"""The ``metric_stats`` module provides an abstract class for storing statistics produced over the course of an experiment and summarizing them. Authors: * Peter Plantinga 2020 * Mirco Ravanelli 2020 * Gaëlle Laperrière 2021 * Sahar Ghannay 2021 """ from typing import Callable, Optional import torch from joblib import Parallel, delayed from speechbrain.dataio.dataio import ( extract_concepts_values, merge_char, split_word, ) from speechbrain.dataio.wer import print_alignments, print_wer_summary from speechbrain.utils.data_utils import undo_padding from speechbrain.utils.edit_distance import ( EDIT_SYMBOLS, _str_equals, wer_details_for_batch, wer_summary, ) class MetricStats: """A default class for storing and summarizing arbitrary metrics. More complex metrics can be created by sub-classing this class. Arguments --------- metric : function The function to use to compute the relevant metric. Should take at least two arguments (predictions and targets) and can optionally take the relative lengths of either or both arguments. Not usually used in sub-classes. n_jobs : int The number of jobs to use for computing the metric. If this is more than one, every sample is processed individually, otherwise the whole batch is passed at once. batch_eval : bool When True it feeds the evaluation metric with the batched input. When False and n_jobs=1, it performs metric evaluation one-by-one in a sequential way. When False and n_jobs>1, the evaluation runs in parallel over the different inputs using joblib. Example ------- >>> from speechbrain.nnet.losses import l1_loss >>> loss_stats = MetricStats(metric=l1_loss) >>> loss_stats.append( ... ids=["utterance1", "utterance2"], ... predictions=torch.tensor([[0.1, 0.2], [0.2, 0.3]]), ... targets=torch.tensor([[0.1, 0.2], [0.1, 0.2]]), ... reduction="batch", ... ) >>> stats = loss_stats.summarize() >>> stats['average'] 0.050... >>> stats['max_score'] 0.100... >>> stats['max_id'] 'utterance2' """ def __init__(self, metric, n_jobs=1, batch_eval=True): self.metric = metric self.n_jobs = n_jobs self.batch_eval = batch_eval self.clear() def clear(self): """Creates empty container for storage, removing existing stats.""" self.scores = [] self.ids = [] self.summary = {} def append(self, ids, *args, **kwargs): """Store a particular set of metric scores. Arguments --------- ids : list List of ids corresponding to utterances. *args : tuple Arguments to pass to the metric function. **kwargs : dict Arguments to pass to the metric function. """ self.ids.extend(ids) # Batch evaluation if self.batch_eval: scores = self.metric(*args, **kwargs).detach() else: if "predict" not in kwargs or "target" not in kwargs: raise ValueError( "Must pass 'predict' and 'target' as kwargs if batch_eval=False" ) if self.n_jobs == 1: # Sequence evaluation (loop over inputs) scores = sequence_evaluation(metric=self.metric, **kwargs) else: # Multiprocess evaluation scores = multiprocess_evaluation( metric=self.metric, n_jobs=self.n_jobs, **kwargs ) self.scores.extend(scores) def summarize(self, field=None): """Summarize the metric scores, returning relevant stats. Arguments --------- field : str If provided, only returns selected statistic. If not, returns all computed statistics. Returns ------- float or dict Returns a float if ``field`` is provided, otherwise returns a dictionary containing all computed stats. """ min_index = torch.argmin(torch.tensor(self.scores)) max_index = torch.argmax(torch.tensor(self.scores)) self.summary = { "average": float(sum(self.scores) / len(self.scores)), "min_score": float(self.scores[min_index]), "min_id": self.ids[min_index], "max_score": float(self.scores[max_index]), "max_id": self.ids[max_index], } if field is not None: return self.summary[field] else: return self.summary def write_stats(self, filestream, verbose=False): """Write all relevant statistics to file. Arguments --------- filestream : file-like object A stream for the stats to be written to. verbose : bool Whether to also print the stats to stdout. """ if not self.summary: self.summarize() message = f"Average score: {self.summary['average']}\n" message += f"Min error: {self.summary['min_score']} " message += f"id: {self.summary['min_id']}\n" message += f"Max error: {self.summary['max_score']} " message += f"id: {self.summary['max_id']}\n" filestream.write(message) if verbose: print(message) def multiprocess_evaluation(metric, predict, target, lengths=None, n_jobs=8): """Runs metric evaluation if parallel over multiple jobs.""" if lengths is not None: lengths = (lengths * predict.size(1)).round().int().cpu() predict = [p[:length].cpu() for p, length in zip(predict, lengths)] target = [t[:length].cpu() for t, length in zip(target, lengths)] while True: try: scores = Parallel(n_jobs=n_jobs, timeout=30)( delayed(metric)(p, t) for p, t in