def get_toy_data_1v1(num_annots=5, num_names=None, **kwargs):
r"""
CommandLine:
python -m ibeis.algo.hots.demobayes --exec-get_toy_data_1v1 --show
Example:
>>> # ENABLE_DOCTEST
>>> from ibeis.algo.hots.demobayes import * # NOQA
>>> toy_data = get_toy_data_1v1()
>>> ut.quit_if_noshow()
>>> import plottool as pt
>>> show_toy_distributions(toy_data['toy_params'])
>>> ut.show_if_requested()
Example1:
>>> # ENABLE_DOCTEST
>>> from ibeis.algo.hots.demobayes import * # NOQA
>>> toy_data = get_toy_data_1v1()
>>> kwargs = {}
>>> initial_aids = toy_data['aids']
>>> initial_nids = toy_data['nids']
>>> num_annots = 1
>>> num_names = 6
>>> toy_data2 = get_toy_data_1v1(num_annots, num_names, initial_aids=initial_aids, initial_nids=initial_nids)
>>> ut.quit_if_noshow()
>>> import plottool as pt
>>> show_toy_distributions(toy_data['toy_params'])
>>> ut.show_if_requested()
Ignore:
>>> num_annots = 1000
>>> num_names = 400
"""
import vtool as vt
tup_ = get_toy_annots(num_annots, num_names, **kwargs)
aids, nids, aids1, nids1, all_aids, all_nids = tup_
rng = vt.ensure_rng(None)
def pairwise_feature(aidx1,
aidx2,
all_nids=all_nids,
toy_params=toy_params):
if aidx1 == aidx2:
score = -1
else:
#rng = np.random.RandomState(int((aidx1 + 13) * (aidx2 + 13)))
nid1 = all_nids[int(aidx1)]
nid2 = all_nids[int(aidx2)]
params = toy_params[nid1 == nid2]
mu, sigma = ut.dict_take(params, ['mu', 'sigma'])
score_ = rng.normal(mu, sigma)
score = np.clip(score_, 0, np.inf)
return score
pairwise_nids = list([tup[::-1] for tup in ut.iprod(nids, nids1)])
pairwise_matches = np.array([nid1 == nid2 for nid1, nid2 in pairwise_nids])
pairwise_aidxs = list([tup[::-1] for tup in ut.iprod(aids, aids1)])
pairwise_features = np.array(
[pairwise_feature(aidx1, aidx2) for aidx1, aidx2 in pairwise_aidxs])
#pairwise_scores_mat = pairwise_scores.reshape(num_annots, num_annots)
is_diag = [r < c for r, c, in pairwise_aidxs]
diag_scores = pairwise_features.compress(is_diag)
diag_aidxs = ut.compress(pairwise_aidxs, is_diag)
import utool
with utool.embed_on_exception_context:
diag_nids = ut.compress(pairwise_nids, is_diag)
diag_labels = pairwise_matches.compress(is_diag)
#import utool
#utool.embed()
toy_data = {
'aids': aids,
'nids': nids,
'all_nids': all_nids,
'all_aids': all_aids,
#'pairwise_aidxs': pairwise_aidxs,
#'pairwise_scores': pairwise_scores,
#'pairwise_matches': pairwise_matches,
'diag_labels': diag_labels,
'diag_scores': diag_scores,
'diag_nids': diag_nids,
'diag_aidxs': diag_aidxs,
'toy_params': toy_params,
}
return toy_data
def get_toy_annots(num_annots,
num_names=None,
initial_aids=None,
initial_nids=None,
nid_sequence=None,
seed=None):
r"""
Args:
num_annots (int):
num_names (int): (default = None)
initial_aids (None): (default = None)
initial_nids (None): (default = None)
nid_sequence (None): (default = None)
seed (None): (default = None)
Returns:
tuple: (aids, nids, aids1, nids1, all_aids, all_nids)
CommandLine:
python -m ibeis.algo.hots.demobayes --exec-get_toy_annots
Example:
>>> # DISABLE_DOCTEST
>>> from ibeis.algo.hots.demobayes import * # NOQA
