def get_grads_reps_pca2_tangent(experiment, nreps, nsel, cores, projector):
experiments = {}
dim = experiment.dim
dimnoise = experiment.dimnoise
for i in range(nreps):
experiments[i] = copy.copy(experiment)
experiments[i].M.selected_points = np.random.choice(list(
range(experiment.n)),
nsel,
replace=False)
tangent_bases = experiments[i].M.get_wlpca_tangent_sel(
experiments[i].Mpca, experiments[i].M.selected_points, dimnoise)
subM = RiemannianManifold(
experiments[i].Mpca.data[experiments[i].M.selected_points], dim)
subM.tb = TangentBundle(subM, tangent_bases)
experiments[i].N.tangent_bundle = TangentBundle(
experiments[i].N,
experiments[i].N.geom.rmetric.embedding_eigenvectors)
#experiments[i].df_M = experiments[i].get_dF_js_idM(experiments[i].Mpca, experiments[i].N, subM.tb, experiments[i].N.tangent_bundle,
# experiments[i].M.selected_points,dimnoise)
#experiments[i].df_M = experiments[i].df_M / np.linalg.norm(experiments[i].df_M, axis=1).sum(axis=0)
#experiments[i].df_M2 = experiments[i].df_M / np.linalg.norm(experiments[i].df_M) ** 2
experiments[i].df_M = np.asarray(
[np.identity(dim) for i in range(nsel)])
experiments[i].dg_x = experiments[i].get_dx_g_full(
experiments[i].M.data[experiments[i].M.selected_points])
experiments[i].W = ShapeSpace(experiments[i].positions,
experiments[i].M.data)
experiments[i].dw = experiments[i].W.get_dw(
cores, experiments[i].atoms3, experiments[i].natoms,
experiments[i].M.selected_points)
experiments[i].dg_w = experiments[i].project(
np.swapaxes(experiments[i].dw, 1, 2),
experiments[i].project(experiments[i].dw, experiments[i].dg_x))
experiments[i].dg_w_pca = np.asarray([
np.matmul(projector,
experiments[i].dg_w[j].transpose()).transpose()
for j in range(nsel)
])
experiments[i].dgw_norm = experiments[i].normalize(
experiments[i].dg_w_pca)
# tb_w_tangent_bases = experiment.project(experiment.dw_pca, np.swapaxes(subM.tb.tangent_bases,1,2))
# experiment.dgw_norm = experiment.normalize(experiment.dg_w)
#experiment[i].dg_M = experiment.project(subM.tb.tangent_bases, experiment.dgw_norm)
experiments[i].dg_M = experiments[i].project(subM.tb.tangent_bases,
experiments[i].dgw_norm)
#experiments[i].dw_pca = np.asarray([np.matmul(projector, experiments[i].dw[j]) for j in range(nsel)])
#experiments[i].dg_w = experiments[i].project(experiments[i].dw_pca, np.swapaxes(experiments[i].dg_x_pca,1,2))
#tb_w_tangent_bases = experiments[i].project(experiments[i].dw_pca, np.swapaxes(subM.tb.tangent_bases, 1, 2))
#experiments[i].dw_norm = experiments[i].normalize(experiments[i].dg_w)
#experiments[i].dg_M = experiments[i].project(np.swapaxes(tb_w_tangent_bases, 1, 2), experiments[i].dw_norm)
#experiments[i].coeffs = experiments[i].get_betas_spam2(experiments[i].xtrain, experiments[i].ytrain, experiments[i].groups, lambdas,
# nsel, experiments[i].q, itermax, tol)
return (experiments)
def get_grads_reps(experiment, nreps, nsel, cores):
experiments = {}
dim = experiment.dim
for i in range(nreps):
experiments[i] = copy.copy(experiment)
experiments[i].M.selected_points = np.random.choice(list(
range(experiment.n)),
nsel,
replace=False)
tangent_bases = experiments[i].M.get_wlpca_tangent_sel(
experiments[i].M, experiments[i].M.selected_points)
subM = RiemannianManifold(
experiments[i].M.data[experiments[i].M.selected_points], dim)
subM.tb = TangentBundle(subM, tangent_bases)
experiments[i].N.tangent_bundle = TangentBundle(
experiments[i].N,
experiments[i].N.geom.rmetric.embedding_eigenvectors)
experiments[i].df_M = experiments[i].get_dF_js_idM(
experiments[i].M, experiments[i].N, subM.tb,
experiments[i].N.tangent_bundle, experiments[i].M.selected_points)
#experiments[i].df_M = experiments[i].df_M / np.linalg.norm(experiments[i].df_M, axis=1).sum(axis=0)
experiments[i].df_M2 = experiments[i].df_M / np.linalg.norm(
experiments[i].df_M)**2
experiments[i].dg_x = experiments[i].get_dx_g_full(
experiments[i].M.data[experiments[i].M.selected_points])
experiments[i].W = ShapeSpace(experiments[i].positions,
experiments[i].M.data)
experiments[i].dw = experiments[i].W.get_dw(
cores, experiments[i].atoms3, experiments[i].natoms,
experiments[i].M.selected_points)
experiments[i].dg_w = experiments[i].project(experiments[i].dw,
experiments[i].dg_x)
tb_w_tangent_bases = experiments[i].project(
experiments[i].dw, np.swapaxes(subM.tb.tangent_bases, 1, 2))
experiments[i].dw_norm = experiments[i].normalize(experiments[i].dg_w)
experiments[i].dg_M = experiments[i].project(
np.swapaxes(tb_w_tangent_bases, 1, 2), experiments[i].dw_norm)
#experiments[i].coeffs = experiments[i].get_betas_spam2(experiments[i].xtrain, experiments[i].ytrain, experiments[i].groups, lambdas,
# nsel, experiments[i].q, itermax, tol)
return (experiments)