zip(predict, target) ) break except Exception as e: print(e) print("Evaluation timeout...... (will try again)") return scores def sequence_evaluation(metric, predict, target, lengths=None): """Runs metric evaluation sequentially over the inputs.""" if lengths is not None: lengths = (lengths * predict.size(1)).round().int().cpu() predict = [p[:length].cpu() for p, length in zip(predict, lengths)] target = [t[:length].cpu() for t, length in zip(target, lengths)] scores = [] for p, t in zip(predict, target): score = metric(p, t) scores.append(score) return scores class ErrorRateStats(MetricStats): """A class for tracking error rates (e.g., WER, PER). Arguments --------- merge_tokens : bool Whether to merge the successive tokens (used for e.g., creating words out of character tokens). See ``speechbrain.dataio.dataio.merge_char``. split_tokens : bool Whether to split tokens (used for e.g. creating characters out of word tokens). See ``speechbrain.dataio.dataio.split_word``. space_token : str The character to use for boundaries. Used with ``merge_tokens`` this represents character to split on after merge. Used with ``split_tokens`` the sequence is joined with this token in between, and then the whole sequence is split. keep_values : bool Whether to keep the values of the concepts or not. extract_concepts_values : bool Process the predict and target to keep only concepts and values. tag_in : str Start of the concept ('<' for example). tag_out : str End of the concept ('>' for example). equality_comparator : Callable[[str, str], bool] The function used to check whether two words are equal. Example ------- >>> cer_stats = ErrorRateStats() >>> i2l = {0: 'a', 1: 'b'} >>> cer_stats.append( ... ids=['utterance1'], ... predict=torch.tensor([[0, 1, 1]]), ... target=torch.tensor([[0, 1, 0]]), ... target_len=torch.ones(1), ... ind2lab=lambda batch: [[i2l[int(x)] for x in seq] for seq in batch], ... ) >>> stats = cer_stats.summarize() >>> stats['WER'] 33.33... >>> stats['insertions'] 0 >>> stats['deletions'] 0 >>> stats['substitutions'] 1 """ def __init__( self, merge_tokens=False, split_tokens=False, space_token="_", keep_values=True, extract_concepts_values=False, tag_in="", tag_out="", equality_comparator: Callable[[str, str], bool] = _str_equals, ): self.clear() self.merge_tokens = merge_tokens self.split_tokens = split_tokens self.space_token = space_token self.extract_concepts_values = extract_concepts_values self.keep_values = keep_values self.tag_in = tag_in self.tag_out = tag_out self.equality_comparator = equality_comparator def append( self, ids, predict, target, predict_len=None, target_len=None, ind2lab=None, ): """Add stats to the relevant containers. * See MetricStats.append() Arguments --------- ids : list List of ids corresponding to utterances. predict : torch.tensor A predicted output, for comparison with the target output target : torch.tensor The correct reference output, for comparison with the prediction. predict_len : torch.tensor The predictions relative lengths, used to undo padding if there is padding present in the predictions. target_len : torch.tensor The target outputs' relative lengths, used to undo padding if there is padding present in the target. ind2lab : callable Callable that maps from indices to labels, operating on batches, for writing alignments. """ self.ids.extend(ids) if predict_len is not None: predict = undo_padding(predict, predict_len) if target_len is not None: target = undo_padding(target, target_len) if ind2lab is not None: predict = ind2lab(predict) target = ind2lab(target) if self.merge_tokens: predict = merge_char(predict, space=self.space_token) target = merge_char(target, space=self.space_token) if self.split_tokens: predict = split_word(predict, space=self.space_token) target = split_word(target, space=self.space_token) if self.extract_concepts_values: predict = extract_concepts_values( predict, self.keep_values, self.tag_in, self.tag_out, space=self.space_token, ) target = extract_concepts_values( target, self.keep_values, self.tag_in, self.tag_out, space=self.space_token, ) scores = wer_details_for_batch( ids, target, predict, compute_alignments=True, equality_comparator=self.equality_comparator, ) self.scores.extend(scores) def summarize(self, field=None): """Summarize the error_rate and return relevant statistics. * See MetricStats.summarize() """ self.summary = wer_summary(self.scores) # Add additional, more generic key self.summary["error_rate"] = self.summary["WER"] if field is not None: return self.summary[field] else: return self.summary def write_stats(self, filestream): """Write all relevant info (e.g., error rate alignments) to file. * See MetricStats.write_stats() """ if not self.summary: self.summarize() print_wer_summary(self.summary, filestream) print_alignments(self.scores, filestream) class