>>> num_annots = 1
>>> num_names = 5
>>> initial_aids = np.array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9], dtype=np.int64)
>>> initial_nids = np.array([0, 0, 1, 2, 2, 1, 1, 1, 2, 3], dtype=np.int64)
>>> nid_sequence = np.array([0, 0, 1, 2, 2, 1, 1], dtype=np.int64)
>>> seed = 0
>>> (aids, nids, aids1, nids1, all_aids, all_nids) = get_toy_annots(num_annots, num_names, initial_aids, initial_nids, nid_sequence, seed)
>>> result = ('(aids, nids, aids1, nids1, all_aids, all_nids) = %s' % (ut.repr2((aids, nids, aids1, nids1, all_aids, all_nids), nl=1),))
>>> print(result)
"""
import vtool as vt
if num_names is None:
num_names = num_annots
print('Generating toy data with num_annots=%r' % (num_annots, ))
if initial_aids is None:
assert initial_nids is None
first_step = True
initial_aids = []
initial_nids = []
else:
first_step = False
assert initial_nids is not None
aids = np.arange(len(initial_aids), num_annots + len(initial_aids))
rng = vt.ensure_rng(seed)
if nid_sequence is None:
nids = rng.randint(0, num_names, num_annots)
else:
unused_from_sequence = max(len(nid_sequence) - len(initial_aids), 0)
if unused_from_sequence == 0:
nids = rng.randint(0, num_names, num_annots)
elif unused_from_sequence > 0 and unused_from_sequence < num_annots:
num_remain = num_annots - unused_from_sequence
nids = np.append(nid_sequence[-unused_from_sequence:],
rng.randint(0, num_names, num_remain))
else:
nids = nid_sequence[-unused_from_sequence]
nids = np.array(
ut.take(
nid_sequence,
range(len(initial_aids),
len(initial_aids) + num_annots)))
if first_step:
aids1 = aids
nids1 = nids
else:
aids1 = initial_aids
nids1 = initial_nids
all_nids = np.append(initial_nids, nids)
all_aids = np.append(initial_aids, aids)
import utool
with utool.embed_on_exception_context:
ut.assert_eq(len(aids), len(nids), 'len new')
ut.assert_eq(len(aids1), len(nids1), 'len comp')
ut.assert_eq(len(all_aids), len(all_nids), 'len all')
return aids, nids, aids1, nids1, all_aids, all_nids
def get_toy_data_1vM(num_annots, num_names=None, **kwargs):
r"""
Args:
num_annots (int):
num_names (int): (default = None)
Kwargs:
initial_aids, initial_nids, nid_sequence, seed
Returns:
tuple: (pair_list, feat_list)
CommandLine:
python -m ibeis.algo.hots.demobayes --exec-get_toy_data_1vM --show
Example:
>>> # DISABLE_DOCTEST
>>> from ibeis.algo.hots.demobayes import * # NOQA
>>> num_annots = 1000
>>> num_names = 40
>>> get_toy_data_1vM(num_annots, num_names)
>>> ut.quit_if_noshow()
>>> import plottool as pt
>>> ut.show_if_requested()
"""
import vtool as vt
tup_ = get_toy_annots(num_annots, num_names, **kwargs)
aids, nids, aids1, nids1, all_aids, all_nids = tup_
rng = vt.ensure_rng(None)
# Test a simple SVM classifier
nid2_nexemp = ut.dict_hist(nids1)
aid2_nid = dict(zip(aids, nids))
ut.fix_embed_globals()
#def add_to_globals(globals_, subdict):
# globals_.update(subdict)
unique_nids = list(nid2_nexemp.keys())
def annot_to_class_feats2(aid, aid2_nid, top=None):
pair_list = []
score_list = []
nexemplar_list = []
for nid in unique_nids:
label = (aid2_nid[aid] == nid)
