aac-metrics 0.4.4

Metrics for evaluating Automated Audio Captioning systems, designed for PyTorch.

Audio Captioning metrics (aac-metrics)

Metrics for evaluating Automated Audio Captioning systems, designed for PyTorch.

Why using this package?

• Easy installation and download
• Same results than caption-evaluation-tools and fense repositories
• Provides the following metrics:
  • BLEU [1]
  • ROUGE-L [2]
  • METEOR [3]
  • CIDEr-D [4]
  • SPICE [5]
  • SPIDEr [6]
  • SPIDEr-max [7]
  • SBERT-sim [8]
  • Fluency Error [8]
  • FENSE [8]
  • SPIDEr-FL [9]

Installation

Install the pip package:

pip install aac-metrics

If you want to check if the package has been installed and the version, you can use this command:

aac-metrics-info

Download the external code and models needed for METEOR, SPICE, SPIDEr, SPIDEr-max, PTBTokenizer, SBERTSim, FluencyError, FENSE and SPIDEr-FL:

aac-metrics-download

Notes:

• The external code for SPICE, METEOR and PTBTokenizer is stored in $HOME/.cache/aac-metrics.
• The weights of the FENSE fluency error detector and the the SBERT model are respectively stored by default in $HOME/.cache/torch/hub/fense_data and $HOME/.cache/torch/sentence_transformers.

Usage

Evaluate default metrics

The full evaluation pipeline to compute AAC metrics can be done with aac_metrics.evaluate function.

from aac_metrics import evaluate

candidates: list[str] = ["a man is speaking"]
mult_references: list[list[str]] = [["a man speaks.", "someone speaks.", "a man is speaking while a bird is chirping in the background"]]

corpus_scores, _ = evaluate(candidates, mult_references)
print(corpus_scores)
# dict containing the score of each metric: "bleu_1", "bleu_2", "bleu_3", "bleu_4", "rouge_l", "meteor", "cider_d", "spice", "spider"
# {"bleu_1": tensor(0.7), "bleu_2": ..., ...}

Evaluate DCASE2023 metrics

To compute metrics for the DCASE2023 challenge, just set the argument metrics="dcase2023" in evaluate function call.

corpus_scores, _ = evaluate(candidates, mult_references, metrics="dcase2023")
print(corpus_scores)
# dict containing the score of each metric: "meteor", "cider_d", "spice", "spider", "spider_fl", "fluerr"

Evaluate a specific metric

Evaluate a specific metric can be done using the aac_metrics.functional.<metric_name>.<metric_name> function or the aac_metrics.classes.<metric_name>.<metric_name> class. Unlike evaluate, the tokenization with PTBTokenizer is not done with these functions, but you can do it manually with preprocess_mono_sents and preprocess_mult_sents functions.

from aac_metrics.functional import cider_d
from aac_metrics.utils.tokenization import preprocess_mono_sents, preprocess_mult_sents

candidates: list[str] = ["a man is speaking"]
mult_references: list[list[str]] = [["a man speaks.", "someone speaks.", "a man is speaking while a bird is chirping in the background"]]

candidates = preprocess_mono_sents(candidates)
mult_references = preprocess_mult_sents(mult_references)

corpus_scores, sents_scores = cider_d(candidates, mult_references)
print(corpus_scores)
# {"cider_d": tensor(0.1)}
print(sents_scores)
# {"cider_d": tensor([0.9, ...])}

Each metrics also exists as a python class version, like aac_metrics.classes.cider_d.CIDErD.

Metrics

Legacy metrics

Metric	Python Class	Origin	Range	Short description
BLEU [1]	BLEU	machine translation	[0, 1]	Precision of n-grams
ROUGE-L [2]	ROUGEL	machine translation	[0, 1]	FScore of the longest common subsequence
METEOR [3]	METEOR	machine translation	[0, 1]	Cosine-similarity of frequencies with synonyms matching
CIDEr-D [4]	CIDErD	image captioning	[0, 10]	Cosine-similarity of TF-IDF computed on n-grams
SPICE [5]	SPICE	image captioning	[0, 1]	FScore of semantic graph
SPIDEr [6]	SPIDEr	image captioning	[0, 5.5]	Mean of CIDEr-D and SPICE

AAC-specific metrics

Metric name	Python Class	Origin	Range	Short description
SPIDEr-max [7]	SPIDErMax	audio captioning	[0, 5.5]	Max of SPIDEr scores for multiples candidates
SBERT-sim [8]	SBERTSim	audio captioning	[-1, 1]	Cosine-similarity of Sentence-BERT embeddings
Fluency Error [8]	FluErr	audio captioning	[0, 1]	Use a pretrained model to detect fluency errors in sentences
FENSE [8]	FENSE	audio captioning	[-1, 1]	Combines SBERT-sim and Fluency Error
SPIDEr-FL [9]	SPIDErFL	audio captioning	[0, 5.5]	Combines SPIDEr and Fluency Error

Requirements

This package has been developped for Ubuntu 20.04, and it is expected to work on most Linux distributions.

