File size: 11,483 Bytes
347d9a3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
"""Formal dynamic evaluation with full inventory denominators and fixed identities."""
from __future__ import annotations
import argparse,copy,gzip,hashlib,json,math,sqlite3,tempfile
from collections import defaultdict
from pathlib import Path
from .frame_score import score as demo_score
from .common import VERSION,rows,write


def safe_score(gold,prediction):
    """Bound malformed coordinates/times before invoking the tested frame scorer."""
    invalid=False;p=copy.deepcopy(prediction)
    if gold['task_id']=='KW_ASR':
        if isinstance(p,str):
            p=p.replace('’',"'").replace('‘',"'")
            if p.strip()=='""':p=''
        result=demo_score(gold,p)
        if not gold['output']:
            correct=isinstance(p,str) and not p.strip()
            result['empty_correct']=correct;result['exact_match']=correct
        return result
    sets=[]
    if isinstance(p,dict):
        if gold['task_id']=='DOA_TRAJECTORY' and isinstance(p.get('tracks'),list):
            sets=[t for t in p['tracks'] if isinstance(t,dict)]
        elif isinstance(p.get('frames'),list):sets=[p]
    maximum=max(gold['expected_times_s'])
    for container in sets:
        if not isinstance(container.get('frames'),list):continue
        filtered=[]
        for f in container['frames']:
            if not isinstance(f,dict):invalid=True;continue
            t=f.get('time_s')
            if isinstance(t,bool) or not isinstance(t,(int,float)) or not math.isfinite(t) or not 0<=t<=maximum+1e-6:
                invalid=True;continue
            if f.get('active') is True:
                key='doa_deg' if gold['task_id']=='DOA_TRAJECTORY' else 'distance_m' if gold['task_id']=='RANGE_TRAJECTORY' else 'xyz_m'
                value=f.get(key);v=[value] if key=='distance_m' else value
                if not isinstance(v,list) or any(isinstance(x,bool) or not isinstance(x,(int,float)) or not math.isfinite(x) or abs(x)>1000 for x in v):
                    invalid=True;continue
                if key=='doa_deg' and (len(v)!=2 or not -180<=v[0]<=180 or not -90<=v[1]<=90):invalid=True;continue
            filtered.append(f)
        container['frames']=filtered
    try:result=demo_score(gold,p)
    except (ValueError,TypeError,OverflowError,IndexError,KeyError):
        result=demo_score(gold,None);invalid=True
    if invalid:result['format_valid']=False
    return result


class Accumulator:
    def __init__(self):self.values=defaultdict(float);self.case_counts=defaultdict(int)
    def add(self,record):
        v=self.values;v['samples']+=1;v['format_valid']+=record['format_valid']
        if record.get('keyword_case') is not None:
            v['reference_words']+=record['reference_words'];v['edits']+=record['edits'];v['exact_match']+=record['exact_match']
            self.case_counts[record['keyword_case']]+=1
            if not record['reference_words']:v['negative_samples']+=1;v['empty_correct']+=bool(record['empty_correct'])
            return
        v['reference_active_frames']+=record['reference_active_frames'];v['covered_frames']+=record['covered_frames']
        n=record['reference_active_frames'] if record['units'].startswith('degrees') else record['covered_frames']
        if n and record['mean_error'] is not None:
            v['error_count']+=n;v['error_sum']+=record['mean_error']*n;v['error_square_sum']+=(record['variance_error']+record['mean_error']**2)*n
        threshold='success_rate_20deg' if record['units'].startswith('degrees') else 'success_rate_0_5m'
        v['successful_frames']+=record[threshold]*record['reference_active_frames'];v['missed_frames']+=record['missed_frames'];v['false_positive_frames']+=record['false_positive_frames']
        v['true_positive_frames']+=record['covered_frames'];v['identity_switches']+=record.get('identity_switches_diagnostic',0)
    def finish(self):
        v=self.values;n=v['samples'];out=dict(samples=int(n),format_valid_fraction=v['format_valid']/max(n,1))
        if self.case_counts:
            refs=v['reference_words'];negs=v['negative_samples']
            out.update(corpus_wer=v['edits']/refs if refs else None,reference_words=int(refs),edits=int(v['edits']),
                exact_match_fraction=v['exact_match']/max(n,1),negative_samples=int(negs),negative_empty_accuracy=v['empty_correct']/negs if negs else None,
                keyword_cases=dict(self.case_counts))
        else:
            count=v['error_count'];mean=v['error_sum']/count if count else None
            out.update(reference_active_frames=int(v['reference_active_frames']),covered_frames=int(v['covered_frames']),
                coverage=v['covered_frames']/max(v['reference_active_frames'],1),success_rate=v['successful_frames']/max(v['reference_active_frames'],1),
                mean_error=mean,variance_error=max(0.,v['error_square_sum']/count-mean**2) if count else None,
                rmse=math.sqrt(v['error_square_sum']/count) if count else None,
                error_count=int(count),active_precision=v['true_positive_frames']/max(v['true_positive_frames']+v['false_positive_frames'],1),
                active_recall=v['true_positive_frames']/max(v['true_positive_frames']+v['missed_frames'],1),
                missed_frames=int(v['missed_frames']),false_positive_frames=int(v['false_positive_frames']),identity_switches_diagnostic=int(v['identity_switches']))
        return out