num_exemplars = nid2_nexemp.get(nid, 0)
if num_exemplars == 0:
continue
params = toy_params[label]
mu, sigma = ut.dict_take(params, ['mu', 'sigma'])
score_ = rng.normal(mu, sigma, size=num_exemplars).max()
score = np.clip(score_, 0, np.inf)
pair_list.append((aid, nid))
score_list.append(score)
nexemplar_list.append(num_exemplars)
rank_list = ut.argsort(score_list, reverse=True)
feat_list = np.array([score_list, rank_list, nexemplar_list]).T
sortx = np.argsort(rank_list)
feat_list = feat_list.take(sortx, axis=0)
pair_list = np.array(pair_list).take(sortx, axis=0)
if top is not None:
feat_list = feat_list[:top]
pair_list = pair_list[0:top]
return pair_list, feat_list
toclass_features = [
annot_to_class_feats2(aid, aid2_nid, top=5) for aid in aids
]
aidnid_pairs = np.vstack(ut.get_list_column(toclass_features, 0))
feat_list = np.vstack(ut.get_list_column(toclass_features, 1))
score_list = feat_list.T[0:1].T
lbl_list = [aid2_nid[aid] == nid for aid, nid in aidnid_pairs]
from sklearn import svm
#clf1 = svm.LinearSVC()
print('Learning classifiers')
clf3 = svm.SVC(probability=True)
clf3.fit(feat_list, lbl_list)
#prob_true, prob_false = clf3.predict_proba(feat_list).T
clf1 = svm.LinearSVC()
clf1.fit(score_list, lbl_list)
# Score new annots against the training database
tup_ = get_toy_annots(num_annots * 2,
num_names,
initial_aids=all_aids,
initial_nids=all_nids)
aids, nids, aids1, nids1, all_aids, all_nids = tup_
aid2_nid = dict(zip(aids, nids))
toclass_features = [annot_to_class_feats2(aid, aid2_nid) for aid in aids]
aidnid_pairs = np.vstack(ut.get_list_column(toclass_features, 0))
feat_list = np.vstack(ut.get_list_column(toclass_features, 1))
lbl_list = np.array([aid2_nid[aid] == nid for aid, nid in aidnid_pairs])
print('Running tests')
score_list = feat_list.T[0:1].T
tp_feat_list = feat_list[lbl_list]
tn_feat_list = feat_list[~lbl_list]
tp_lbls = lbl_list[lbl_list]
tn_lbls = lbl_list[~lbl_list]
print('num tp: %d' % len(tp_lbls))
print('num fp: %d' % len(tn_lbls))
tp_score_list = score_list[lbl_list]
tn_score_list = score_list[~lbl_list]
print('tp_feat' +
ut.repr3(ut.get_stats(tp_feat_list, axis=0), precision=2))
print('tp_feat' +
ut.repr3(ut.get_stats(tn_feat_list, axis=0), precision=2))
print('tp_score' + ut.repr2(ut.get_stats(tp_score_list), precision=2))
print('tp_score' + ut.repr2(ut.get_stats(tn_score_list), precision=2))
tp_pred3 = clf3.predict(tp_feat_list)
tn_pred3 = clf3.predict(tn_feat_list)
print((tp_pred3.sum(), tp_pred3.shape))
print((tn_pred3.sum(), tn_pred3.shape))
tp_score3 = clf3.score(tp_feat_list, tp_lbls)
tn_score3 = clf3.score(tn_feat_list, tn_lbls)
tp_pred1 = clf1.predict(tp_score_list)
tn_pred1 = clf1.predict(tn_score_list)
print((tp_pred1.sum(), tp_pred1.shape))
print((tn_pred1.sum(), tn_pred1.shape))
tp_score1 = clf1.score(tp_score_list, tp_lbls)
tn_score1 = clf1.score(tn_score_list, tn_lbls)
print('tp score with rank = %r' % (tp_score3, ))
print('tn score with rank = %r' % (tn_score3, ))
print('tp score without rank = %r' % (tp_score1, ))
print('tn score without rank = %r' % (tn_score1, ))
