importance between RGB and location.
The smaller it is, bigger the RGB effect is / more irregular the superpixels are.
prior_count: determines the weight of Inverse-Wishart prior of space covariance(ex:1,5,10)
nChannels: the number of channels of the input image (gray:1, LAB/RGB: 3)
"""
if nChannels not in (1,3):
raise NotImplementedError(nChannels)
dim_i=nChannels
dim_s=2
self.dim_i=dim_i
self.dim_s=dim_s
self.nSuperpixels=nSuperpixels
self.s_std, self.i_std, self.prior_count = s_std,i_std,prior_count
mu_s = CpuGpuArray.zeros((nSuperpixels,dim_s))
mu_i = CpuGpuArray.zeros((nSuperpixels,dim_i))
Sigma_s = CpuGpuArray.zeros(shape = (nSuperpixels,dim_s,dim_s))
J_s = CpuGpuArray.zeros_like(Sigma_s)
Sigma_i = CpuGpuArray.zeros((nSuperpixels,dim_i,dim_i))
J_i = CpuGpuArray.zeros_like(Sigma_i)
logdet_Sigma_i = CpuGpuArray.zeros((nSuperpixels,1)) # scalars
logdet_Sigma_s = CpuGpuArray.zeros((nSuperpixels,1))
# start with unnormalized counts (uniform)
counts = np.ones(nSuperpixels,dtype=np.int32)
counts = CpuGpuArray(counts)
self.params = Bunch()
self.params.mu_i = mu_i
self.params.mu_s = mu_s
self.params.Sigma_i = Sigma_i
self.params.Sigma_s = Sigma_s
self.params.prior_sigma_s_sum = Sigma_s
self.params.J_i = J_i
self.params.J_s = J_s
self.params.logdet_Sigma_i = logdet_Sigma_i
self.params.logdet_Sigma_s = logdet_Sigma_s
self.params.counts = counts
# set those parameters related to covariance
self.initialize_params()
# intermediate arrays needed for the Gaussian parameter calculation on GPU
self.gpu_helper = Bunch()
self.gpu_helper.mu_i_helper = gpuarray.zeros((nSuperpixels,dim_i),dtype=np.int32)
self.gpu_helper.mu_s_helper = gpuarray.zeros((nSuperpixels,dim_s),dtype=np.int32)
self.gpu_helper.prior_sigma_s = self.params.prior_sigma_s_sum.gpu.copy()
self.gpu_helper.sigma_s_helper = gpuarray.zeros((nSuperpixels,3),dtype=np.int64)
self.gpu_helper.log_count_helper = gpuarray.zeros((nSuperpixels,1),dtype=np.double)
def example(img=None,
tess='I',
eval_cell_idx=True,
eval_v=True,
show_downsampled_pts=True,
valid_outside=True,
base=[1, 1],
scale_spatial=.1,
scale_value=100,
permute_cell_idx_for_display=True,
nLevels=3,
vol_preserve=False,
zero_v_across_bdry=[0, 0],
use_lims_when_plotting=True):
show_downsampled_pts = bool(show_downsampled_pts)
eval_cell_idx = bool(eval_cell_idx)
eval_v = bool(eval_cell_idx)
valid_outside = bool(valid_outside)
permute_cell_idx_for_display = bool(permute_cell_idx_for_display)
vol_preserve = bool(vol_preserve)
if img is None:
img = Img(get_std_test_img())
else:
img = Img(img)
img = img[:, :, ::-1] # bgr2rgb
tw = TransformWrapper(
nRows=img.shape[0],
nCols=img.shape[1],
nLevels=nLevels,
base=base,
scale_spatial=scale_spatial, # controls the prior's smoothness
scale_value=scale_value, # controls the prior's variance
tess=tess,
vol_preserve=vol_preserve,
zero_v_across_bdry=zero_v_across_bdry,
valid_outside=valid_outside)