def freeze_predictions(path,known,database,strict=False):
    database.execute('CREATE TABLE predictions (sample_id TEXT PRIMARY KEY, output TEXT)')
    bad=set();counts=defaultdict(int);digest=hashlib.sha256()
    with Path(path).open('rb') as f:
        for number,line in enumerate(f,1):
            digest.update(line)
            if not line.strip():continue
            try:
                r=json.loads(line);sid=r.get('sample_id');output=r.get('output')
                if sid not in known:raise ValueError('Unknown sample_id')
                if sid in bad:counts['duplicate_rows']+=1;continue
                if database.execute('SELECT 1 FROM predictions WHERE sample_id=?',(sid,)).fetchone():
                    counts['duplicate_rows']+=1;bad.add(sid);database.execute('DELETE FROM predictions WHERE sample_id=?',(sid,))
                    if strict:raise ValueError('Duplicate sample_id')
                    continue
                database.execute('INSERT INTO predictions VALUES (?,?)',(sid,json.dumps(output)))
            except (ValueError,TypeError,AttributeError) as e:
                counts['invalid_rows']+=1
                if strict:raise ValueError('Invalid prediction row '+str(number)+': '+str(e)) from e
            if number%2000==0:database.commit()
    database.commit();counts['accepted_rows']=database.execute('SELECT COUNT(*) FROM predictions').fetchone()[0]
    return dict(prediction_sha256=digest.hexdigest(),counts=dict(counts),duplicate_ids=len(bad))


def evaluate(root,predictions,output,strict=False):
    root=Path(root);output=Path(output);output.mkdir(parents=True,exist_ok=True)
    known={r['sample_id'] for r in rows(root/'public/test/index.jsonl')}
    buckets=defaultdict(Accumulator);missing=total=0;kw_pairs={}
    with tempfile.TemporaryDirectory(prefix='dish-evaluate-') as temp:
        db=sqlite3.connect(str(Path(temp)/'predictions.sqlite'))
        inventory=freeze_predictions(predictions,known,db,strict)
        # Gold is first opened only after the prediction inventory has been frozen.
        with gzip.open(output/'per_sample.jsonl.gz','wt') as per_sample:
            for g in rows(root/'private/test/gold.jsonl'):
                found=db.execute('SELECT output FROM predictions WHERE sample_id=?',(g['sample_id'],)).fetchone()
                p=json.loads(found[0]) if found else None;missing+=found is None;total+=1
                result=safe_score(g,p);record=dict(sample_id=g['sample_id'],scene_id=g['scene_id'],task_id=g['task_id'],setting=g['setting'],
                    branch=g['branch'],difficulty=g['difficulty'],**result)
                record.update(target_human_motion=g.get('target_human_motion','unspecified'),observed_receiver_motion=g.get('observed_receiver_motion','unspecified'))
                per_sample.write(json.dumps(record,allow_nan=False,separators=(',',':'))+'\n')
                keys=[g['setting']+'/'+g['task_id'],g['setting']+'/'+g['task_id']+'/'+g['difficulty'],
                    g['setting']+'/'+g['task_id']+'/branch/'+g['branch'],g['setting']+'/'+g['task_id']+'/robots/'+str(g['num_robots']),
                    g['setting']+'/'+g['task_id']+'/humans/'+str(g['num_humans']),g['setting']+'/'+g['task_id']+'/rt_bin/'+str(g['rt_bin'])]
                keys.append(g['setting']+'/'+g['task_id']+'/motion/target_'+g.get('target_human_motion','unspecified')+'/receiver_'+g.get('observed_receiver_motion','unspecified'))
                if g['task_id']=='KW_ASR':keys.append(g['setting']+'/KW_ASR/case/'+g.get('keyword_observation_case',g['keyword_case']))
                for key in keys:buckets[key].add(result)
                if g['task_id']=='KW_ASR':kw_pairs.setdefault(g['query_group_id'],[]).append(dict(setting=g['setting'],wer=result['wer'],correct=result['exact_match']))
        db.close()
    if total!=len(known):raise ValueError('Gold/public inventory mismatch')
    gains=[];accuracy_gains=[]
    for values in kw_pairs.values():
        edges=[r for r in values if r['setting']=='edge'];cloud=next(r for r in values if r['setting']=='cloud')
        if cloud['wer'] is not None:gains.append(sum(e['wer'] for e in edges)/len(edges)-cloud['wer'])
        accuracy_gains.append(float(cloud['correct'])-sum(float(e['correct']) for e in edges)/len(edges))
    report=dict(scorer_version=VERSION+'.1',samples=total,missing_predictions=missing,inventory=inventory,
        summary={k:v.finish() for k,v in sorted(buckets.items())},
        paired_kw_asr=dict(query_groups=len(kw_pairs),positive_groups=len(gains),
            mean_cloud_wer_improvement_over_mean_edge=sum(gains)/len(gains) if gains else None,
            mean_cloud_exact_match_improvement_over_mean_edge=sum(accuracy_gains)/len(accuracy_gains) if accuracy_gains else None),
        definitions=dict(doa='Whole-clip fixed identity matching; active-frame spherical error, mean/population variance/RMSE. Missing DOA positions count as 180 degrees.',
            distance_and_world='Error statistics on covered active frames, always accompanied by all-reference-frame coverage and 0.5m success.',
            kw_asr='Positive corpus WER; negative empty-string accuracy; invalid and missing outputs remain failures.',
            error_statistics='Frame-pooled statistics, not averages of per-example variances.',
            sampling_unit='Paired motion group/room and original source/speaker; views, questions and overlapping speech crops are correlated. No cross-task overall score.',
            motion_strata='Target-human and observed-receiver motion are reported separately. Motion elsewhere in the scene does not imply either selected entity moves.'))
    write(output/'metrics.json',report);print('SCORED',total,'samples;',missing,'missing',flush=True)
    return report


if __name__=='__main__':
    p=argparse.ArgumentParser();p.add_argument('--root',type=Path,required=True);p.add_argument('--predictions',type=Path,required=True);p.add_argument('--output',type=Path,required=True);p.add_argument('--strict',action='store_true')
    a=p.parse_args();evaluate(a.root,a.predictions,a.output,a.strict)