def test_get_targ(self, other, T = 5e-3, outlierprior=1e-1, outlierfrac=1e-2, outliercutoff=1e-2, norm_iters = DEFAULT_NORM_ITERS):
self.transform_points()
other.transform_points()
init_prob_nm(self.pt_ptrs, other.pt_ptrs,
self.pt_w_ptrs, other.pt_w_ptrs,
self.dims_gpu, other.dims_gpu,
self.N, outlierprior, outlierfrac, T,
self.corr_cm_ptrs, self.corr_rm_ptrs)
norm_prob_nm(self.corr_cm_ptrs, self.corr_rm_ptrs,
self.dims_gpu, other.dims_gpu, self.N, outlierfrac, norm_iters,
self.r_coef_ptrs, self.c_coef_rn_ptrs, self.c_coef_cn_ptrs)
get_targ_pts(self.pt_ptrs, other.pt_ptrs,
self.pt_w_ptrs, other.pt_w_ptrs,
self.corr_cm_ptrs, self.corr_rm_ptrs,
self.r_coef_ptrs, self.c_coef_rn_ptrs, self.c_coef_cn_ptrs,
self.dims_gpu, other.dims_gpu,
outliercutoff, self.N,
self.pt_t_ptrs, other.pt_t_ptrs)
n, m = self.dims[0], other.dims[0]
x = self.pts[0].get()[:n]
xw = self.pts_w[0].get()[:n]
xt = self.pts_t[0].get()[:n]
y = other.pts[0].get()[:m]
yw = other.pts_w[0].get()[:m]
yt = other.pts_t[0].get()[:m]
init_corr = self.corr_rm[0].get()
init_corr = init_corr.flatten()[:(n + 1) * (m + 1)].reshape(n+1, m+1).astype(np.float32)
gpu_c_cn_coefs = self.c_coefs_cn[0].get().flatten()[:m + 1].reshape(m+1)
gpu_c_rn_coefs = self.c_coefs_rn[0].get().flatten()[:m + 1].reshape(m+1)
gpu_r_coefs = self.r_coefs[0].get().flatten()[:n + 1].reshape(n+1)
rn_corr = (init_corr * gpu_c_rn_coefs[None, :]) * gpu_r_coefs[:, None]
cn_corr = (init_corr * gpu_c_cn_coefs[None, :]) * gpu_r_coefs[:, None]
rn_corr = rn_corr[:n, :m]
cn_corr = cn_corr[:n, :m]
wt_n = rn_corr.sum(axis=1)
inlier = wt_n > outliercutoff
xtarg = np.empty((n, DATA_DIM))
xtarg[inlier, :] = rn_corr.dot(y)[inlier, :]
xtarg[~inlier, :] = xw[~inlier, :]
if not np.allclose(xtarg, xt):
print "xt values differ"
ipy.embed()
sys.exit(1)
wt_m = cn_corr.sum(axis=0)
inliner = wt_m > outliercutoff
ytarg = np.empty((m, DATA_DIM))
ytarg[inlier, :] = cn_corr.T.dot(x)[inliner, :]
ytarg[~inlier, :] = yw[~inlier, :]
if not np.allclose(ytarg, yt):
print "yt values differ"
ipy.embed()
sys.exit(1)
def get_target_points(
self, other, outlierprior=1e-1, outlierfrac=1e-2, outliercutoff=1e-2, T=5e-3, norm_iters=DEFAULT_NORM_ITERS
):
"""
computes the target points for self and other
using the current warped points for both
"""
init_prob_nm(
self.pt_ptrs,
other.pt_ptrs,
self.pt_w_ptrs,
other.pt_w_ptrs,
self.dims_gpu,
other.dims_gpu,
self.N,
outlierprior,
outlierfrac,
T,
self.corr_cm_ptrs,
self.corr_rm_ptrs,
)
sync()
norm_prob_nm(
self.corr_cm_ptrs,
self.corr_rm_ptrs,
self.dims_gpu,
other.dims_gpu,
self.N,
outlierfrac,
norm_iters,
self.r_coef_ptrs,
self.c_coef_rn_ptrs,
self.c_coef_cn_ptrs,
)
sync()
get_targ_pts(
self.pt_ptrs,
other.pt_ptrs,
self.pt_w_ptrs,
other.pt_w_ptrs,
self.corr_cm_ptrs,
self.corr_rm_ptrs,
self.r_coef_ptrs,
self.c_coef_rn_ptrs,
self.c_coef_cn_ptrs,
self.dims_gpu,
other.dims_gpu,
outliercutoff,
self.N,
self.pt_t_ptrs,
other.pt_t_ptrs,
)
sync()
def test_norm_corr(self, other, T = 5e-3, outlierprior=1e-1, outlierfrac=1e-2, outliercutoff=1e-2, norm_iters = DEFAULT_NORM_ITERS):
import sys
self.transform_points()
other.transform_points()
init_prob_nm(self.pt_ptrs, other.pt_ptrs,
self.pt_w_ptrs, other.pt_w_ptrs,
self.dims_gpu, other.dims_gpu,
self.N, outlierprior, outlierfrac, T,
self.corr_cm_ptrs, self.corr_rm_ptrs)
n, m = self.dims[0], other.dims[0]
init_corr = self.corr_rm[0].get()
init_corr = init_corr.flatten()[:(n + 1) * (m + 1)].reshape(n+1, m+1).astype(np.float32)
a_N = np.ones((n+1),dtype = np.float32)
a_N[n] = m*outlierfrac
b_M = np.ones((m+1), dtype = np.float32)
b_M[m] = n*outlierfrac
old_r_coefs = np.ones(n+1, dtype=np.float32)
old_c_coefs = np.ones(m+1, dtype=np.float32)
for n_iter in range(1, norm_iters):
init_prob_nm(self.pt_ptrs, other.pt_ptrs,
self.pt_w_ptrs, other.pt_w_ptrs,
self.dims_gpu, other.dims_gpu,
self.N, outlierprior, outlierfrac, T,
self.corr_cm_ptrs, self.corr_rm_ptrs)
norm_prob_nm(self.corr_cm_ptrs, self.corr_rm_ptrs,
self.dims_gpu, other.dims_gpu, self.N, outlierfrac, n_iter,
self.r_coef_ptrs, self.c_coef_rn_ptrs, self.c_coef_cn_ptrs)
new_r_coefs = a_N/init_corr.dot(old_c_coefs[:m+1])
new_c_coefs = b_M/new_r_coefs[:n+1].dot(init_corr)
gpu_c_coefs = self.c_coefs_cn[0].get().flatten()[:m + 1].reshape(m+1)
gpu_r_coefs = self.r_coefs[0].get().flatten()[:n + 1].reshape(n+1)
if not np.allclose(new_r_coefs, gpu_r_coefs):
print "row coeficients don't match", n_iter
ipy.embed()
sys.exit(1)
if not np.allclose(new_c_coefs, gpu_c_coefs):
print "column coeficients don't match", n_iter
ipy.embed()
sys.exit(1)
# old_r_coefs = gpu_r_coefs
# old_c_coefs = gpu_c_coefs
old_r_coefs = new_r_coefs
old_c_coefs = new_c_coefs