print tw
# You probably want to do that: padding image border with zeros
border_width = 1
img[:border_width] = 0
img[-border_width:] = 0
img[:, :border_width] = 0
img[:, -border_width:] = 0
# The tw.calc_T_fwd (or tw.calc_T_inv) is always done in gpu.
# After using it to compute new pts,
# you may want to use remap (to warp an image accordingly).
# If you will use tw.remap_fwd (or tw.remap_inv), which is done in gpu,
# then the image type can be either float32 or float64.
# But if you plan to use tw.tw.remap_fwd_opencv (or tw.remap_inv_opencv),
# which is done in cpu (hence slightly lower) but supports better
# interpolation methods, then the image type must be np.float32.
# img_original = CpuGpuArray(img.copy().astype(np.float32))
img_original = CpuGpuArray(img.copy().astype(np.float64))
img_wrapped_fwd = CpuGpuArray.zeros_like(img_original)
img_wrapped_bwd = CpuGpuArray.zeros_like(img_original)
seed = 0
np.random.seed(seed)
ms_Avees = tw.get_zeros_PA_all_levels()
ms_theta = tw.get_zeros_theta_all_levels()
for level in range(tw.ms.nLevels):
if level == 0:
tw.sample_gaussian(level,
ms_Avees[level],
ms_theta[level],
mu=None) # zero mean
else:
tw.sample_from_the_ms_prior_coarse2fine_one_level(ms_Avees,
ms_theta,
level_fine=level)
print('\nimg shape: {}\n'.format(img_original.shape))
# You don't have use these. You can use any 2d array
# that has two columns (regardless of the number of rows).
pts_src = tw.pts_src_dense
# Create buffers for the output
pts_fwd = CpuGpuArray.zeros_like(pts_src)
pts_inv = CpuGpuArray.zeros_like(pts_src)
for level in range(tw.ms.nLevels):
#######################################################################
# instead of the tw.sample_from_the_ms_prior() above,
# you may want to use one of the following.
# 1)
# tw.sample_gaussian(level,ms_Avees[level],ms_theta[level],mu=None)# zero mean
# 2)
# tw.sample_gaussian(level,ms_Avees[level],ms_theta[level],mu=some_user_specified_mu)
# The following should be used only for level>0 :
# 3)
# tw.sample_normal_in_one_level_using_the_coarser_as_mean(Avees_coarse=ms_Avees[level-1],
# Avees_fine=ms_Avees[level],
# theta_fine=ms_theta[level],
# level_fine=level)
#
#######################################################################
# You can also change the values this way:
# cpa_space = tw.ms.L_cpa_space[level]
# theta = cpa_space.get_zeros_theta()
# theta[:] = some values
# Avees = cpa_space.get_zeros_PA()
# cpa_space.theta2Avees(theta,Avees)
# cpa_space.update_pat(Avees)
# This step is important and must be done
# before are trying to "use" the new values of
# the (vectorized) A's.
tw.update_pat_from_Avees(ms_Avees[level], level)
if eval_v:
# Evaluating the velocity field.
# You don't have to do it in unless you want to visualize v.
# (when evaluting the treansformation, v will be internally
# evaluated anyway -- but its result won't be stored)
tw.calc_v(level=level)
# optional, if you want to time it
timer_gpu_T_fwd = GpuTimer()
# Simply calling
# tic = time.clock()
# and then
# tic = time.clock()
# won't work.
# In fact, most likely you will get that toc-tic is zero.
# You need to use the GpuTimer object. When you do that,
# one side effect is that suddenly the toc-tic from above will
# give you a more realistic result.
tic = time.clock()
timer_gpu_T_fwd.tic()
tw.calc_T_fwd(pts_src, pts_fwd, level=level)
timer_gpu_T_fwd.toc()
toc = time.clock()
print 'Time, in sec, for computing T_fwd:'
print timer_gpu_T_fwd.secs
print toc - tic # likely to be 0, unless you also used the GpuTimer.
