def get_solver(self, x_na, K_nn, bend_coefs, rot_coef): n,d = x_na.shape assert len(bend_coefs) <= len(self.bend_coefs) assert n <= self.max_N if not self.cur_solver is None: self.cur_solver.valid = False Q = np.c_[np.ones((n, 1)), x_na, K_nn] A = np.r_[np.zeros((d+1, d+1)), np.c_[np.ones((n, 1)), x_na]].T R = np.zeros((n+d+1, d)) R[1:d+1, :d] = np.diag(rot_coef) n_cnts = A.shape[0] _u,_s,_vh = np.linalg.svd(A.T) N = _u[:,n_cnts:].copy() N_gpu = self.N_gpu[:(n+d+1)*n].reshape(n+d+1, n) N_gpu.set_async(N) QN = Q.dot(N) QN_gpu = self.QN_gpu[:n*n].reshape(n, n) QN_gpu.set_async(QN) WQN_gpu = self.WQN_gpu[:n*n].reshape(n, n) NHN_gpu = self.NHN_gpu[:n*n].reshape(n, n) NR = N.T.dot(R) N_arr_gpu = [] O_gpu = [] ON_gpu = [] NON_gpu = [] for i, b in enumerate(bend_coefs): O_b = np.zeros((n+d+1, n+d+1), np.float64) O_b[d+1:, d+1:] += b * K_nn O_b[1:d+1, 1:d+1] += np.diag(rot_coef) offset = i * (n+d+1)*(n+d+1) O_gpu.append(self.O_gpu[offset:offset + (n+d+1)*(n+d+1)].reshape(n+d+1, n+d+1)) O_gpu[-1].set(O_b) offset = i * (n)*(n+d+1) ON_gpu.append(self.ON_gpu[offset:offset + n*(n+d+1)].reshape(n+d+1, n)) offset = i * n * n NON_gpu.append(self.NON_gpu[offset:offset + n*n].reshape(n, n)) N_arr_gpu.append(N_gpu) O_ptrs = get_gpu_ptrs(O_gpu) ON_ptrs = get_gpu_ptrs(ON_gpu) NON_ptrs = get_gpu_ptrs(NON_gpu) N_ptrs = get_gpu_ptrs(N_arr_gpu) dot_batch_nocheck(O_gpu, N_arr_gpu, ON_gpu, O_ptrs, N_ptrs, ON_ptrs, b = 0) dot_batch_nocheck(N_arr_gpu, ON_gpu, NON_gpu, N_ptrs, ON_ptrs, NON_ptrs, transa='T', b = 0) NON_gpu = dict(zip(bend_coefs, NON_gpu)) NON = dict([(b, non.get_async()) for b, non in NON_gpu.iteritems()]) self.cur_solver = TPSSolver(bend_coefs, N, QN, NON, NR, x_na, K_nn, rot_coef, QN_gpu, WQN_gpu, NON_gpu, NHN_gpu) return self.cur_solver
def transform_points(self): """ computes the warp of self.pts under the current tps params """ fill_mat(self.pt_w_ptrs, self.trans_d_ptrs, self.dims_gpu, self.N) dot_batch_nocheck(self.pts, self.lin_dd, self.pts_w, self.pt_ptrs, self.lin_dd_ptrs, self.pt_w_ptrs) dot_batch_nocheck(self.kernels, self.w_nd, self.pts_w, self.kernel_ptrs, self.w_nd_ptrs, self.pt_w_ptrs) sync()
def update_transform(self, b): """ computes the TPS associated with the current target pts """ self.set_tps_params(self.offset_mats[b]) dot_batch_nocheck( self.proj_mats[b], self.pts_t, self.tps_params, self.proj_mat_ptrs[b], self.pt_t_ptrs, self.tps_param_ptrs ) sync()
def bending_cost(self, b=DEFAULT_LAMBDA[1]): ## b * w_nd' * K * w_nd ## use pts_w as temporary storage dot_batch_nocheck(self.kernels, self.w_nd, self.pts_w, self.kernel_ptrs, self.w_nd_ptrs, self.pt_w_ptrs, b=0) dot_batch_nocheck( self.pts_w, self.w_nd, self.bend_res, self.pt_w_ptrs, self.w_nd_ptrs, self.bend_res_ptrs, transa="T", b=0 ) bend_res = self.bend_res_mat.get() return b * np.array([np.trace(bend_res[i * DATA_DIM : (i + 1) * DATA_DIM]) for i in range(self.N)])
