def _cached_pairwise_features(extr, edges):
"""
Create pairwise features for annotations in a test inference object
based on the features used to learn here
TODO: need a more systematic way of specifying which feature dimensions
need to be computed
Notes:
Given a edge (u, v), we need to:
* Check which classifiers we have
* Check which feat-cols the classifier needs,
and construct a configuration that can acheive that.
* Construct the chip/feat config
* Construct the vsone config
* Additional LNBNN enriching config
* Pairwise feature construction config
* Then we can apply the feature to the classifier
edges = [(1, 2)]
"""
edges = list(edges)
if extr.verbose:
print('[pairfeat] Requesting {} cached pairwise features'.format(
len(edges)))
# TODO: use object properties
if len(edges) == 0:
assert extr.feat_dims is not None, 'no edges and unset feat dims'
index = nxu.ensure_multi_index([], ('aid1', 'aid2'))
feats = pd.DataFrame(columns=extr.feat_dims, index=index)
return feats
else:
use_cache = not extr.need_lnbnn and len(edges) > 2
cache_dir = join(extr.ibs.get_cachedir(), 'infr_bulk_cache')
feat_cfgstr = extr._make_cfgstr(edges)
cacher = ub.Cacher('bulk_pairfeats_v3',
feat_cfgstr,
enabled=use_cache,
dpath=cache_dir,
verbose=extr.verbose - 3)
# if cacher.exists() and extr.verbose > 3:
# fpath = cacher.get_fpath()
# print('Load match cache size: {}'.format(
# ut.get_file_nBytes_str(fpath)))
data = cacher.tryload()
if data is None:
data = extr._make_pairwise_features(edges)
cacher.save(data)
# if cacher.enabled and extr.verbose > 3:
# fpath = cacher.get_fpath()
# print('Save match cache size: {}'.format(
# ut.get_file_nBytes_str(fpath)))
matches, feats = data
feats = extr._postprocess_feats(feats)
return feats
def predict_proba_df(verif, edges):
"""
CommandLine:
python -m ibeis.algo.graph.demo DummyVerif.predict_edges
Example:
>>> # ENABLE_DOCTEST
>>> from ibeis.algo.graph.demo import * # NOQA
>>> from ibeis.algo.graph import demo
>>> import networkx as nx
>>> kwargs = dict(num_pccs=40, size=2)
>>> infr = demo.demodata_infr(**kwargs)
>>> verif = infr.dummy_verif
>>> edges = list(infr.graph.edges())
>>> probs = verif.predict_proba_df(edges)
>>> #print('scores = %r' % (scores,))
>>> #hashid = ut.hash_data(scores)
>>> #print('hashid = %r' % (hashid,))
>>> #assert hashid == 'cdlkytilfeqgmtsihvhqwffmhczqmpil'
"""
infr = verif.infr
edges = list(it.starmap(verif.infr.e_, edges))
prob_cache = infr.task_probs['match_state']
is_miss = np.array([e not in prob_cache for e in edges])
# is_hit = ~is_miss
if np.any(is_miss):
miss_edges = ut.compress(edges, is_miss)
miss_truths = [verif._get_truth(edge) for edge in miss_edges]
grouped_edges = ut.group_items(miss_edges,
miss_truths,
sorted_=False)
# Need to make this determenistic too
states = [POSTV, NEGTV, INCMP]
for key in sorted(grouped_edges.keys()):
group = grouped_edges[key]
probs0 = randn(shape=[len(group)],
rng=verif.rng,
a_max=1,
a_min=0,
**verif.dummy_params[key])
# Just randomly assign other probs
probs1 = verif.rng.rand(len(group)) * (1 - probs0)
probs2 = 1 - (probs0 + probs1)
for edge, probs in zip(group, zip(probs0, probs1, probs2)):
prob_cache[edge] = ut.dzip(states, probs)
from ibeis.algo.graph import nx_utils as nxu
import pandas as pd
probs = pd.DataFrame(ut.take(prob_cache, edges),
index=nxu.ensure_multi_index(
edges, ('aid1', 'aid2')))
return probs
def match_state_df(infr, index):
""" Returns groundtruth state based on ibeis controller """