# You can also time the inv of course. Results will be similar.
tw.calc_T_inv(pts_src, pts_inv, level=level)
if eval_cell_idx:
# cell_idx is computed here just for display.
cell_idx = CpuGpuArray.zeros(len(pts_src), dtype=np.int32)
tw.calc_cell_idx(pts_src,
cell_idx,
level,
permute_for_disp=permute_cell_idx_for_display)
# If may also want ro to time the remap.
# However, the remap is usually very fast (e.g, about 2 milisec).
# timer_gpu_remap_fwd = GpuTimer()
# tic = time.clock()
# timer_gpu_remap_fwd.tic()
# tw.remap_fwd(pts_inv=pts_inv,img=img_original,img_wrapped_fwd=img_wrapped_fwd)
tw.remap_fwd(pts_inv=pts_inv,
img=img_original,
img_wrapped_fwd=img_wrapped_fwd)
# timer_gpu_remap_fwd.toc()
# toc = time.clock()
# If the img type is np.float32, you may also use
# tw.remap_fwd_opencv instead of tw.remap_fw. The differences between
# the two methods are explained above
tw.remap_inv(pts_fwd=pts_fwd,
img=img_original,
img_wrapped_inv=img_wrapped_bwd)
# For display purposes, do gpu2cpu transfer
print("For display purposes, do gpu2cpu transfer")
if eval_cell_idx:
cell_idx.gpu2cpu()
if eval_v:
tw.v_dense.gpu2cpu()
pts_fwd.gpu2cpu()
pts_inv.gpu2cpu()
img_wrapped_fwd.gpu2cpu()
img_wrapped_bwd.gpu2cpu()
figsize = (12, 12)
plt.figure(figsize=figsize)
if eval_v:
plt.subplot(332)
tw.imshow_vx()
plt.title('vx')
plt.subplot(333)
tw.imshow_vy()
plt.title('vy')
if eval_cell_idx:
plt.subplot(331)
cell_idx_disp = cell_idx.cpu.reshape(img.shape[0], -1)
plt.imshow(cell_idx_disp)
plt.title('tess (type {})'.format(tess))
if show_downsampled_pts:
ds = 20
pts_src_grid = pts_src.cpu.reshape(tw.nRows, -1, 2)
pts_src_ds = pts_src_grid[::ds, ::ds].reshape(-1, 2)
pts_fwd_grid = pts_fwd.cpu.reshape(tw.nRows, -1, 2)
pts_fwd_ds = pts_fwd_grid[::ds, ::ds].reshape(-1, 2)
pts_inv_grid = pts_inv.cpu.reshape(tw.nRows, -1, 2)
pts_inv_ds = pts_inv_grid[::ds, ::ds].reshape(-1, 2)
use_lims = use_lims_when_plotting
# return tw
plt.subplot(334)
plt.plot(pts_src_ds[:, 0], pts_src_ds[:, 1], 'r.')
plt.title('pts ds')
tw.config_plt()
plt.subplot(335)
plt.plot(pts_fwd_ds[:, 0], pts_fwd_ds[:, 1], 'g.')
plt.title('fwd(pts)')
tw.config_plt(axis_on_or_off='on', use_lims=use_lims)
plt.subplot(336)
plt.plot(pts_inv_ds[:, 0], pts_inv_ds[:, 1], 'b.')
plt.title('inv(pts)')
tw.config_plt(axis_on_or_off='on', use_lims=use_lims)
plt.subplot(337)
plt.imshow(img_original.cpu.astype(np.uint8))
plt.title('img')
# plt.axis('off')
plt.subplot(338)
plt.imshow(img_wrapped_fwd.cpu.astype(np.uint8))
# plt.axis('off')
plt.title('fwd(img)')
plt.subplot(339)
plt.imshow(img_wrapped_bwd.cpu.astype(np.uint8))
# plt.axis('off')
plt.title('inv(img)')
return tw