WeightedErrorRateStats(MetricStats): """Metric that reweighs the WER from :class:`~ErrorRateStats` with any chosen method. This does not edit the sequence of found edits (insertion/deletion/substitution) but multiplies their impact on the metric by a value between 0 and 1 as returned by the cost function. Arguments --------- base_stats : ErrorRateStats The base WER calculator to use. cost_function : Callable[[str, Optional[str], Optional[str]], float] Cost function of signature `fn(edit_symbol, a, b) -> float`, where the returned value, between 0 and 1, is the weight that should be assigned to a particular edit in the weighted WER calculation. In the case of insertions and deletions, either of `a` or `b` may be `None`. In the case of substitutions, `a` and `b` will never be `None`. weight_name : str Prefix to be prepended to each metric name (e.g. `xxx_wer`) """ def __init__( self, base_stats: ErrorRateStats, cost_function: Callable[[str, Optional[str], Optional[str]], float], weight_name: str = "weighted", ): self.clear() self.base_stats = base_stats self.cost_function = cost_function self.weight_name = weight_name def append(self, *args, **kwargs): """Append function, which should **NOT** be used for the weighted error rate stats. Please append to the specified `base_stats` instead. `WeightedErrorRateStats` reuses the scores from the base :class:`~ErrorRateStats` class. Arguments --------- *args : tuple Ignored. **kwargs : dict Ignored. """ raise ValueError( "Cannot append to a WeightedErrorRateStats. " "You should only append to the base ErrorRateStats." ) def summarize(self, field=None): """Returns a dict containing some detailed WER statistics after weighting every edit with a weight determined by `cost_function` (returning `0.0` for no error, `1.0` for the default error behavior, and anything in between). Does not require :meth:`~ErrorRateStats.summarize` to have been called. Full set of fields, **each of which are prepended with `<weight_name_specified_at_init>_`**: - `wer`: Weighted WER (ratio `*100`) - `insertions`: Weighted insertions - `substitutions`: Weighted substitutions - `deletions`: Weighted deletions - `num_edits`: Sum of weighted insertions/substitutions/deletions Additionally, a `scores` list is populated by this function for each pair of sentences. Each entry of that list is a dict, with the fields: - `key`: the ID of the utterance. - `WER`, `insertions`, `substitutions`, `deletions`, `num_edits` with the same semantics as described above, but at sentence level rather than global. Arguments --------- field : str, optional The field to return, if you are only interested in one of them. If specified, a single `float` is returned, otherwise, a dict is. Returns ------- dict from str to float, if `field is None` A dictionary of the fields documented above. float, if `field is not None` The single field selected by `field`. """ weighted_insertions = 0.0 weighted_substitutions = 0.0 weighted_deletions = 0.0 total = 0.0 for i, utterance in enumerate(self.base_stats.scores): utt_weighted_insertions = 0.0 utt_weighted_substitutions = 0.0 utt_weighted_deletions = 0.0 utt_total = 0.0 for edit_symbol, a_idx, b_idx in utterance["alignment"]: a = ( utterance["ref_tokens"][a_idx] if a_idx is not None else None ) b = ( utterance["hyp_tokens"][b_idx] if b_idx is not None else None ) if edit_symbol != EDIT_SYMBOLS["eq"]: pair_score = self.cost_function(edit_symbol, a, b) if edit_symbol == EDIT_SYMBOLS["ins"]: utt_weighted_insertions += pair_score elif edit_symbol == EDIT_SYMBOLS["del"]: utt_weighted_deletions += pair_score elif edit_symbol == EDIT_SYMBOLS["sub"]: utt_weighted_substitutions += pair_score utt_total += 1.0 utt_weighted_edits = ( utt_weighted_insertions + utt_weighted_substitutions + utt_weighted_deletions ) utt_weighted_wer_ratio = utt_weighted_edits / utt_total self.scores.append( { "key": self.base_stats.ids[i], "WER": utt_weighted_wer_ratio * 100.0, "insertions": utt_weighted_insertions, "substitutions": utt_weighted_substitutions, "deletions": utt_weighted_deletions, "num_edits": utt_weighted_edits, } ) weighted_insertions += utt_weighted_insertions weighted_substitutions += utt_weighted_substitutions weighted_deletions += utt_weighted_deletions total += utt_total weighted_edits = ( weighted_insertions + weighted_substitutions + weighted_deletions ) weighted_wer_ratio = weighted_edits / total self.summary = { f"{self.weight_name}_wer": weighted_wer_ratio * 100.0, f"{self.weight_name}_insertions": weighted_insertions, f"{self.weight_name}_substitutions": weighted_substitutions, f"{self.weight_name}_deletions": weighted_deletions, f"{self.weight_name}_num_edits": weighted_edits, } if field is not None: return self.summary[field] else: return self.summary def write_stats(self, filestream): """Write