index = ensure_multi_index(index, ('aid1', 'aid2'))
aid_pairs = np.asarray(index.tolist())
aid_pairs = vt.ensure_shape(aid_pairs, (None, 2))
is_same = infr.is_same(aid_pairs)
is_comp = infr.is_comparable(aid_pairs)
match_state_df = pd.DataFrame.from_items([
(NEGTV, ~is_same & is_comp),
(POSTV, is_same & is_comp),
(INCMP, ~is_comp),
])
match_state_df.index = index
return match_state_df
def _make_pairwise_features(extr, edges):
"""
Construct matches and their pairwise features
CommandLine:
python -m ibeis.algo.verif.pairfeat _make_pairwise_features
Doctest:
>>> from ibeis.algo.verif.pairfeat import *
>>> from ibeis.algo.graph import demo
>>> infr = demo.demodata_mtest_infr()
>>> extr = PairwiseFeatureExtractor(ibs=infr.ibs)
>>> match_config = {'K': 1, 'Knorm': 3, 'affine_invariance': True,
>>> 'augment_orientation': True, 'checks': 20,
>>> 'ratio_thresh': 0.8, 'refine_method': 'homog',
>>> 'sv_on': True, 'sver_xy_thresh': 0.01,
>>> 'symmetric': True, 'weight': 'fgweights'}
>>> local_keys = [
>>> 'fgweights', 'match_dist', 'norm_dist', 'norm_x1', 'norm_x2',
>>> 'norm_y1', 'norm_y2', 'ratio_score', 'scale1', 'scale2',
>>> 'sver_err_ori', 'sver_err_scale', 'sver_err_xy',
>>> 'weighted_norm_dist', 'weighted_ratio_score']
>>> pairfeat_cfg = {
>>> 'bin_key': 'ratio',
>>> 'bins': [0.6, 0.7, 0.8],
>>> 'indices': [],
>>> 'local_keys': local_keys,
>>> 'sorters': [],
>>> 'summary_ops': {'len', 'mean', 'sum'}
>>> }
>>> global_keys = ['gps', 'qual', 'time', 'view']
>>> ibs = infr.ibs
>>> extr = PairwiseFeatureExtractor(ibs, match_config=match_config,
>>> pairfeat_cfg=pairfeat_cfg,
>>> global_keys=global_keys)
>>> multi_index = True
>>> edges = [(1, 2), (2, 3)]
>>> matches, X = extr._make_pairwise_features(edges)
>>> featinfo = vt.AnnotPairFeatInfo(X.columns)
>>> print(featinfo.get_infostr())
>>> match = matches[0]
>>> glob_X = match._make_global_feature_vector(global_keys)
>>> assert len(glob_X) == 19
"""
edges = ut.lmap(tuple, ut.aslist(edges))
if len(edges) == 0:
return [], []
matches = extr._enriched_pairwise_matches(edges)
# ---------------
# Try different feature constructions
print('[extr] building pairwise features')
pairfeat_cfg = extr.pairfeat_cfg.copy()
use_na = pairfeat_cfg.pop('use_na')
pairfeat_cfg['summary_ops'] = set(pairfeat_cfg['summary_ops'])
X = pd.DataFrame([
m.make_feature_vector(**pairfeat_cfg)
for m in ut.ProgIter(matches, label='making pairwise feats')
])
multi_index = True
if multi_index:
# Index features by edges
uv_index = nxu.ensure_multi_index(edges, ('aid1', 'aid2'))
X.index = uv_index
X[pd.isnull(X)] = np.nan
X[np.isinf(X)] = np.nan
# Re-order column names to ensure dimensions are consistent
X = X.reindex_axis(sorted(X.columns), axis=1)
# hack to fix feature validity
if 'global(speed)' in X.columns:
if np.any(np.isinf(X['global(speed)'])):
flags = np.isinf(X['global(speed)'])
numer = X.loc[flags, 'global(gps_delta)']
denom = X.loc[flags, 'global(time_delta)']
newvals = np.full(len(numer), np.nan)
newvals[(numer == 0) & (denom == 0)] = 0
X.loc[flags, 'global(speed)'] = newvals
aid_pairs_ = [(m.annot1['aid'], m.annot2['aid']) for m in matches]
assert aid_pairs_ == edges, 'edge ordering changed'
if not use_na:
# Fill nan values with very large values to workaround lack of nan
# support in sklearn master.
X[pd.isnull(X)] = (2 ** 30) - 1
return matches, X
def ensure_priority_scores(infr, priority_edges):
"""
Ensures that priority attributes are assigned to the edges.
