def test_fg11_top_bandit():
L = lfw.FG11Bandit()
config = stochastic.sample(L.template, np.random.RandomState(0))
config['decisions'] = None
config['slm'] = stochastic.sample(pyll.as_apply(params.fg11_top),
np.random.RandomState(0))
config['comparison'] = 'sqrtabsdiff'
rec = L.evaluate(config, hyperopt.base.Ctrl(None))
assert np.abs(rec['loss'] - .194) < 1e-2
return rec
def get_oneobj_latents(tdict, models, categories):
rng = np.random.RandomState(seed=int(clock()))
latents = []
tname = tdict['name']
template = tdict['template']
n_ex = tdict['n_ex']
for model in models:
for _ind in range(n_ex):
l = stochastic.sample(template, rng)
l['obj'] = [model]
l['category'] = [categories[model][0]]
l['id'] = get_image_id(l)
rec = (l['bgname'],
float(l['bgphi']),
float(l['bgpsi']),
float(l['bgscale']),
l['category'],
l['obj'],
[float(l['ryz'])],
[float(l['rxz'])],
[float(l['rxy'])],
[float(l['ty'])],
[float(l['tz'])],
[float(l['tx'])],
[float(l['s'])],
[None],
[None],
tname,
l['id'])
latents.append(rec)
return latents
def get_oneobj_latents(tdict, models, categories):
rng = np.random.RandomState(seed=int(clock()))
latents = []
tname = tdict["name"]
template = tdict["template"]
n_ex = tdict["n_ex"]
for model in models:
for _ind in range(n_ex):
l = stochastic.sample(template, rng)
l["obj"] = [model]
l["category"] = [categories[model][0]]
l["id"] = get_image_id(l)
rec = (
l["bgname"],
float(l["bgphi"]),
float(l["bgpsi"]),
float(l["bgscale"]),
l["category"],
l["obj"],
[float(l["ryz"])],
[float(l["rxz"])],
[float(l["rxy"])],
[float(l["ty"])],
[float(l["tz"])],
[float(l["tx"])],
[float(l["s"])],
[None],
[None],
tname,
l["id"],
)
latents.append(rec)
return latents
def get_textured_oneobj_latents(tdict, models, categories):
rng = np.random.RandomState(seed=0)
latents = []
tname = tdict['name']
template = tdict['template']
n_ex = tdict['n_ex']
for model in models:
# print('Generating meta for %s' % model)
for _ind in range(n_ex):
l = stochastic.sample(template, rng)
l['obj'] = [model]
l['category'] = [categories[model][0]]
l['id'] = model + str(_ind)
rec = (l['bgname'],
float(l['bgphi']),
float(l['bgpsi']),
float(l['bgscale']),
l['category'],
l['obj'],
[float(l['ryz'])],
[float(l['rxz'])],
[float(l['rxy'])],
[float(l['ty'])],
[float(l['tz'])],
[float(l['tx'])],
[float(l['s'])],
l['texture'],
[float(l['texture_mode'])],
tname,
l['id'])
latents.append(rec)
return latents
def get_canon_latents(tdict, models, categories):
rng = np.random.RandomState(seed=int(clock()))
latents = []
tname = tdict['name']
template = tdict['template']
n_ex = tdict['n_ex']
for model in models:
objmodel = obj.REVERSED_OBJECT_DICT[model]
for _ind in range(n_ex):
lr = obj.get_canonical_angle(model)
lr['s'] = obj.CANONICAL_SIZES[objmodel] / 100.0
l = stochastic.sample(template, rng)
l['obj'] = [model]
l['category'] = [categories[model][0]]
l['id'] = objmodel + '_nolabel'
rec = (l['bgname'],
float(l['bgphi']),
float(l['bgpsi']),
float(l['bgscale']),
l['category'],
l['obj'],
[float(lr['ryz'])],
[float(lr['rxz'])],
[float(lr['rxy'])],
[float(l['ty'])],
[float(l['tz'])],
[float(l['tx'])],
[float(lr['s'])],
l.get('texture'),
l.get('texture_mode'),
tname,
l['id'])
latents.append(rec)
return latents
def forint(self, seed):
config = stochastic.sample(
dict(
slm=[[params.pf_lnorm()]],
preproc=params.pf_preproc()[1]),
np.random.RandomState(seed))
fuzz_config(config)
def test_main_bandit():
L = lfw.MainBandit()
config = stochastic.sample(L.template, np.random.RandomState(999))
print config['model']['slm'][0]
assert np.allclose(
config['model']['slm'][0][0][1]['kwargs']['threshold'],
16.894686245715995)
rec = L.evaluate(config, hyperopt.base.Ctrl(None))
print rec['test_accuracy']
print rec['train_accuracy']
print rec['loss']
assert np.allclose(rec['test_accuracy'], 58.3)
assert np.allclose(rec['train_accuracy'], 84.9090909091)
assert np.allclose(rec['loss'], 0.417)
return rec
def get_multiobj_latents(tdict, models, categories):
rng = np.random.RandomState(seed=int(clock()))
latents = []
tname = tdict['name']
template = tdict['template']
n_ex = tdict['n_ex']
n_obj = 2
model_tuples = list(itertools.combinations(models, n_obj))
for models in model_tuples:
model1 = models[0]
model2 = models[1]
for _ind in range(n_ex):
l = stochastic.sample(template, rng)
OBJS = [model1, model2]
rng.shuffle(OBJS)
CATS = [categories[model1][0], categories[model2][0]]
imgid = get_image_id(l)
rec = (l['bgname'],
float(l['bgphi']),
float(l['bgpsi']),
float(l['bgscale']),
CATS,
OBJS,
[float(i) for i in l['ryz']],
[float(i) for i in l['rxz']],
[float(i) for i in l['rxy']],
[float(i) for i in l['ty']],
[float(i) for i in l['tz']],
[float(i) for i in l['tx']],
[float(i) for i in l['s']],
[None, None],
[None, None],
tname,
imgid )
latents.append(rec)
return latents
def forint(self, seed):
L = lfw.MainBandit()
config = stochastic.sample(L.template,
np.random.RandomState(seed))
fuzz_config(config['model'])
import hyperopt
import hyperopt.base
import eccv12.lfw as lfw
import eccv12.bandits as bandits
import eccv12.experiments as experiments
import os
test_params = {
'slm': [[('lnorm', params.lnorm)]],
'preproc': {
'global_normalize': 0,
'crop': params.crop_choice,
'size': [20, 20]}}
config_tiny_rnd0 = stochastic.sample(
test_params,
np.random.RandomState(0))
config_tiny_rnd2 = stochastic.sample(
test_params,
np.random.RandomState(2))
def test_sampling_lockdown():
"""
Test that pyll samples the same way as when we ran these experiments.
"""
# done with commit
# f11026e235e57f85e9dd4f3a23a67dac2b33e8db
print config_tiny_rnd0
print config_tiny_rnd2
assert config_tiny_rnd0 == {
