def learn(self,layer_start=0):
# learn additional layers:
for mind, m in enumerate(self.model_sequence[layer_start:]):
if not mind:
l = learners.SGD(model=m,datasource=self.datasource,display_every=20000,batchsize=self.batchsize)
l_firstlayer = l
else:
l = learners.SGD_layer(first_layer_learner=l_firstlayer,model=m,datasource=self.datasource,display_every=20000,batchsize=self.batchsize,model_sequence=self.model_sequence[:mind],layer_params=self.layer_params)
whitenpatches = self.layer_params[mind]['whitenpatches']
databatch = l.get_databatch(whitenpatches)
l.model.learn_whitening(databatch)
l.model.setup()
sched_list = self.schedules[mind]
for sdict in sched_list:
if sdict.has_key('change_target'):
l.change_target(sdict['change_target'])
if sdict.has_key('batchsize'):
l.batchsize *= sdict['batchsize']
if sdict.has_key('iterations'):
l.learn(iterations=sdict['iterations'])
else:
l.learn()
from display import display_final
display_final(self.model_sequence[mind])
def fit(self,X,y=None):
if not self._setup_complete: self.setup(X)
databatch = self._learner.get_databatch(self.whitenpatches)
self._learner.model.learn_whitening(databatch)
self._learner.model.setup()
sched_list = self.get_learning_schedule()
for sdict in sched_list:
if sdict.has_key('change_target'):
self._learner.change_target(sdict['change_target'])
if sdict.has_key('batchsize'):
self._learner.batchsize = sdict['batchsize']
if sdict.has_key('iterations'):
self._learner.learn(iterations=sdict['iterations'])
else:
self._learner.learn()
display_final(self._model)
self._learning_complete = True
return self
def fit(self, X, y=None):
if not self._setup_complete: self.setup(X)
databatch = self._learner.get_databatch(self.whitenpatches)
self._learner.model.learn_whitening(databatch)
display_final(self._model)
self._learning_complete = True
return self
def learn(self,layer_start=0):
l_firstlayer = None
# learn additional layers:
for mind, m in enumerate(self.model_sequence):
if not mind:
if self.ipython_profile is None:
l = learners.SGD(model=m,datasource=self.datasource,display_every=self.display_every,save_every=self.save_every,batchsize=self.batchsize)
else:
l = parallel_learners.SGD(model=m,datasource=self.datasource,display_every=self.display_every,save_every=self.save_every,batchsize=self.batchsize,ipython_profile=self.ipython_profile)
l_firstlayer = l
else:
if self.ipython_profile is None:
l = learners.SGD_layer(first_layer_learner=l_firstlayer,model=m,datasource=self.datasource,display_every=self.display_every,save_every=self.save_every,batchsize=self.batchsize,model_sequence=self.model_sequence[:mind],layer_params=self.layer_params)
else:
l = parallel_learners.SGD_layer(first_layer_learner=l_firstlayer,model=m,datasource=self.datasource,display_every=self.display_every,save_every=self.save_every,batchsize=self.batchsize,model_sequence=self.model_sequence[:mind],layer_params=self.layer_params,ipython_profile=self.ipython_profile)
if layer_start > mind:
self.model = models.HierarchicalModel(model_sequence=self.model_sequence[:mind+1],layer_params=self.layer_params[:mind+1])
self.evaluate()
continue
whitenpatches = self.layer_params[mind]['whitenpatches']
databatch = l.get_databatch(whitenpatches)
l.model.learn_whitening(databatch)
l.model.setup()
# update self.model to use the sequence up to this layer
self.model = models.HierarchicalModel(model_sequence=self.model_sequence[:mind+1],layer_params=self.layer_params[:mind+1])
sched_list = self.schedules[mind]
iter0 = self.iter
self.evaluate()
for sdict in sched_list:
if sdict.has_key('change_target'):
l.change_target(sdict['change_target'])
if sdict.has_key('batchsize'):
l.batchsize = sdict['batchsize']
if sdict.has_key('iterations'):
l.learn(iterations=sdict['iterations'])
else:
l.learn()
self.iter = iter0 + l.iter
self.evaluate()
from display import display_final
display_final(self.model_sequence[mind])
def learn(self,layer_start=0):
l_firstlayer = None
# learn additional layers:
for mind, m in enumerate(self.model_sequence):
if not mind:
if self.ipython_profile is None:
l = learners.SGD(model=m,
datasource=self.datasource,
display_every=self.display_every,
save_every=self.save_every,
batchsize=self.batchsize,
layer_params=self.layer_params,
eta_target_maxupdate=self.start_eta_target_maxupdates[mind],
eta=self.start_etas[mind],
**self.extra_learner_kargs)
else:
l = parallel_learners.SGD(model=m,
datasource=self.datasource,
display_every=self.display_every,
save_every=self.save_every,
batchsize=self.batchsize,
layer_params=self.layer_params,
ipython_profile=self.ipython_profile,
eta_target_maxupdate=self.start_eta_target_maxupdates[mind],
eta=self.start_etas[mind],
**self.extra_learner_kargs)
l_firstlayer = l
else:
if self.ipython_profile is None:
l = learners.SGD_layer(first_layer_learner=l_firstlayer,
model=m,
datasource=self.datasource,
display_every=self.display_every,
save_every=self.save_every,
batchsize=self.batchsize,
model_sequence=self.model_sequence[:mind],
layer_params=self.layer_params,
eta_target_maxupdate=self.start_eta_target_maxupdates[mind],
eta=self.start_etas[mind],
**self.extra_learner_kargs)
else:
l = parallel_learners.SGD_layer(first_layer_learner=l_firstlayer,
model=m,
datasource=self.datasource,
display_every=self.display_every,
save_every=self.save_every,
batchsize=self.batchsize,
model_sequence=self.model_sequence[:mind],
layer_params=self.layer_params,
ipython_profile=self.ipython_profile,
eta_target_maxupdate=self.start_eta_target_maxupdates[mind],
eta=self.start_etas[mind],
**self.extra_learner_kargs)
if layer_start > mind:
self.model = models.HierarchicalModel(model_sequence=self.model_sequence[:mind+1],layer_params=self.layer_params[:mind+1])
if self.evaluate_unlearned_layers:
self.evaluate(l)
continue
print 'Begin learning layer', mind
print 'Learn whitening...'
whitenpatches = self.layer_params[mind]['whitenpatches']
databatch = l.get_databatch(whitenpatches)
l.model.learn_whitening(databatch)
print 'Done.'
print 'Setup model...'
l.model.setup()
print 'Done.'
# update self.model to use the sequence up to this layer
self.model = models.HierarchicalModel(model_sequence=self.model_sequence[:mind+1],layer_params=self.layer_params[:mind+1])
sched_list = self.schedules[mind]
iter0 = self.iter
self.evaluate(l)
for sdict in sched_list:
if sdict.has_key('change_target'):
l.change_target(sdict['change_target'])
if sdict.has_key('batchsize'):
l.batchsize = sdict['batchsize']
if sdict.has_key('center_basis_functions') and hasattr(l.model,'center_basis_functions'):
l.model.center_basis_functions = sdict['center_basis_functions']
if sdict.has_key('iterations'):
l.learn(iterations=sdict['iterations'])
else:
l.learn()
self.iter = iter0 + l.iter
self.evaluate(l)
from display import display_final
display_final(self.model_sequence[mind])
l.model.save(save_txt=True)
l.save()