This does not change the state of the queue.
Doctest:
>>> import ibeis
>>> ibs = ibeis.opendb('PZ_MTEST')
>>> infr = ibeis.AnnotInference(ibs, aids='all')
>>> infr.ensure_mst()
>>> priority_edges = list(infr.edges())[0:1]
>>> infr.ensure_priority_scores(priority_edges)
Doctest:
>>> import ibeis
>>> ibs = ibeis.opendb('PZ_MTEST')
>>> infr = ibeis.AnnotInference(ibs, aids='all')
>>> infr.ensure_mst()
>>> infr.load_published()
>>> priority_edges = list(infr.edges())
>>> infr.ensure_priority_scores(priority_edges)
Doctest:
>>> from ibeis.algo.graph import demo
>>> infr = demo.demodata_infr(num_pccs=6, p_incon=.5, size_std=2)
>>> edges = list(infr.edges())
>>> infr.ensure_priority_scores(edges)
"""
infr.print('Checking for verifiers: %r' % (infr.verifiers, ))
if infr.verifiers:
infr.print(
'Prioritizing {} edges with one-vs-one probs'.format(
len(priority_edges)), 1)
infr.print('Using thresholds: %r' % (infr.task_thresh, ))
infr.print('Using infr.params[autoreview.enabled] : %r' %
(infr.params['autoreview.enabled'], ))
infr.print('Using infr.params[autoreview.prioritize_nonpos]: %r' %
(infr.params['autoreview.prioritize_nonpos'], ))
infr.ensure_task_probs(priority_edges)
primary_task = 'match_state'
match_probs = infr.task_probs[primary_task]
primary_thresh = infr.task_thresh[primary_task]
# Read match_probs into a DataFrame
primary_probs = pd.DataFrame(ut.take(match_probs, priority_edges),
index=nxu.ensure_multi_index(
priority_edges, ('aid1', 'aid2')))
# Convert match-state probabilities into priorities
prob_match = primary_probs[POSTV]
# Initialize priorities to probability of matching
default_priority = prob_match.copy()
# If the edges are currently between the same individual, then
# prioritize by non-positive probability (because those edges might
# expose an inconsistency)
already_pos = [
infr.pos_graph.node_label(u) == infr.pos_graph.node_label(v)
for u, v in priority_edges
]
default_priority[already_pos] = 1 - default_priority[already_pos]
if infr.params['autoreview.enabled']:
if infr.params['autoreview.prioritize_nonpos']:
# Give positives that pass automatic thresholds high priority
_probs = primary_probs[POSTV]
flags = _probs > primary_thresh[POSTV]
default_priority[flags] = np.maximum(
default_priority[flags], _probs[flags]) + 1
# Give negatives that pass automatic thresholds high priority
_probs = primary_probs[NEGTV]
flags = _probs > primary_thresh[NEGTV]
default_priority[flags] = np.maximum(
default_priority[flags], _probs[flags]) + 1
# Give not-comps that pass automatic thresholds high priority
_probs = primary_probs[INCMP]
flags = _probs > primary_thresh[INCMP]
default_priority[flags] = np.maximum(
default_priority[flags], _probs[flags]) + 1
infr.set_edge_attrs('prob_match', prob_match.to_dict())
infr.set_edge_attrs('default_priority', default_priority.to_dict())
metric = 'default_priority'
priority = default_priority
elif infr.cm_list is not None:
infr.print(
'Prioritizing {} edges with one-vs-vsmany scores'.format(
len(priority_edges), 1))
# Not given any deploy classifier, this is the best we can do
scores = infr._make_lnbnn_scores(priority_edges)
metric = 'normscore'
priority = scores
else:
infr.print('WARNING: No verifiers to prioritize {} edge(s)'.format(
len(priority_edges)))
metric = 'random'
priority = np.zeros(len(priority_edges)) + 1e-6
infr.set_edge_attrs(metric, ut.dzip(priority_edges, priority))
return metric, priority