all relevant info to file; here, only the weighted info as returned by `summarize`. See :meth:`~ErrorRateStats.write_stats`. """ if not self.summary: self.summarize() print(f"Weighted WER metrics ({self.weight_name}):", file=filestream) for k, v in self.summary.items(): print(f"{k}: {v}", file=filestream) class EmbeddingErrorRateSimilarity: """Implements the similarity function from the EmbER metric as defined by https://www.isca-archive.org/interspeech_2022/roux22_interspeech.pdf This metric involves a dictionary to map a token to a single word embedding. Substitutions in the WER get weighted down when the embeddings are similar enough. The goal is to reduce the impact of substitution errors with small semantic impact. Only substitution errors get weighted. This is done by computing the cosine similarity between the two embeddings, then weighing the substitution with `low_similarity_weight` if `similarity >= threshold` or with `high_similarity_weight` otherwise (e.g. a substitution with high similarity could be weighted down to matter 10% as much as a substitution with low similarity). .. note :: The cited paper recommended `(1.0, 0.1, 0.4)` as defaults for fastTexst French embeddings, chosen empirically. When using different embeddings, you might want to test other values; thus we don't provide defaults. Arguments --------- embedding_function : Callable[[str], Optional[torch.Tensor]] Function that returns an embedding (as a :class:`torch.Tensor`) from a word. If no corresponding embedding could be found for the word, should return `None`. In that case, `low_similarity_weight` will be chosen. low_similarity_weight : float Weight applied to the substitution if `cosine_similarity < threshold`. high_similarity_weight : float Weight applied to the substitution if `cosine_similarity >= threshold`. threshold : float Cosine similarity threshold used to select by how much a substitution error should be weighed for this word. """ def __init__( self, embedding_function: Callable[[str], Optional[torch.Tensor]], low_similarity_weight: float, high_similarity_weight: float, threshold: float, ): self.embedding_function = embedding_function self.low_similarity_weight = low_similarity_weight self.high_similarity_weight = high_similarity_weight self.threshold = threshold def __call__( self, edit_symbol: str, a: Optional[str], b: Optional[str] ) -> float: """Returns the weight that should be associated with a specific edit in the WER calculation. Compatible candidate for the cost function of :class:`~WeightedErrorRateStats` so an instance of this class can be passed as a `cost_function`. Arguments --------- edit_symbol: str Edit symbol as assigned by the WER functions, see `EDIT_SYMBOLS`. a: str, optional First word to compare (if present) b: str, optional Second word to compare (if present) Returns ------- float Weight to assign to the edit. For actual edits, either `low_similarity_weight` or `high_similarity_weight` depending on the embedding distance and threshold. """ if edit_symbol in (EDIT_SYMBOLS["ins"], EDIT_SYMBOLS["del"]): return 1.0 if edit_symbol == EDIT_SYMBOLS["sub"]: if a is None or a == "": return self.low_similarity_weight if b is None or b == "": return self.low_similarity_weight a_emb = self.embedding_function(a) if a_emb is None: return self.low_similarity_weight b_emb = self.embedding_function(b) if b_emb is None: return self.low_similarity_weight similarity = torch.nn.functional.cosine_similarity( a_emb, b_emb, dim=0 ).item() if similarity >= self.threshold: return self.high_similarity_weight return self.low_similarity_weight # eq return 0.0 class BinaryMetricStats(MetricStats): """Tracks binary metrics, such as precision, recall, F1, EER, etc.""" def __init__(self, positive_label=1): self.clear() self.positive_label = positive_label def clear(self): """Clears the stored metrics.""" self.ids = [] self.scores = [] self.labels = [] self.summary = {} def append(self, ids, scores, labels): """Appends scores and labels to internal lists. Does not compute metrics until time of summary, since automatic thresholds (e.g., EER) need full set of scores. Arguments --------- ids : list The string ids for the samples. scores : list The scores corresponding to the ids. labels : list The labels corresponding to the ids. """ self.ids.extend(ids) self.scores.extend(scores.detach()) self.labels.extend(labels.detach()) def summarize( self, field=None, threshold=None, max_samples=None, beta=1, eps=1e-8 ): """Compute statistics using a full set of scores. Full set of fields: - TP - True Positive - TN - True Negative - FP - False Positive - FN - False Negative - FAR - False Acceptance Rate - FRR - False Rejection Rate - DER - Detection Error Rate (EER if no threshold passed) - threshold - threshold (EER threshold if no threshold