def transform_trajs(self): """ computes the warp of l_traj and r_traj under current tps params """ fill_mat(self.l_traj_w_ptrs, self.trans_d_ptrs, self.l_traj_dims_gpu, self.N) fill_mat(self.r_traj_w_ptrs, self.trans_d_ptrs, self.r_traj_dims_gpu, self.N) dot_batch_nocheck( self.l_traj, self.lin_dd, self.l_traj_w, self.l_traj_ptrs, self.lin_dd_ptrs, self.l_traj_w_ptrs ) dot_batch_nocheck( self.r_traj, self.lin_dd, self.r_traj_w, self.r_traj_ptrs, self.lin_dd_ptrs, self.r_traj_w_ptrs ) dot_batch_nocheck( self.l_traj_K, self.w_nd, self.l_traj_w, self.l_traj_K_ptrs, self.w_nd_ptrs, self.l_traj_w_ptrs ) dot_batch_nocheck( self.r_traj_K, self.w_nd, self.r_traj_w, self.r_traj_K_ptrs, self.w_nd_ptrs, self.r_traj_w_ptrs ) sync()
def check_transform_pts(ctx, i=0): import scikits.cuda.linalg as la n = ctx.dims[i] w_nd = ctx.w_nd[i].get()[:n] lin_dd = ctx.lin_dd[i].get() trans_d = ctx.trans_d[i].get() k_nn = ctx.kernels[i].get()[:n, :n].reshape(n, n).copy() x_nd = ctx.pts[i].get()[:n] xw_nd = ctx.pts_w[i].get()[:n] _k_gpu = gpuarray.to_gpu(k_nn) _x_gpu = gpuarray.to_gpu(x_nd) _lin_gpu = gpuarray.to_gpu(lin_dd) _trans_gpu = gpuarray.to_gpu(trans_d) _w_gpu = gpuarray.to_gpu(w_nd) fill_mat(ctx.pt_w_ptrs, ctx.trans_d_ptrs, ctx.dims_gpu, ctx.N) dot_batch_nocheck(ctx.pts, ctx.lin_dd, ctx.pts_w, ctx.pt_ptrs, ctx.lin_dd_ptrs, ctx.pt_w_ptrs) xw_nd = ctx.pts_w[i].get()[:n] cpu_xw_nd = np.dot(x_nd, lin_dd) + trans_d[None, :] # assert np.allclose(xw_nd, cpu_xw_nd) dot_batch_nocheck(ctx.kernels, ctx.w_nd, ctx.pts_w, ctx.kernel_ptrs, ctx.w_nd_ptrs, ctx.pt_w_ptrs) xw_nd = ctx.pts_w[i].get()[:n] cpu_xw_nd = cpu_xw_nd + np.dot(k_nn, w_nd) # print "w_nd\n", w_nd[:3], np.max(w_nd) # print "lin_dd\n", lin_dd[:3] # print "trans_d\n", trans_d # print "k_nn\n", k_nn[:3, :3] # print "x_nd\n", x_nd[:3, :3] # print cpu_xw_nd[:3] if not (np.allclose(xw_nd, cpu_xw_nd)): print "k dot w_nd is difference on cpu and gpu" k_dot_w = np.dot(k_nn, w_nd) k_gpu = [gpuarray.to_gpu(k_nn)] w_gpu = [gpuarray.to_gpu(w_nd)] res_gpu = [gpuarray.zeros((n, DATA_DIM), np.float32)] k_ptrs = get_gpu_ptrs(k_gpu) w_ptrs = get_gpu_ptrs(w_gpu) res_ptrs = get_gpu_ptrs(res_gpu) dot_batch_nocheck(k_gpu, w_gpu, res_gpu, k_ptrs, w_ptrs, res_ptrs) res = res_gpu[0].get() single_gpu = la.dot(_k_gpu, _w_gpu) print "retry success {}".format(np.allclose(res, k_dot_w)) print "gpu success {}".format(np.allclose(single_gpu.get(), res)) assert np.allclose(single_gpu.get(), res) raw_input("go?")