def check_the_grad(nos=100, idim=30, hdim=10, eps=1e-8, verbose=False):
"""
Check gradient computation for Neural Networks.
"""
#
from opt import check_grad
from misc import sigmoid
from losses import ssd, mia
# number of input samples (nos)
# with dimension ind each
ins = np.random.randn(nos, idim)
structure = dict()
structure["hdim"] = hdim
structure["af"] = sigmoid
structure["score"] = ssd
structure["beta"] = 0.7
structure["rho"] = 0.01
weights = np.zeros(idim*hdim + hdim + idim)
weights[:idim*hdim] = 0.001 * np.random.randn(idim*hdim)
args = dict()
args["inputs"] = ins
args["structure"] = structure
#
delta = check_grad(true_score, grad, weights, args, eps=eps, verbose=verbose)
assert delta < 1e-4, "[sae.py] check_the_grad FAILED. Delta is %f" % delta
return True
def check_the_grad(nos=1, ind=30, outd=5, bxe=False,
eps=1e-6, verbose=False):
"""
"""
from opt import check_grad
#
weights = 0.1 * np.random.randn(ind*outd + outd)
ins = np.random.randn(nos, ind)
if bxe:
score = score_mia
grad = grad_mia
outs = 1.*(np.random.rand(nos, outd) > 0.5)
else:
outs = np.random.random_integers(outd, size=(nos)) - 1
score = score_xe
grad = grad_xe
cg = dict()
cg["inputs"] = ins
cg["targets"] = outs
#
delta = check_grad(f=score, fprime=grad, x0=weights,
args=cg, eps=eps, verbose=verbose)
assert delta < 1e-4, "[logreg.py] check_the_gradient FAILED. Delta is %f" % delta
return True
def check_the_grad(nos=1, ind=30, outd=5, bxe=False, eps=1e-6, verbose=False):
"""
"""
from opt import check_grad
#
weights = 0.1 * np.random.randn(ind * outd + outd)
ins = np.random.randn(nos, ind)
if bxe:
score = score_mia
grad = grad_mia
outs = 1. * (np.random.rand(nos, outd) > 0.5)
else:
outs = np.random.random_integers(outd, size=(nos)) - 1
score = score_xe
grad = grad_xe
cg = dict()
cg["inputs"] = ins
cg["targets"] = outs
#
delta = check_grad(f=score,
fprime=grad,
x0=weights,
args=cg,
eps=eps,
verbose=verbose)
assert delta < 1e-4, "[logreg.py] check_the_gradient FAILED. Delta is %f" % delta
return True
def check_the_grad(nos=100, ind=30, outd=10, eps=1e-8, verbose=False):
"""
Check gradient computation.
_nos_: number of samples
_ind_: dimension of one sample
_outd_: number of filters
"""
from opt import check_grad
from misc import sqrtsqr
ins = np.random.randn(nos, ind)
weights = 0.001 * np.random.randn(ind, outd).ravel()
structure = dict()
structure["af"] = sqrtsqr
structure["eps"] = 1e-8
args = dict()
args["inputs"] = ins
args["structure"] = structure
delta = check_grad(score, grad, weights, args, eps=eps, verbose=verbose)
assert delta < 1e-4, "[sf.py] check_the_grad FAILED. Delta is %f" % delta
return True
def check_the_grad(regression=True, nos=1, ind=30,
outd=3, bxe=False, eps=1e-8,verbose=False):
"""
Check gradient computation for Neural Networks.
"""
#
from opt import check_grad
from misc import sigmoid
from losses import xe, ssd, mia
# number of input samples (nos)
# with dimension ind each
ins = np.random.randn(nos, ind)
#
structure = dict()
if regression:
# Regression
# Network with one hidden layer
structure["layers"] = [(ind, 15), (15, outd)]
structure["activs"] = [np.tanh]
structure["score"] = ssd
outs = np.random.randn(nos, outd)
else:
# Classification
# _outd_ is interpreted as number of classes
structure["layers"] = [(ind, 15), (15, outd)]
structure["activs"] = [sigmoid]
if bxe:
structure["score"] = mia
outs = 1.*(np.random.rand(nos, outd) > 0.5)
else:
structure["score"] = xe
outs = np.random.random_integers(outd, size=(nos)) - 1
# weight decay
structure["l2"] = 0.1
weights = init_weights(structure)
cg = dict()
cg["inputs"] = ins
cg["targets"] = outs
cg["structure"] = structure
#
delta = check_grad(score, grad, weights, cg, eps=eps, verbose=verbose)
assert delta < 1e-4, "[nn.py] check_the_grad FAILED. Delta is %f" % delta
return True
def check_the_grad(loss,
nos=1,
idim=30,
hdim=10,
odim=5,
eps=1e-8,
verbose=False):
"""
Check gradient computation for Neural Networks.