passed) - precision - Precision (positive predictive value) - recall - Recall (sensitivity) - F-score - Balance of precision and recall (equal if beta=1) - MCC - Matthews Correlation Coefficient Arguments --------- field : str A key for selecting a single statistic. If not provided, a dict with all statistics is returned. threshold : float If no threshold is provided, equal error rate is used. max_samples: float How many samples to keep for positive/negative scores. If no max_samples is provided, all scores are kept. Only effective when threshold is None. beta : float How much to weight precision vs recall in F-score. Default of 1. is equal weight, while higher values weight recall higher, and lower values weight precision higher. eps : float A small value to avoid dividing by zero. Returns ------- summary if field is specified, only returns the score for that field. if field is None, returns the full set of fields. """ if isinstance(self.scores, list): self.scores = torch.stack(self.scores) self.labels = torch.stack(self.labels) if threshold is None: positive_scores = self.scores[ (self.labels == self.positive_label).nonzero(as_tuple=True) ] negative_scores = self.scores[ (self.labels != self.positive_label).nonzero(as_tuple=True) ] if max_samples is not None: if len(positive_scores) > max_samples: positive_scores, _ = torch.sort(positive_scores) positive_scores = positive_scores[ [ i for i in range( 0, len(positive_scores), int(len(positive_scores) / max_samples), ) ] ] if len(negative_scores) > max_samples: negative_scores, _ = torch.sort(negative_scores) negative_scores = negative_scores[ [ i for i in range( 0, len(negative_scores), int(len(negative_scores) / max_samples), ) ] ] eer, threshold = EER(positive_scores, negative_scores) pred = (self.scores > threshold).float() true = self.labels TP = self.summary["TP"] = float(pred.mul(true).sum()) TN = self.summary["TN"] = float((1.0 - pred).mul(1.0 - true).sum()) FP = self.summary["FP"] = float(pred.mul(1.0 - true).sum()) FN = self.summary["FN"] = float((1.0 - pred).mul(true).sum()) self.summary["FAR"] = FP / (FP + TN + eps) self.summary["FRR"] = FN / (TP + FN + eps) self.summary["DER"] = (FP + FN) / (TP + TN + eps) self.summary["threshold"] = threshold self.summary["precision"] = TP / (TP + FP + eps) self.summary["recall"] = TP / (TP + FN + eps) self.summary["F-score"] = ( (1.0 + beta**2.0) * TP / ((1.0 + beta**2.0) * TP + beta**2.0 * FN + FP) ) self.summary["MCC"] = (TP * TN - FP * FN) / ( (TP + FP) * (TP + FN) * (TN + FP) * (TN + FN) + eps ) ** 0.5 if field is not None: return self.summary[field] else: return self.summary def EER(positive_scores, negative_scores): """Computes the EER (and its threshold). Arguments --------- positive_scores : torch.tensor The scores from entries of the same class. negative_scores : torch.tensor The scores from entries of different classes. Returns ------- EER : float The EER score. threshold : float The corresponding threshold for the EER score. Example ------- >>> positive_scores = torch.tensor([0.6, 0.7, 0.8, 0.5]) >>> negative_scores = torch.tensor([0.4, 0.3, 0.2, 0.1]) >>> val_eer, threshold = EER(positive_scores, negative_scores) >>> val_eer 0.0 """ # Computing candidate thresholds thresholds, _ = torch.sort(torch.cat([positive_scores, negative_scores])) thresholds = torch.unique(thresholds) # Adding intermediate thresholds intermediate_thresholds = (thresholds[0:-1] + thresholds[1:]) / 2 thresholds, _ = torch.sort(torch.cat([thresholds, intermediate_thresholds])) # Variable to store the min FRR, min FAR and their corresponding index min_index = 0 final_FRR = 0 final_FAR = 0 for i, cur_thresh in enumerate(thresholds): pos_scores_threshold = positive_scores <= cur_thresh FRR = (pos_scores_threshold.sum(0)).float() / positive_scores.shape[0] del pos_scores_threshold neg_scores_threshold = negative_scores > cur_thresh FAR = (neg_scores_threshold.sum(0)).float() / negative_scores.shape[0] del neg_scores_threshold # Finding the threshold for EER if (FAR - FRR).abs().item() < abs(final_FAR - final_FRR) or i == 0: min_index = i final_FRR = FRR.item() final_FAR = FAR.item() # It is possible that eer != fpr != fnr. We return (FAR + FRR) / 2 as EER. EER = (final_FAR + final_FRR) / 2 return float(EER), float(thresholds[min_index]) def minDCF( positive_scores, negative_scores, c_miss=1.0, c_fa=1.0, p_target=0.01 ): """Computes the minDCF metric normally used to evaluate speaker verification systems. The min_DCF is the minimum of the following C_det function computed within the defined threshold range: C_det = c_miss * p_miss * p_target + c_fa * p_fa * (1 -p_target) where p_miss is the missing probability and p_fa is the probability of having a false alarm. Arguments --------- positive_scores : torch.tensor The scores from entries of the same