toy_data = {}
return toy_data
def get_toy_data_1v1(num_annots=5, num_names=None, **kwargs):
r"""
CommandLine:
python -m ibeis.algo.hots.demobayes --exec-get_toy_data_1v1 --show
Example:
>>> # ENABLE_DOCTEST
>>> from ibeis.algo.hots.demobayes import * # NOQA
>>> toy_data = get_toy_data_1v1()
>>> ut.quit_if_noshow()
>>> import plottool as pt
>>> show_toy_distributions(toy_data['toy_params'])
>>> ut.show_if_requested()
Example1:
>>> # ENABLE_DOCTEST
>>> from ibeis.algo.hots.demobayes import * # NOQA
>>> toy_data = get_toy_data_1v1()
>>> kwargs = {}
>>> initial_aids = toy_data['aids']
>>> initial_nids = toy_data['nids']
>>> num_annots = 1
>>> num_names = 6
>>> toy_data2 = get_toy_data_1v1(num_annots, num_names, initial_aids=initial_aids, initial_nids=initial_nids)
>>> ut.quit_if_noshow()
>>> import plottool as pt
>>> show_toy_distributions(toy_data['toy_params'])
>>> ut.show_if_requested()
Ignore:
>>> num_annots = 1000
>>> num_names = 400
"""
import vtool as vt
tup_ = get_toy_annots(num_annots, num_names, **kwargs)
aids, nids, aids1, nids1, all_aids, all_nids = tup_
rng = vt.ensure_rng(None)
def pairwise_feature(aidx1, aidx2, all_nids=all_nids, toy_params=toy_params):
if aidx1 == aidx2:
score = -1
else:
#rng = np.random.RandomState(int((aidx1 + 13) * (aidx2 + 13)))
nid1 = all_nids[int(aidx1)]
nid2 = all_nids[int(aidx2)]
params = toy_params[nid1 == nid2]
mu, sigma = ut.dict_take(params, ['mu', 'sigma'])
score_ = rng.normal(mu, sigma)
score = np.clip(score_, 0, np.inf)
return score
pairwise_nids = list([tup[::-1] for tup in ut.iprod(nids, nids1)])
pairwise_matches = np.array(
[nid1 == nid2 for nid1, nid2 in pairwise_nids])
pairwise_aidxs = list([tup[::-1] for tup in ut.iprod(aids, aids1)])
pairwise_features = np.array(
[pairwise_feature(aidx1, aidx2) for aidx1, aidx2 in pairwise_aidxs])
#pairwise_scores_mat = pairwise_scores.reshape(num_annots, num_annots)
is_diag = [r < c for r, c, in pairwise_aidxs]
diag_scores = pairwise_features.compress(is_diag)
diag_aidxs = ut.compress(pairwise_aidxs, is_diag)
import utool
with utool.embed_on_exception_context:
diag_nids = ut.compress(pairwise_nids, is_diag)
diag_labels = pairwise_matches.compress(is_diag)
#import utool
#utool.embed()
toy_data = {
'aids': aids,
'nids': nids,
'all_nids': all_nids,
'all_aids': all_aids,
#'pairwise_aidxs': pairwise_aidxs,
#'pairwise_scores': pairwise_scores,
#'pairwise_matches': pairwise_matches,
'diag_labels': diag_labels,
'diag_scores': diag_scores,
'diag_nids': diag_nids,
'diag_aidxs': diag_aidxs,
'toy_params': toy_params,
}
return toy_data
def get_toy_annots(num_annots, num_names=None, initial_aids=None, initial_nids=None, nid_sequence=None, seed=None):
r"""
Args:
num_annots (int):
num_names (int): (default = None)
initial_aids (None): (default = None)
initial_nids (None): (default = None)
nid_sequence (None): (default = None)
seed (None): (default = None)
Returns:
tuple: (aids, nids, aids1, nids1, all_aids, all_nids)
CommandLine:
python -m ibeis.algo.hots.demobayes --exec-get_toy_annots
Example:
>>> # DISABLE_DOCTEST
