def main(): sarcos_dir = os.path.join(os.getenv("SIGVISA_HOME"), "papers", "product_tree", "sarcos") sarcos_train_X, sarcos_train_y, sarcos_test_X, sarcos_test_y, hyperparams = load_sarcos(sdir=sarcos_dir) np.save(os.path.join(sarcos_dir, "testX.npy"), sarcos_test_X) np.save(os.path.join(sarcos_dir, "testy.npy"), sarcos_test_y) np.save(os.path.join(sarcos_dir, "hyperparams.npy"), hyperparams) print "loaded sarcos data and converted to numpy format" train_realdata_gp(sarcos_dir, sarcos_train_X, sarcos_train_y, hyperparams) print "trained model" test_predict(sarcos_dir) print "evaluated predictions" eval_gp(bdir=sarcos_dir, test_n=100) print "timings finished"
def cov_timing(cv_dir, lscale, order, n_azi_buckets): sgp = SparseGP(fname=os.path.join(cv_dir, "fold_00.gp%d_%d_%d" % (lscale, order, n_azi_buckets))) X = np.loadtxt(os.path.join(cv_dir, "X.txt")) test_idx = np.array([int(z) for z in np.loadtxt(os.path.join(cv_dir, "fold_00_test.txt"))]) testX = X[test_idx] resultfile = os.path.join(cv_dir, "results_gp%d_%d_%d.txt" % (lscale, order, n_azi_buckets)) errorfile = os.path.join(cv_dir, "error_gp%d_%d_%d.npz" % (lscale, order, n_azi_buckets)) eval_gp(gp=sgp, testX=testX, resultfile=resultfile, errorfile=errorfile) poly = LinearBasisModel(fname=os.path.join(cv_dir, "fold_00.poly_%d_%d" % (order, n_azi_buckets))) resultfile_poly = os.path.join(cv_dir, "results_poly_%d_%d.txt" % (order, n_azi_buckets)) f = open(resultfile_poly, 'w') test_n = len(test_idx) poly_covars = np.zeros((test_n,)) t0 = time.time() for i in range(test_n): poly_covars[i] = poly.covariance(testX[i:i+1,:]) t1 = time.time() f.write("cov time %f\n" % ((t1-t0)/test_n)) f.close()
def learn_gp(X, y): p = np.random.permutation(len(y)) train_n = int(len(y) * 0.2) trainX = X[p[:train_n], :] trainy = y[p[:train_n]] testX = X[p[train_n:], :] testy = y[p[train_n:]] fitz_dir = os.path.join(os.getenv("SIGVISA_HOME"), "papers", "product_tree", "fitz_learned") # hyperparams = np.array([0.5, 3.0, 50.0, 50.0], dtype=float) # hyperparams = learn_hyperparams(fitz_dir, trainX, trainy, dfn_str='lld', hyperparams=hyperparams, sparse_invert=False, basisfns = [lambda x : 1,], param_cov=np.array(((10000,),)), param_mean = np.array((0,)), k=1000) # print "got hyperparams", hyperparams # hyperparams = np.array([1.16700753, 2.53145332, 212.46536884,157.68719303], dtype=float) np.save(os.path.join(fitz_dir, "testX.npy"), testX) np.save(os.path.join(fitz_dir, "testy.npy"), testy) np.save(os.path.join(fitz_dir, "hyperparams.npy"), hyperparams) print "loaded data" train_realdata_gp( fitz_dir, trainX, trainy, hyperparams=hyperparams, sparse_invert=False, basisfns=[lambda x: 1], param_cov=np.array(((10000,),)), param_mean=np.array((0,)), dfn_str="lld", ) test_predict(fitz_dir) eval_gp(bdir=fitz_dir, test_n=100)
def main(n_max=18000): rundir = sys.argv[1] task_name = sys.argv[2] basedir = os.path.join(os.getenv('SIGVISA_HOME'), 'papers', 'product_tree', 'run', rundir) basename = os.path.join(basedir, task_name) X_train = np.loadtxt(basename + '_X_train.txt', delimiter=',') y_train = np.loadtxt(basename + '_y_train.txt', delimiter=',') n_X = X_train.shape[0] if n_X > n_max: X_train = np.array(X_train[:n_max,:], copy=True) y_train = np.array(y_train[:n_max], copy=True) print "using restricted subset of %d training points" % (n_max) actual_n = min(n_X, n_max) X_test = np.loadtxt(basename + '_X_test.txt', delimiter=',') y_test = np.loadtxt(basename + '_y_test.txt', delimiter=',') """ #for nu in (90,): nu = 90 csficbase = basename + '_csfic_%d' % nu if nu is not None else basename + '_csfic' csficmodel_dir = basename + "_py_csfic_%d" % nu if nu is not None else basename + "_py_csfic" print csficmodel_dir #if not os.path.exists(csficbase): # continue csgp = os.path.join(csficmodel_dir, 'trained_%d.gp' % actual_n) mkdir_p(csficmodel_dir) if os.path.exists(csgp): gp_csfic = SparseGP_CSFIC(fname=csgp, build_tree=True, leaf_bin_size=15) else: gp_csfic = load_matlab_csficmodel(csficbase, X_train, y_train) gp_csfic.save_trained_model(csgp) #print "testing predictions" #test_predict(csficmodel_dir, sgp=gp_csfic, testX=X_test, testy=y_test) #print "testing cutoff rule 0" #eval_gp(bdir=csficmodel_dir, testX=X_test, test_n=200, gp=gp_csfic, cutoff_rule=0) #print "testing cutoff rule 1" #eval_gp(bdir=csficmodel_dir, testX=X_test, test_n=200, gp=gp_csfic, cutoff_rule=1) #print "testing cutoff rule 2" #eval_gp(bdir=csficmodel_dir, testX=X_test, test_n=200, gp=gp_csfic, cutoff_rule=2) print "testing leaf bins 15" gp_csfic10 = SparseGP_CSFIC(fname=csgp, build_tree=True, leaf_bin_size=10) eval_gp(bdir=csficmodel_dir, testX=X_test, test_n=200, gp=gp_csfic10, cutoff_rule=2, flag="_bin15") #print "testing leaf bins 100" #gp_csfic50 = SparseGP_CSFIC(fname=csgp, build_tree=True, leaf_bin_size=50) #eval_gp(bdir=csficmodel_dir, testX=X_test, test_n=200, gp=gp_csfic50, cutoff_rule=2, flag="_bin50") """ semodel_dir = basename + "_py_se" segp = os.path.join(semodel_dir, 'trained.gp') mkdir_p(semodel_dir) if os.path.exists(segp): gp_se = SparseGP(fname=segp, build_tree=True, leaf_bin_size=0, sparse_invert=False) else: gp_se = load_matlab_semodel(basename, X_train, y_train) gp_se.save_trained_model(segp) test_predict(semodel_dir, sgp=gp_se, testX=X_test, testy=y_test) eval_gp(bdir=semodel_dir, testX=X_test, test_n=200, gp=gp_se)