class. negative_scores : torch.tensor The scores from entries of different classes. c_miss : float Cost assigned to a missing error (default 1.0). c_fa : float Cost assigned to a false alarm (default 1.0). p_target: float Prior probability of having a target (default 0.01). Returns ------- minDCF : float The minDCF score. threshold : float The corresponding threshold for the minDCF score. Example ------- >>> positive_scores = torch.tensor([0.6, 0.7, 0.8, 0.5]) >>> negative_scores = torch.tensor([0.4, 0.3, 0.2, 0.1]) >>> val_minDCF, threshold = minDCF(positive_scores, negative_scores) >>> val_minDCF 0.0 """ # Computing candidate thresholds thresholds, _ = torch.sort(torch.cat([positive_scores, negative_scores])) thresholds = torch.unique(thresholds) # Adding intermediate thresholds intermediate_thresholds = (thresholds[0:-1] + thresholds[1:]) / 2 thresholds, _ = torch.sort(torch.cat([thresholds, intermediate_thresholds])) # Computing False Rejection Rate (miss detection) positive_scores = torch.cat( len(thresholds) * [positive_scores.unsqueeze(0)] ) pos_scores_threshold = positive_scores.transpose(0, 1) <= thresholds p_miss = (pos_scores_threshold.sum(0)).float() / positive_scores.shape[1] del positive_scores del pos_scores_threshold # Computing False Acceptance Rate (false alarm) negative_scores = torch.cat( len(thresholds) * [negative_scores.unsqueeze(0)] ) neg_scores_threshold = negative_scores.transpose(0, 1) > thresholds p_fa = (neg_scores_threshold.sum(0)).float() / negative_scores.shape[1] del negative_scores del neg_scores_threshold c_det = c_miss * p_miss * p_target + c_fa * p_fa * (1 - p_target) c_min, min_index = torch.min(c_det, dim=0) return float(c_min), float(thresholds[min_index]) class ClassificationStats(MetricStats): """Computes statistics pertaining to multi-label classification tasks, as well as tasks that can be loosely interpreted as such for the purpose of evaluations. Example ------- >>> import sys >>> from speechbrain.utils.metric_stats import ClassificationStats >>> cs = ClassificationStats() >>> cs.append( ... ids=["ITEM1", "ITEM2", "ITEM3", "ITEM4"], ... predictions=[ ... "M EY K AH", ... "T EY K", ... "B AE D", ... "M EY K", ... ], ... targets=[ ... "M EY K", ... "T EY K", ... "B AE D", ... "M EY K", ... ], ... categories=[ ... "make", ... "take", ... "bad", ... "make" ... ] ... ) >>> cs.write_stats(sys.stdout) Overall Accuracy: 75% <BLANKLINE> Class-Wise Accuracy ------------------- bad -> B AE D : 1 / 1 (100.00%) make -> M EY K: 1 / 2 (50.00%) take -> T EY K: 1 / 1 (100.00%) <BLANKLINE> Confusion --------- Target: bad -> B AE D -> B AE D : 1 / 1 (100.00%) Target: make -> M EY K -> M EY K : 1 / 2 (50.00%) -> M EY K AH: 1 / 2 (50.00%) Target: take -> T EY K -> T EY K : 1 / 1 (100.00%) >>> summary = cs.summarize() >>> summary['accuracy'] 0.75 >>> summary['classwise_stats'][('bad', 'B AE D')] {'total': 1.0, 'correct': 1.0, 'accuracy': 1.0} >>> summary['classwise_stats'][('make', 'M EY K')] {'total': 2.0, 'correct': 1.0, 'accuracy': 0.5} >>> summary['keys'] [('bad', 'B AE D'), ('make', 'M EY K'), ('take', 'T EY K')] >>> summary['predictions'] ['B AE D', 'M EY K', 'M EY K AH', 'T EY K'] >>> summary['classwise_total'] {('bad', 'B AE D'): 1.0, ('make', 'M EY K'): 2.0, ('take', 'T EY K'): 1.0} >>> summary['classwise_correct'] {('bad', 'B AE D'): 1.0, ('make', 'M EY K'): 1.0, ('take', 'T EY K'): 1.0} >>> summary['classwise_accuracy'] {('bad', 'B AE D'): 1.0, ('make', 'M EY K'): 0.5, ('take', 'T EY K'): 1.0} """ def __init__(self): super() self.clear() self.summary = None def append(self, ids, predictions, targets, categories=None): """ Appends inputs, predictions and targets to internal lists Arguments --------- ids: list the string IDs for the samples predictions: list the model's predictions (human-interpretable, preferably strings) targets: list the ground truths (human-interpretable, preferably strings) categories: list an additional way to classify training samples. If available, the categories will be combined with targets """ self.ids.extend(ids) self.predictions.extend(predictions) self.targets.extend(targets) if categories is not None: self.categories.extend(categories) def summarize(self, field=None): """Summarize the classification metric scores The following statistics are computed: accuracy: the overall accuracy (# correct / # total) confusion_matrix: a dictionary of type {(target, prediction): num_entries} representing the confusion matrix classwise_stats: computes the total number of samples, the number of correct classifications and accuracy for each class keys: all available class keys, which can be either target classes or (category, target) tuples predictions: all available predictions all predictions the model has made Arguments --------- field : str If provided, only returns selected