def train(self, images, annotations, metadata=None, **kwargs):
## load images
ntrain = len(images)
xs = images
zs = annotations
if metadata is None:
metas = [None for i in range(ntrain)]
else:
metas = metadata
## initial w
w0 = kwargs.pop('w0', 1)
w = w0
## initial latent variable
y0s = kwargs.pop('init_latents', None)
## initial objective
svm_objective0 = 0
for niter in range(1, self.niter_max + 1):
logger.debug('iteration (latent) #{}'.format(niter))
## annotation consistent inference
logger.debug('annotation consistent inference (with w={})'\
.format(self.print_w(w)))
ys = self.all_aci(w, xs, zs, metas, y0s=y0s)
## convec struct svm
logger.debug('start struct svm')
struct_svm = StructSVM(loss_function=self.user_loss,
psi=self.user_psi,
most_violated_constraint=self.user_mvc,
**self.struct_params)
w, xi, struct_info = struct_svm.train(xs,
ys,
metadata=metas,
w0=w,
**kwargs)
## Stop condition
svm_objective = struct_svm.objective(w, xi)
strw = self.print_w(w)
if (svm_objective - svm_objective0) < self.epsilon:
logger.debug('latent svm stopping with: w={}, xi={:.04}'\
.format(strw, float(xi)))
info = {'converged': True}
return w, xi, ys, info
## end current iteration
logger.debug('done iteration #{}, with: w={}, xi={:.04}'.format(
niter, strw, float(xi)))
logger.debug('objective={:.04}'.format(svm_objective))
svm_objective0 = svm_objective
else:
info = {'converged': False}
logger.info("max number of iterations reached ({})".format(
self.niter_max))
return w, xi, ys, info
def train(self, x, z, balanced=True, **kwargs):
import struct_svm
struct_svm.logger = struct_svm.utils_logging.get_logger(
'svm', struct_svm.utils_logging.INFO)
from struct_svm import StructSVM
C = kwargs.pop('C', 1.)
if balanced:
n0 = np.where(np.asarray(z) == 0)[0]
n1 = np.where(np.asarray(z) == 1)[0]
n = np.minimum(len(n0), len(n1))
in0 = np.random.permutation(n0)[:n]
in1 = np.random.permutation(n1)[:n]
S = [(x[i], z[i]) for i in in0]
S += [(x[i], z[i]) for i in in1]
else:
S = [(x[i], z[i]) for i in range(len(z))]
svm = StructSVM(
S,
self.loss,
self.psi,
self.mvc,
C=C,
)
## train svm
w, xi, info = svm.train()
self.w = w
self.xi = xi
def train(self, images, annotations, metadata=None, **kwargs):
## load images
ntrain = len(images)
xs = images
zs = annotations
if metadata is None:
metas = [None for i in range(ntrain)]
else: metas = metadata
## initial w
w0 = kwargs.pop('w0',1)
w = w0
## initial latent variable
y0s = kwargs.pop('init_latents', None)
## initial objective
svm_objective0 = 0
for niter in range(1,self.niter_max+1):
logger.debug('iteration (latent) #{}'.format(niter))
## annotation consistent inference
logger.debug('annotation consistent inference (with w={})'\
.format(self.print_w(w)))
ys = self.all_aci(w, xs, zs, metas, y0s=y0s)
## convec struct svm
logger.debug('start struct svm')
struct_svm = StructSVM(
loss_function=self.user_loss,
psi=self.user_psi,
most_violated_constraint=self.user_mvc,
**self.struct_params
)
w,xi,struct_info = struct_svm.train(xs, ys, metadata=metas, w0=w, **kwargs)
## Stop condition
svm_objective = struct_svm.objective(w,xi)
strw = self.print_w(w)
if (svm_objective - svm_objective0) < self.epsilon:
logger.debug('latent svm stopping with: w={}, xi={:.04}'\
.format(strw, float(xi)))
info = {'converged': True}
return w, xi, ys, info
## end current iteration
logger.debug('done iteration #{}, with: w={}, xi={:.04}'.format(niter, strw, float(xi)))
logger.debug('objective={:.04}'.format(svm_objective))
svm_objective0 = svm_objective
else:
info = {'converged': False}
logger.info("max number of iterations reached ({})".format(self.niter_max))
return w, xi, ys, info
def train(self, x, z, balanced=True, **kwargs):
import struct_svm
struct_svm.logger = struct_svm.utils_logging.get_logger('svm',struct_svm.utils_logging.INFO)
from struct_svm import StructSVM
C = kwargs.pop('C', 1.)
if balanced:
n0 = np.where(np.asarray(z)==0)[0]
n1 = np.where(np.asarray(z)==1)[0]
n = np.minimum(len(n0), len(n1))
in0 = np.random.permutation(n0)[:n]
in1 = np.random.permutation(n1)[:n]
S = [(x[i],z[i]) for i in in0]
S += [(x[i],z[i]) for i in in1]
else:
S = [(x[i],z[i]) for i in range(len(z))]
svm = StructSVM(S, self.loss, self.psi, self.mvc,C=C,)
## train svm
w,xi,info = svm.train()
self.w = w
self.xi = xi
