#!/usr/bin/env python # encoding: utf-8 # The MIT License (MIT) # Copyright (c) 2017-2020 CNRS # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE # SOFTWARE. # AUTHORS # Herve BREDIN - http://herve.niderb.fr """ Evaluation Usage: pyannote-metrics detection [--subset= --collar= --skip-overlap] pyannote-metrics segmentation [--subset= --tolerance=] pyannote-metrics overlap [--subset= --collar=] pyannote-metrics diarization [--subset= --greedy --collar= --skip-overlap] pyannote-metrics identification [--subset= --collar= --skip-overlap] pyannote-metrics spotting [--subset= --latency=... --filter=...] pyannote-metrics -h | --help pyannote-metrics --version Options: Set evaluation protocol (e.g. "Etape.SpeakerDiarization.TV") --subset= Evaluated subset (train|developement|test) [default: test] --collar= Collar, in seconds [default: 0.0]. --skip-overlap Do not evaluate overlap regions. --tolerance= Tolerance, in seconds [default: 0.5]. --greedy Use greedy diarization error rate. --latency= Evaluate with fixed latency. --filter= Filter out target trials that do not match the expression; e.g. use --filter="speech>10" to skip target trials with less than 10s of speech from the target. -h --help Show this screen. --version Show version. All modes but "spotting" expect hypothesis using the RTTM file format. RTTM files contain one line per speech turn, using the following convention: SPEAKER {uri} 1 {start_time} {duration} {speaker_id} * uri: file identifier (as given by pyannote.database protocols) * start_time: speech turn start time in seconds * duration: speech turn duration in seconds * speaker_id: speaker identifier "spotting" mode expects hypothesis using the following JSON file format. It should contain a list of trial hypothesis, using the same trial order as pyannote.database speaker spotting protocols (e.g. protocol.test_trial()) [ {'uri': '', 'model_id': '', 'scores': [[, ], [, ], ... [, ]]}, {'uri': '', 'model_id': '', 'scores': [[, ], [, ], ... [, ]]}, {'uri': '', 'model_id': '', 'scores': [[, ], [, ], ... [, ]]}, ... {'uri': '', 'model_id': '', 'scores': [[, ], [, ], ... [, ]]}, ] * uri: file identifier (as given by pyannote.database protocols) * model_id: target identifier (as given by pyannote.database protocols) * [ti, vi]: [time, value] pair indicating that the system has output the score vi at time ti (e.g. [10.2, 0.2] means that the system gave a score of 0.2 at time 10.2s). Calling "spotting" mode will create a bunch of files. * contains DET curve using the following raw file format: * contains latency curves using this format: """ import functools import json import sys import warnings import numpy as np import pandas as pd # command line parsing from docopt import docopt from pyannote.core import Annotation from pyannote.core import Timeline # evaluation protocols from pyannote.database import get_protocol from pyannote.database.util import get_annotated from pyannote.database.util import load_rttm from tabulate import tabulate from pyannote.metrics.detection import DetectionAccuracy from pyannote.metrics.detection import DetectionErrorRate from pyannote.metrics.detection import DetectionPrecision from pyannote.metrics.detection import DetectionRecall from pyannote.metrics.diarization import DiarizationCoverage from pyannote.metrics.diarization import DiarizationErrorRate from pyannote.metrics.diarization import DiarizationPurity from pyannote.metrics.diarization import GreedyDiarizationErrorRate from pyannote.metrics.identification import IdentificationErrorRate from pyannote.metrics.identification import IdentificationPrecision from pyannote.metrics.identification import IdentificationRecall from pyannote.metrics.segmentation import SegmentationCoverage from pyannote.metrics.segmentation import SegmentationPrecision from pyannote.metrics.segmentation import SegmentationPurity from pyannote.metrics.segmentation import SegmentationRecall from pyannote.metrics.spotting import LowLatencySpeakerSpotting showwarning_orig = warnings.showwarning def showwarning(message, category, *args, **kwargs): print(category.