statistic. If not, returns all computed statistics. Returns ------- float or dict Returns a float if ``field`` is provided, otherwise returns a dictionary containing all computed stats. """ self._build_lookups() confusion_matrix = self._compute_confusion_matrix() self.summary = { "accuracy": self._compute_accuracy(), "confusion_matrix": confusion_matrix, "classwise_stats": self._compute_classwise_stats(confusion_matrix), "keys": self._available_keys, "predictions": self._available_predictions, } for stat in ["total", "correct", "accuracy"]: self.summary[f"classwise_{stat}"] = { key: key_stats[stat] for key, key_stats in self.summary["classwise_stats"].items() } if field is not None: return self.summary[field] else: return self.summary def _compute_accuracy(self): return sum( prediction == target for prediction, target in zip(self.predictions, self.targets) ) / len(self.ids) def _build_lookups(self): self._available_keys = self._get_keys() self._available_predictions = list( sorted(set(prediction for prediction in self.predictions)) ) self._keys_lookup = self._index_lookup(self._available_keys) self._predictions_lookup = self._index_lookup( self._available_predictions ) def _compute_confusion_matrix(self): confusion_matrix = torch.zeros( len(self._available_keys), len(self._available_predictions) ) for key, prediction in self._get_confusion_entries(): key_idx = self._keys_lookup[key] prediction_idx = self._predictions_lookup[prediction] confusion_matrix[key_idx, prediction_idx] += 1 return confusion_matrix def _compute_classwise_stats(self, confusion_matrix): total = confusion_matrix.sum(dim=-1) # This can be used with "classes" that are not # statically determined; for example, they could # be constructed from seq2seq predictions. As a # result, one cannot use the diagonal key_targets = ( self._available_keys if not self.categories else [target for _, target in self._available_keys] ) correct = torch.tensor( [ ( confusion_matrix[idx, self._predictions_lookup[target]] if target in self._predictions_lookup else 0 ) for idx, target in enumerate(key_targets) ] ) accuracy = correct / total return { key: { "total": item_total.item(), "correct": item_correct.item(), "accuracy": item_accuracy.item(), } for key, item_total, item_correct, item_accuracy in zip( self._available_keys, total, correct, accuracy ) } def _get_keys(self): if self.categories: keys = zip(self.categories, self.targets) else: keys = self.targets return list(sorted(set(keys))) def _get_confusion_entries(self): if self.categories: result = ( ((category, target), prediction) for category, target, prediction in zip( self.categories, self.targets, self.predictions ) ) else: result = zip(self.targets, self.predictions) result = list(result) return result def _index_lookup(self, items): return {item: idx for idx, item in enumerate(items)} def clear(self): """Clears the collected statistics""" self.ids = [] self.predictions = [] self.targets = [] self.categories = [] def write_stats(self, filestream): """Outputs the stats to the specified filestream in a human-readable format Arguments --------- filestream: file a file-like object """ if self.summary is None: self.summarize() print( f"Overall Accuracy: {self.summary['accuracy']:.0%}", file=filestream ) print(file=filestream) self._write_classwise_stats(filestream) print(file=filestream) self._write_confusion(filestream) def _write_classwise_stats(self, filestream): self._write_header("Class-Wise Accuracy", filestream=filestream) key_labels = { key: self._format_key_label(key) for key in self._available_keys } longest_key_label = max(len(label) for label in key_labels.values()) for key in self._available_keys: stats = self.summary["classwise_stats"][key] padded_label = self._pad_to_length( self._format_key_label(key), longest_key_label ) print( f"{padded_label}: {int(stats['correct'])} / {int(stats['total'])} ({stats['accuracy']:.2%})", file=filestream, ) def _write_confusion(self, filestream): self._write_header("Confusion", filestream=filestream) longest_prediction = max( len(prediction) for prediction in self._available_predictions ) confusion_matrix = self.summary["confusion_matrix"].int() totals = confusion_matrix.sum(dim=-1) for key, key_predictions, total in zip( self._available_keys, confusion_matrix, totals ): target_label = self._format_key_label(key) print(f"Target: {target_label}", file=filestream) (indexes,) = torch.where(key_predictions > 0) total = total.item() for index in indexes: count = key_predictions[index].item() prediction = self._available_predictions[index] padded_label = self._pad_to_length( prediction, longest_prediction ) print( f" -> {padded_label}: {count} / {total} ({count / total:.2%})", file=filestream, ) def _write_header(self, header, filestream): print(header, file=filestream) print("-" * len(header), file=filestream) def _pad_to_length(self, label, length): padding = max(0, length - len(label)) return label + (" " * padding) def _format_key_label(self, key): if self.categories: category, target = key label = f"{category} -> {target}" else: label = key return label class MultiMetricStats: """A wrapper that evaluates multiple metrics simultaneously Arguments --------- metric : function The function to use to compute the relevant metrics. Should take at least two arguments (predictions and targets) and can optionally take the relative lengths of either or both arguments. The function should return a dict or a namedtuple n_jobs : int The number of jobs to use for computing the metric. If this is more than one, every sample is processed individually, otherwise the whole batch is passed at once. batch_eval : bool When True it feeds the evaluation metric with the batched input. When False and n_jobs=1, it performs metric evaluation one-by-one in a sequential way. When False and n_jobs>1, the evaluation runs in parallel over the different inputs using joblib. Example ------- >>> def metric(a, b): ... return { ... "sum": a + b, ... "diff": a - b, ... "sum_sq": a**2 + b**2 ... } >>> multi_metric = MultiMetricStats(metric, batch_eval=True) >>> multi_metric.append([1, 2], a=torch.tensor([2.0, 1.0]), b=torch.tensor([1.0, 2.0])) >>> multi_metric.append([3, 4], a=torch.tensor([4.0, 5.0]), b=torch.tensor([0.0, 1.0])) >>> multi_metric.append([5, 6], a=torch.tensor([2.0, 4.0]), b=torch.tensor([4.0, 2.0])) >>> multi_metric.append([7, 8], a=torch.tensor([2.0, 4.0]), b=torch.tensor([4.0, 2.0])) >>> multi_metric.summarize() #doctest: +NORMALIZE_WHITESPACE {'sum': {'average': 5.0, 'min_score': 3.0, 'min_id': 1, 'max_score': 6.0, 'max_id': 4}, 'diff': {'average': 1.0, 'min_score': -2.0, 'min_id': 5, 'max_score': 4.0, 'max_id': 3}, 'sum_sq': {'average': 16.5, 'min_score': 5.0, 'min_id': 1, 'max_score': 26.0, 'max_id': 4}} >>> multi_metric.summarize(flat=True) #doctest: +NORMALIZE_WHITESPACE {'sum_average': 5.0, 'sum_min_score': 3.0, 'sum_min_id': 1, 'sum_max_score': 6.0, 'sum_max_id': 4, 'diff_average': 1.0, 'diff_min_score': -2.0, 'diff_min_id': 5, 'diff_max_score': 4.0, 'diff_max_id': 3, 'sum_sq_average': 16.5, 'sum_sq_min_score': 5.0, 'sum_sq_min_id': 1, 'sum_sq_max_score': 26.0, 'sum_sq_max_id': 4} """ def __init__(self, metric, n_jobs=1, batch_eval=False): self.metric = _dictify(metric) self.n_jobs = n_jobs self.batch_eval = batch_eval self.ids = [] self.metrics = {} def append(self, ids, *args, **kwargs): """Store a particular set of metric scores. Arguments --------- ids : list List of ids corresponding to utterances. *args : tuple Arguments to pass to the metric function. **kwargs : dict Arguments to pass to the metric function. """ self.ids.extend(ids) # Batch evaluation if self.batch_eval: scores = self.eval_simple(*args, **kwargs) else: if "predict" not in kwargs or "target" not in kwargs: raise ValueError( "Must pass 'predict' and 'target' as kwargs if batch_eval=False" ) if self.n_jobs == 1: # Sequence evaluation (loop over inputs) scores_raw = sequence_evaluation(self.metric, **kwargs) else: # Multiprocess evaluation scores_raw = multiprocess_evaluation( metric=self.metric, n_jobs=self.n_jobs, **kwargs ) keys = scores_raw[0].keys() scores = { key: torch.tensor([score[key] for score in scores_raw]) for key in keys } for key, metric_scores in scores.items(): if key not in self.metrics: self.metrics[key] = MetricStats(lambda x: x, batch_eval=True) self.metrics[key].append(ids, metric_scores) def eval_simple(self, *args, **kwargs): """Evaluates the metric in a simple, sequential manner""" scores = self.metric(*args, **kwargs) return {key: score.detach() for key, score in scores.items()} def summarize(self, field=None, flat=False): """Summarize the metric scores, returning relevant stats. Arguments --------- field : str If provided, only returns selected statistic. If not, returns all computed statistics. flat : bool whether to flatten the dictionary Returns ------- dict Returns a dictionary of all computed stats """ result = { key: metric.summarize(field) for key, metric in self.metrics.items() } if flat: result = { f"{key}_{field}": value for key, fields in result.items() for field, value in fields.items() } return result def _dictify(f): """A wrapper that converts functions returning namedtuples to functions returning dicts while leaving functions returning dicts intact Arguments --------- f : callable a function Returns ------- result : callable a wrapped function """ has_asdict = None def wrapper(*args, **kwargs): """The wrapper function""" nonlocal has_asdict result = f(*args, **kwargs) if has_asdict is None: has_asdict = hasattr(result, "_asdict") return result._asdict() if has_asdict else result return wrapper
Memory