>>> from ibeis.algo.hots.demobayes import * # NOQA
>>> num_annots = 1
>>> num_names = 5
>>> initial_aids = np.array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9], dtype=np.int64)
>>> initial_nids = np.array([0, 0, 1, 2, 2, 1, 1, 1, 2, 3], dtype=np.int64)
>>> nid_sequence = np.array([0, 0, 1, 2, 2, 1, 1], dtype=np.int64)
>>> seed = 0
>>> (aids, nids, aids1, nids1, all_aids, all_nids) = get_toy_annots(num_annots, num_names, initial_aids, initial_nids, nid_sequence, seed)
>>> result = ('(aids, nids, aids1, nids1, all_aids, all_nids) = %s' % (ut.repr2((aids, nids, aids1, nids1, all_aids, all_nids), nl=1),))
>>> print(result)
"""
import vtool as vt
if num_names is None:
num_names = num_annots
print('Generating toy data with num_annots=%r' % (num_annots,))
if initial_aids is None:
assert initial_nids is None
first_step = True
initial_aids = []
initial_nids = []
else:
first_step = False
assert initial_nids is not None
aids = np.arange(len(initial_aids), num_annots + len(initial_aids))
rng = vt.ensure_rng(seed)
if nid_sequence is None:
nids = rng.randint(0, num_names, num_annots)
else:
unused_from_sequence = max(len(nid_sequence) - len(initial_aids), 0)
if unused_from_sequence == 0:
nids = rng.randint(0, num_names, num_annots)
elif unused_from_sequence > 0 and unused_from_sequence < num_annots:
num_remain = num_annots - unused_from_sequence
nids = np.append(nid_sequence[-unused_from_sequence:], rng.randint(0, num_names, num_remain))
else:
nids = nid_sequence[-unused_from_sequence]
nids = np.array(ut.take(nid_sequence, range(len(initial_aids), len(initial_aids) + num_annots)))
if first_step:
aids1 = aids
nids1 = nids
else:
aids1 = initial_aids
nids1 = initial_nids
all_nids = np.append(initial_nids, nids)
all_aids = np.append(initial_aids, aids)
import utool
with utool.embed_on_exception_context:
ut.assert_eq(len(aids), len(nids), 'len new')
ut.assert_eq(len(aids1), len(nids1), 'len comp')
ut.assert_eq(len(all_aids), len(all_nids), 'len all')
return aids, nids, aids1, nids1, all_aids, all_nids
def get_toy_data_1vM(num_annots, num_names=None, **kwargs):
r"""
Args:
num_annots (int):
num_names (int): (default = None)
Kwargs:
initial_aids, initial_nids, nid_sequence, seed
Returns:
tuple: (pair_list, feat_list)
CommandLine:
python -m ibeis.algo.hots.demobayes --exec-get_toy_data_1vM --show
Example:
>>> # DISABLE_DOCTEST
>>> from ibeis.algo.hots.demobayes import * # NOQA
>>> num_annots = 1000
>>> num_names = 40
>>> get_toy_data_1vM(num_annots, num_names)
>>> ut.quit_if_noshow()
>>> import plottool as pt
>>> ut.show_if_requested()
"""
import vtool as vt
tup_ = get_toy_annots(num_annots, num_names, **kwargs)
aids, nids, aids1, nids1, all_aids, all_nids = tup_
rng = vt.ensure_rng(None)
# Test a simple SVM classifier
nid2_nexemp = ut.dict_hist(nids1)
aid2_nid = dict(zip(aids, nids))
ut.fix_embed_globals()
#def add_to_globals(globals_, subdict):
# globals_.update(subdict)
unique_nids = list(nid2_nexemp.keys())
def annot_to_class_feats2(aid, aid2_nid, top=None):
pair_list = []
score_list = []
nexemplar_list = []
for nid in unique_nids:
label = (aid2_nid[aid] == nid)
num_exemplars = nid2_nexemp.get(nid, 0)
if num_exemplars == 0:
continue
params = toy_params[label]
mu, sigma = ut.dict_take(params, ['mu', 'sigma'])
score_ = rng.normal(mu, sigma, size=num_exemplars).max()
score = np.clip(score_, 0, np.inf)
pair_list.append((aid, nid))
score_list.append(score)
nexemplar_list.append(num_exemplars)
rank_list = ut.argsort(score_list, reverse=True)
feat_list = np.array([score_list, rank_list, nexemplar_list]).T
sortx = np.argsort(rank_list)
feat_list = feat_list.take(sortx, axis=0)
pair_list = np.array(pair_list).take(sortx, axis=0)
if top is not None:
feat_list = feat_list[:top]
pair_list = pair_list[0:top]
return pair_list, feat_list
toclass_features = [annot_to_class_feats2(aid, aid2_nid, top=5) for aid in aids]
aidnid_pairs = np.vstack(ut.get_list_column(toclass_features, 0))
feat_list = np.vstack(ut.get_list_column(toclass_features, 1))
score_list = feat_list.T[0:1].T
lbl_list = [aid2_nid[aid] == nid for aid, nid in aidnid_pairs]
from sklearn import svm
#clf1 = svm.LinearSVC()
print('Learning classifiers')
clf3 = svm.SVC()
clf3.fit(feat_list, lbl_list)
clf1 = svm.LinearSVC()
clf1.fit(score_list, lbl_list)
# Score new annots against the training database
tup_ = get_toy_annots(num_annots * 2, num_names, initial_aids=all_aids, initial_nids=all_nids)
aids, nids, aids1, nids1, all_aids, all_nids = tup_
aid2_nid = dict(zip(aids, nids))
toclass_features = [annot_to_class_feats2(aid, aid2_nid) for aid in aids]
aidnid_pairs = np.vstack(ut.get_list_column(toclass_features, 0))
feat_list = np.vstack(ut.get_list_column(toclass_features, 1))
lbl_list = np.array([aid2_nid[aid] == nid for aid, nid in aidnid_pairs])
print('Running tests')
score_list = feat_list.T[0:1].T
tp_feat_list = feat_list[lbl_list]
tn_feat_list = feat_list[~lbl_list]
tp_lbls = lbl_list[lbl_list]
tn_lbls = lbl_list[~lbl_list]
print('num tp: %d' % len(tp_lbls))
print('num fp: %d' % len(tn_lbls))
tp_score_list = score_list[lbl_list]
tn_score_list = score_list[~lbl_list]
print('tp_feat' + ut.repr3(ut.get_stats(tp_feat_list, axis=0), precision=2))
print('tp_feat' + ut.repr3(ut.get_stats(tn_feat_list, axis=0), precision=2))
print('tp_score' + ut.repr2(ut.get_stats(tp_score_list), precision=2))
print('tp_score' + ut.repr2(ut.get_stats(tn_score_list), precision=2))
tp_pred3 = clf3.predict(tp_feat_list)
tn_pred3 = clf3.predict(tn_feat_list)
print((tp_pred3.sum(), tp_pred3.shape))
print((tn_pred3.sum(), tn_pred3.shape))
tp_score3 = clf3.score(tp_feat_list, tp_lbls)
tn_score3 = clf3.score(tn_feat_list, tn_lbls)
tp_pred1 = clf1.predict(tp_score_list)
tn_pred1 = clf1.predict(tn_score_list)
print((tp_pred1.sum(), tp_pred1.shape))
print((tn_pred1.sum(), tn_pred1.shape))
tp_score1 = clf1.score(tp_score_list, tp_lbls)
tn_score1 = clf1.score(tn_score_list, tn_lbls)
print('tp score with rank = %r' % (tp_score3,))
print('tn score with rank = %r' % (tn_score3,))
print('tp score without rank = %r' % (tp_score1,))
print('tn score without rank = %r' % (tn_score1,))
toy_data = {}
return toy_data