__name__ + ":", str(message)) warnings.showwarning = showwarning def to_overlap(current_file: dict) -> Annotation: """Get overlapped speech reference annotation Parameters ---------- current_file : `dict` File yielded by pyannote.database protocols. Returns ------- overlap : `pyannote.core.Annotation` Overlapped speech reference. """ reference = current_file["annotation"] overlap = Timeline(uri=reference.uri) for (s1, t1), (s2, t2) in reference.co_iter(reference): l1 = reference[s1, t1] l2 = reference[s2, t2] if l1 == l2: continue overlap.add(s1 & s2) return overlap.support().to_annotation() def get_hypothesis(hypotheses, current_file): """Get hypothesis for given file Parameters ---------- hypotheses : `dict` Speaker diarization hypothesis provided by `load_rttm`. current_file : `dict` File description as given by pyannote.database protocols. Returns ------- hypothesis : `pyannote.core.Annotation` Hypothesis corresponding to `current_file`. """ uri = current_file["uri"] if uri in hypotheses: return hypotheses[uri] # if the exact 'uri' is not available in hypothesis, # look for matching substring tmp_uri = [u for u in hypotheses if u in uri] # no matching speech turns. return empty annotation if len(tmp_uri) == 0: msg = f'Could not find hypothesis for file "{uri}"; assuming empty file.' warnings.warn(msg) return Annotation(uri=uri, modality="speaker") # exactly one matching file. return it if len(tmp_uri) == 1: hypothesis = hypotheses[tmp_uri[0]] hypothesis.uri = uri return hypothesis # more that one matching file. error. msg = f'Found too many hypotheses matching file "{uri}" ({tmp_uri}).' raise ValueError(msg.format(uri=uri, uris=tmp_uri)) def process_one(item, hypotheses=None, metrics=None): reference = item["annotation"] hypothesis = get_hypothesis(hypotheses, item) uem = get_annotated(item) return { key: metric(reference, hypothesis, uem=uem) for key, metric in metrics.items() } def get_reports(protocol, subset, hypotheses, metrics): process = functools.partial(process_one, hypotheses=hypotheses, metrics=metrics) for item in getattr(protocol, subset)(): process(item) return {key: metric.report(display=False) for key, metric in metrics.items()} def reindex(report): """Reindex report so that 'TOTAL' is the last row""" index = list(report.index) i = index.index("TOTAL") return report.reindex(index[:i] + index[i + 1 :] + ["TOTAL"]) def detection(protocol, subset, hypotheses, collar=0.0, skip_overlap=False): options = {"collar": collar, "skip_overlap": skip_overlap} metrics = { "error": DetectionErrorRate(**options), "accuracy": DetectionAccuracy(**options), "precision": DetectionPrecision(**options), "recall": DetectionRecall(**options), } reports = get_reports(protocol, subset, hypotheses, metrics) report = metrics["error"].report(display=False) accuracy = metrics["accuracy"].report(display=False) precision = metrics["precision"].report(display=False) recall = metrics["recall"].report(display=False) report["accuracy", "%"] = accuracy[metrics["accuracy"].name, "%"] report["precision", "%"] = precision[metrics["precision"].name, "%"] report["recall", "%"] = recall[metrics["recall"].name, "%"] report = reindex(report) columns = list(report.columns) report = report[[columns[0]] + columns[-3:] + columns[1:-3]] summary = "Detection (collar = {0:g} ms{1})".format( 1000 * collar, ", no overlap" if skip_overlap else "" ) headers = ( [summary] + [report.columns[i][0] for i in range(4)] + ["%" if c[1] == "%" else c[0] for c in report.columns[4:]] ) print( tabulate( report, headers=headers, tablefmt="simple", floatfmt=".2f", numalign="decimal", stralign="left", missingval="", showindex="default", disable_numparse=False, ) ) def segmentation(protocol, subset, hypotheses, tolerance=0.5): options = {"tolerance": tolerance} metrics = { "coverage": SegmentationCoverage(**options), "purity": SegmentationPurity(**options), "precision": SegmentationPrecision(**options), "recall": SegmentationRecall(**options), } reports = get_reports(protocol, subset, hypotheses, metrics) coverage = metrics["coverage"].report(display=False) purity = metrics["purity"].report(display=False) precision = metrics["precision"].report(display=False) recall = metrics["recall"].report(display=False) coverage = coverage[metrics["coverage"].name] purity = purity[metrics["purity"].name] precision = precision[metrics["precision"].name] recall = recall[metrics["recall"].name] report = pd.concat([coverage, purity, precision, recall], axis=1) report = reindex(report) headers = [ "Segmentation (tolerance = {0:g} ms)".format(1000 * tolerance), "coverage", "purity", "precision", "recall", ] print( tabulate( report, headers=headers, tablefmt="simple", floatfmt=".2f", numalign="decimal", stralign="left", missingval="", showindex="default", disable_numparse=False, ) ) def diarization( protocol, subset, hypotheses, greedy=False, collar=0.0, skip_overlap=False ): options = {"collar": collar, "skip_overlap": skip_overlap} metrics = { "purity": DiarizationPurity(**options), "coverage": DiarizationCoverage(**options), } if greedy: metrics["error"] = GreedyDiarizationErrorRate(**options) else: metrics["error"] = DiarizationErrorRate(**options) reports = get_reports(protocol, subset, hypotheses, metrics) report = metrics["error"].report(display=False) purity = metrics["purity"].report(display=False) coverage = metrics["coverage"].report(display=False) report["purity", "%"] = purity[metrics["purity"].name, "%"] report["coverage", "%"] = coverage[metrics["coverage"].name, "%"] columns = list(report.columns) report = report[[columns[0]] + columns[-2:] + columns[1:-2]] report = reindex(report) summary = "Diarization ({0:s}collar = {1:g} ms{2})".format( "greedy, " if greedy else "", 1000 * collar, ", no overlap" if skip_overlap else "", ) headers = ( [summary] + [report.columns[i][0] for i in range(3)] + ["%" if c[1] == "%" else c[0] for c in report.columns[3:]] ) print( tabulate( report, headers=headers, tablefmt="simple", floatfmt=".2f", numalign="decimal", stralign="left", missingval="", showindex="default", disable_numparse=False, ) ) def identification(protocol, subset, hypotheses, collar=0.0, skip_overlap=False): options = {"collar": collar, "skip_overlap": skip_overlap} metrics = { "error": IdentificationErrorRate(**options), "precision": IdentificationPrecision(**options), "recall": IdentificationRecall(**options), } reports = get_reports(protocol, subset, hypotheses, metrics) report = metrics["error"].report(display=False) precision = metrics["precision"].report(display=False) recall = metrics["recall"].report(display=False) report["precision", "%"] = precision[metrics["precision"].name, "%"] report["recall", "%"] = recall[metrics["recall"].name, "%"] columns = list(report.columns) report = report[[columns[0]] + columns[-2:] + columns[1:-2]] report = reindex(report) summary = "Identification (collar = {0:g} ms{1})".format( 1000 * collar, ", no overlap" if skip_overlap else "" ) headers = ( [summary] + [report.columns[i][0] for i in range(3)] + ["%" if c[1] == "%" else c[0] for c in report.columns[3:]] ) print( tabulate( report, headers=headers, tablefmt="simple", floatfmt=".2f", numalign="decimal", stralign="left", missingval="", showindex="default", disable_numparse=False, ) ) def spotting(protocol, subset, latencies, hypotheses, output_prefix, filter_func=None): if not latencies: Scores = [] protocol.diarization = False trials = getattr(protocol, "{subset}_trial".format(subset=subset))() for i, (current_trial, hypothesis) in enumerate(zip(trials, hypotheses)): # check trial/hypothesis target consistency try: assert current_trial["model_id"] == hypothesis["model_id"] except AssertionError as e: msg = ( "target mismatch in trial #{i} " "(found: {found}, should be: {should_be})" ) raise ValueError( msg.format( i=i, found=hypothesis["model_id"], should_be=current_trial["model_id"], ) ) # check trial/hypothesis file consistency try: assert current_trial["uri"] == hypothesis["uri"] except AssertionError as e: msg = ( "file mismatch in trial #{i} " "(found: {found}, should be: {should_be})" ) raise ValueError( msg.format(i=i, found=hypothesis["uri"], should_be=current_trial["uri"]) ) # check at least one score is provided try: assert len(hypothesis["scores"]) > 0 except AssertionError as e: msg = "empty list of scores in trial #{i}." raise ValueError(msg.format(i=i)) timestamps, scores = zip(*hypothesis["scores"]) if not latencies: Scores.append(scores) # check trial/hypothesis timerange consistency try_with = current_trial["try_with"] try: assert min(timestamps) >= try_with.start except AssertionError as e: msg = ( "incorrect timestamp in trial #{i} " "(found: {found:g}, should be: >= {should_be:g})" ) raise ValueError( msg.format(i=i, found=min(timestamps), should_be=try_with.start) ) if not latencies: # estimate best set of thresholds scores = np.concatenate(Scores) epsilons = np.array( [n * 10 ** (-e) for e in range(4, 1, -1) for n in range(1, 10)] ) percentile = np.concatenate( [epsilons, np.arange(0.1, 100.0, 0.1), 100 - epsilons[::-1]] ) thresholds = np.percentile(scores, percentile) if not latencies: metric = LowLatencySpeakerSpotting(thresholds=thresholds) else: metric = LowLatencySpeakerSpotting(latencies=latencies) trials = getattr(protocol, "{subset}_trial".format(subset=subset))() for i, (current_trial, hypothesis) in enumerate(zip(trials, hypotheses)): if filter_func is not None: speech = current_trial["reference"].duration() target_trial = speech > 0 if target_trial and filter_func(speech): continue reference = current_trial["reference"] metric(reference, hypothesis["scores"]) if not latencies: thresholds, fpr, fnr, eer, _ = metric.det_curve(return_latency=False) # save DET curve to hypothesis.det.txt det_path = "{output_prefix}.det.txt".format(output_prefix=output_prefix) det_tmpl = "{t:.9f} {p:.9f} {n:.9f}\n" with open(det_path, mode="w") as fp: fp.write("# threshold false_positive_rate false_negative_rate\n") for t, p, n in zip(thresholds, fpr, fnr): line = det_tmpl.format(t=t, p=p, n=n) fp.write(line) print("> {det_path}".format(det_path=det_path)) thresholds, fpr, fnr, _, _, speaker_lcy, absolute_lcy = metric.det_curve( return_latency=True ) # save DET curve to hypothesis.det.txt lcy_path = "{output_prefix}.lcy.txt".format(output_prefix=output_prefix) lcy_tmpl = "{t:.9f} {p:.9f} {n:.9f} {s:.6f} {a:.6f}\n" with open(lcy_path, mode="w") as fp: fp.write( "# threshold false_positive_rate false_negative_rate speaker_latency absolute_latency\n" ) for t, p, n, s, a in zip(thresholds, fpr, fnr, speaker_lcy, absolute_lcy): if p == 1: continue if np.isnan(s): continue line = lcy_tmpl.format(t=t, p=p, n=n, s=s, a=a) fp.write(line) print("> {lcy_path}".format(lcy_path=lcy_path)) print() print("EER% = {eer:.2f}".format(eer=100 * eer)) else: results = metric.det_curve() logs = [] for key in sorted(results): result = results[key] log = {"latency": key} for latency in latencies: thresholds, fpr, fnr, eer, _ = result[latency] # print('EER @ {latency}s = {eer:.2f}%'.format(latency=latency, # eer=100 * eer)) log[latency] = eer # save DET curve to hypothesis.det.{lcy}s.txt det_path = "{output_prefix}.det.{key}.{latency:g}s.txt".format( output_prefix=output_prefix, key=key, latency=latency ) det_tmpl = "{t:.9f} {p:.9f} {n:.9f}\n" with open(det_path, mode="w") as fp: fp.write("# threshold false_positive_rate false_negative_rate\n") for t, p, n in zip(thresholds, fpr, fnr): line = det_tmpl.format(t=t, p=p, n=n) fp.write(line) logs.append(log) det_path = "{output_prefix}.det.{key}.XXs.txt".format( output_prefix=output_prefix, key=key ) print("> {det_path}".format(det_path=det_path)) print() df = 100 * pd.DataFrame.from_dict(logs).set_index("latency")[latencies] print( tabulate( df, tablefmt="simple", headers=["latency"] + ["EER% @ {l:g}s".format(l=l) for l in latencies], floatfmt=".2f", numalign="decimal", stralign="left", missingval="", showindex="default", disable_numparse=False, ) ) def main(): arguments = docopt(__doc__, version="Evaluation") collar = float(arguments["--collar"]) skip_overlap = arguments["--skip-overlap"] tolerance = float(arguments["--tolerance"]) # protocol protocol_name = arguments[""] preprocessors = dict() if arguments["overlap"]: if skip_overlap: msg = ( "Option --skip-overlap is not supported " "when evaluating overlapped speech detection." ) sys.exit(msg) preprocessors = {"annotation": to_overlap} protocol = get_protocol(protocol_name, preprocessors=preprocessors) # subset (train, development, or test) subset = arguments["--subset"] if arguments["spotting"]: hypothesis_json = arguments[""] with open(hypothesis_json, mode="r") as fp: hypotheses = json.load(fp) output_prefix = hypothesis_json[:-5] latencies = [float(l) for l in arguments["--latency"]] filters = arguments["--filter"] if filters: from sympy import sympify, lambdify, symbols speech = symbols("speech") filter_funcs = [] filter_funcs = [ lambdify([speech], sympify(expression)) for expression in filters ] filter_func = lambda speech: any(~func(speech) for func in filter_funcs) else: filter_func = None spotting( protocol, subset, latencies, hypotheses, output_prefix, filter_func=filter_func, ) sys.exit(0) hypothesis_rttm = arguments[""] try: hypotheses = load_rttm(hypothesis_rttm) except FileNotFoundError: msg = f"Could not find file {hypothesis_rttm}." sys.exit(msg) except: msg = ( f"Failed to load {hypothesis_rttm}, please check its format " f"(only RTTM files are supported)." ) sys.exit(msg) if arguments["detection"]: detection( protocol, subset, hypotheses, collar=collar, skip_overlap=skip_overlap ) if arguments["overlap"]: detection( protocol, subset, hypotheses, collar=collar, skip_overlap=skip_overlap ) if arguments["segmentation"]: segmentation(protocol, subset, hypotheses, tolerance=tolerance) if arguments["diarization"]: greedy = arguments["--greedy"] diarization( protocol, subset, hypotheses, greedy=greedy, collar=collar, skip_overlap=skip_overlap, ) if arguments["identification"]: identification( protocol, subset, hypotheses, collar=collar, skip_overlap=skip_overlap ) if __name__ == "__main__": main()