"""
#
from opt import check_grad
from misc import sigmoid
from losses import ssd, mia, xe
# number of input samples (nos)
# with dimension ind each
ins = np.random.rand(nos, idim)
if loss is ssd:
outs = np.random.randn(nos, odim)
if loss is mia:
outs = 1. * (np.random.rand(nos, odim) > 0.5)
if loss is xe:
outs = np.random.random_integers(odim, size=(nos)) - 1
structure = dict()
structure["hdim"] = hdim
structure["odim"] = odim
structure["af"] = sigmoid
structure["score"] = loss
structure["l2"] = 1e-2
weights = np.zeros(idim * hdim + hdim + hdim * odim + odim)
weights[:idim * hdim] = 0.001 * np.random.randn(idim * hdim)
weights[idim * hdim + hdim:-odim] = 0.001 * np.random.randn(hdim * odim)
args = dict()
args["inputs"] = ins
args["targets"] = outs
args["structure"] = structure
#
delta = check_grad(score, grad, weights, args, eps=eps, verbose=verbose)
assert delta < 1e-4, "[flann.py] check_the_grad FAILED. Delta is %f" % delta
return True
def check_the_grad(loss, nos=1, idim=30, hdim=10, odim=5, eps=1e-8, verbose=False):
"""
Check gradient computation for Neural Networks.
"""
#
from opt import check_grad
from misc import sigmoid
from losses import ssd, mia, xe
# number of input samples (nos)
# with dimension ind each
ins = np.random.rand(nos, idim)
if loss is ssd:
outs = np.random.randn(nos, odim)
if loss is mia:
outs = 1.*(np.random.rand(nos, odim) > 0.5)
if loss is xe:
outs = np.random.random_integers(odim, size=(nos))-1
structure = dict()
structure["hdim"] = hdim
structure["odim"] = odim
structure["af"] = sigmoid
structure["score"] = loss
structure["l2"] = 1e-2
weights = np.zeros(idim*hdim + hdim + hdim*odim + odim)
weights[:idim*hdim] = 0.001 * np.random.randn(idim*hdim)
weights[idim*hdim+hdim:-odim] = 0.001 * np.random.randn(hdim*odim)
args = dict()
args["inputs"] = ins
args["targets"] = outs
args["structure"] = structure
#
delta = check_grad(score, grad, weights, args, eps=eps, verbose=verbose)
assert delta < 1e-4, "[flann.py] check_the_grad FAILED. Delta is %f" % delta
return True
def check_the_grad(nos=1, ind=30, outd=10, eps=1e-8, verbose=False):
"""
Check gradient computation.
"""
from opt import check_grad
from misc import logcosh, sqrtsqr
# number of input samples (nos)
# with dimension ind each
ins = np.random.randn(nos, ind)
weights = 0.001 * np.random.randn(ind*outd)
structure = dict()
structure["l1"] = sqrtsqr
structure["lmbd"] = 1
args = dict()
args["inputs"] = ins
args["structure"] = structure
delta = check_grad(score, grad, weights, args, eps=eps, verbose=verbose)
assert delta < 1e-4, "[ica.py] check_the_grad FAILED. Delta is %f" % delta
return True
def check_the_grad(nos=1, ind=30, outd=10, eps=1e-8, verbose=False):
"""
Check gradient computation.
"""
from opt import check_grad
from misc import logcosh, sqrtsqr
# number of input samples (nos)
# with dimension ind each
ins = np.random.randn(nos, ind)
weights = 0.001 * np.random.randn(ind * outd)
structure = dict()
structure["l1"] = sqrtsqr
structure["lmbd"] = 1
args = dict()
args["inputs"] = ins
args["structure"] = structure
delta = check_grad(score, grad, weights, args, eps=eps, verbose=verbose)
assert delta < 1e-4, "[ica.py] check_the_grad FAILED. Delta is %f" % delta
return True