def __init__(self, deployFile=None, session=None):
self.isDeployed = False
if session is None:
if deployFile is not None:
self.matlab = transplant.Matlab(executable=deployFile,
arguments=['-nodisplay'])
self.isDeployed = True
else:
self.matlab = transplant.Matlab()
else:
self.matlab = session
def __init__(self,
pde,
theta=0.2,
maxit=50,
step=0,
maxdof=1e5,
n=3,
p=1,
q=3,
sigma=None,
multieigs=False,
matlab=False,
resultdir='~/'):
self.multieigs = multieigs
self.sigma = sigma
self.pde = pde
self.step = step
self.theta = theta
self.maxit = maxit
self.maxdof = maxdof
self.p = p
self.q = q
self.numrefine = n
self.resultdir = resultdir
self.picard = False
if matlab is True:
self.matlab = transplant.Matlab()
else:
self.matlab = False
def __init__(self, root):
super(ECO_H, self).__init__(name="ECO-H")
self.root_path = root + "/external/ECO"
self.eng = transplant.Matlab()
genpath = self.eng.genpath(self.root_path)
self.eng.addpath(genpath)
self.eng.vl_setupnn()
def __init__(self, matlab=None):
if matlab is None:
self.matlab = transplant.Matlab()
else:
self.matlab = matlab
def restart_matlab(self, line):
import transplant
self.m = transplant.Matlab()
def __init__(self, shell):
super(MatlabMagic, self).__init__(shell)
import transplant
self.m = transplant.Matlab()
def __init__(self, root):
super(MCCT_H, self).__init__(name="MCCT-H")
self.root_path = root + "/external/MCCT/TrackerMCCT-H"
self.eng = transplant.Matlab()
genpath = self.eng.genpath(self.root_path)
self.eng.addpath(genpath)
import transplant
import numpy as np
matlab = transplant.Matlab()
matlab.addpath("/home/george/matlab_projects/code_nr_alg3_book/TabGenGam/")
g_dft, g_mag, mag2 = matlab.Tabulate_gain_functions(1, 0.6)
np.save("gain.npy", g_mag)
def _collect_stats(self,
stats_loader,
cuda,
criterion=None,
stats_type='marginal',
run_logger=None):
self.eval()
if stats_type == 'jacobian':
data, target = next(iter(stats_loader))
if cuda:
data, target = data.cuda(), target.cuda()
data.requires_grad_()
stats, _ = self(data, stats_type=stats_type)
stats = tuple(map(lambda el: el.cpu().detach().numpy(), stats))
stats = np.dstack(stats)
stats = np.swapaxes(stats, 0, -1)
stats = np.linalg.svd(stats, full_matrices=False, compute_uv=False)
elif stats_type == 'smoothness':
data, target = next(iter(stats_loader))
if cuda:
data, target = data.cuda(), target.cuda()
x, output = self(data, stats_type=stats_type)
loss = criterion(output, target)
# Collect statistics
lambda_max_Fisher = SmoothnessFisher(loss, output, self)
lambda_max_AbsHessian = SmoothnessAbsHessian(loss, output, self)
if run_logger is not None:
run_logger.log_scalar("curvature.SmoothnessFisher",
float(lambda_max_Fisher))
run_logger.log_scalar("curvature.SmoothnessAbsHessian",
float(lambda_max_AbsHessian))
stats = [lambda_max_Fisher, lambda_max_AbsHessian]
elif stats_type == 'condition_number':
data, target = next(iter(stats_loader))
if cuda:
data, target = data.cuda(), target.cuda()
x, output = self(data, stats_type=stats_type)
loss = criterion(output, target)
# Collect statistics
lambda_max_AbsHessian = SmoothnessAbsHessian(loss, output, self)
lambda_min_AbsHessian = StrongConvexity(
loss, output, self, lambda_max=lambda_max_AbsHessian)
if run_logger is not None:
run_logger.log_scalar("curvature.SmoothnessAbsHessian",
float(lambda_max_AbsHessian))
run_logger.log_scalar("curvature.SCAbsHessian",
float(lambda_min_AbsHessian))
run_logger.log_scalar(
"curvature.ConditionNumber",
float(lambda_max_AbsHessian / lambda_min_AbsHessian))
stats = [lambda_max_AbsHessian, lambda_min_AbsHessian]
else:
stats_temp = []
stats = []
for batch_idx, (data, target) in enumerate(
tqdm(stats_loader, desc='Stat batches', ascii=True)):
data, target = data.cuda(), target.cuda()
stat, _ = self(data, stats_type=stats_type)
stats_temp += stat
stats_temp = np.hstack(stats_temp)
stats_temp = np.swapaxes(stats_temp, 1, 2)
import transplant
matlab = transplant.Matlab()
matlab.addpath("/home/***/projects/isonetry/projectionpursuit")
for i in range(stats_temp.shape[0]):
_, projdata, _ = matlab.NGCA(
stats_temp[0].astype(np.double),
transplant.MatlabStruct({
"nbng": 1,
"sphered_proj": 0
}))
stats += [projdata]
matlab.exit()
stats = np.vstack(stats)
return stats