Python packages

The pip requirements are automatically installed when using pip install on this repository.

torch >= 1.10.1
numpy >= 1.21.2
pyyaml >= 6.0
tqdm >= 4.64.0
sentence-transformers >= 2.2.2

External requirements

• java >= 1.8 and <= 1.13 is required to compute METEOR, SPICE and use the PTBTokenizer. Most of these functions can specify a java executable path with java_path argument.

• unzip command to extract SPICE zipped files.

Additional notes

CIDEr or CIDEr-D ?

The CIDEr metric differs from CIDEr-D because it applies a stemmer to each word before computing the n-grams of the sentences. In AAC, only the CIDEr-D is reported and used for SPIDEr in caption-evaluation-tools, but some papers called it "CIDEr".

Does metrics work on multi-GPU ?

No. Most of these metrics use numpy or external java programs to run, which prevents multi-GPU testing for now.

SPIDEr-max metric

SPIDEr-max [7] is a metric based on SPIDEr that takes into account multiple candidates for the same audio. It computes the maximum of the SPIDEr scores for each candidate to balance the high sensitivity to the frequency of the words generated by the model. For more detail, please see the documentation about SPIDEr-max.

References

BLEU

[1] K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu, “BLEU: a method for automatic evaluation of machine translation,” in Proceed- ings of the 40th Annual Meeting on Association for Computational Linguistics - ACL ’02. Philadelphia, Pennsylvania: Association for Computational Linguistics, 2001, p. 311. [Online]. Available: http://portal.acm.org/citation.cfm?doid=1073083.1073135

ROUGE-L

[2] C.-Y. Lin, “ROUGE: A package for automatic evaluation of summaries,” in Text Summarization Branches Out. Barcelona, Spain: Association for Computational Linguistics, Jul. 2004, pp. 74–81. [Online]. Available: https://aclanthology.org/W04-1013

METEOR

[3] M. Denkowski and A. Lavie, “Meteor Universal: Language Specific Translation Evaluation for Any Target Language,” in Proceedings of the Ninth Workshop on Statistical Machine Translation. Baltimore, Maryland, USA: Association for Computational Linguistics, 2014, pp. 376–380. [Online]. Available: http://aclweb.org/anthology/W14-3348

CIDEr

[4] R. Vedantam, C. L. Zitnick, and D. Parikh, “CIDEr: Consensus-based Image Description Evaluation,” arXiv:1411.5726 [cs], Jun. 2015, arXiv: 1411.5726. [Online]. Available: http://arxiv.org/abs/1411.5726

SPICE

[5] P. Anderson, B. Fernando, M. Johnson, and S. Gould, “SPICE: Semantic Propositional Image Caption Evaluation,” arXiv:1607.08822 [cs], Jul. 2016, arXiv: 1607.08822. [Online]. Available: http://arxiv.org/abs/1607.08822

SPIDEr

[6] S. Liu, Z. Zhu, N. Ye, S. Guadarrama, and K. Murphy, “Improved Image Captioning via Policy Gradient optimization of SPIDEr,” 2017 IEEE Inter- national Conference on Computer Vision (ICCV), pp. 873–881, Oct. 2017, arXiv: 1612.00370. [Online]. Available: http://arxiv.org/abs/1612.00370

SPIDEr-max

[7] E. Labbé, T. Pellegrini, and J. Pinquier, “Is my automatic audio captioning system so bad? spider-max: a metric to consider several caption candidates,” Nov. 2022. [Online]. Available: https://hal.archives-ouvertes.fr/hal-03810396

FENSE

[8] Z. Zhou, Z. Zhang, X. Xu, Z. Xie, M. Wu, and K. Q. Zhu, Can Audio Captions Be Evaluated with Image Caption Metrics? arXiv, 2022. [Online]. Available: http://arxiv.org/abs/2110.04684

SPIDEr-FL

[9] DCASE website task6a description: https://dcase.community/challenge2023/task-automated-audio-captioning#evaluation

Citation

If you use SPIDEr-max, you can cite the following paper using BibTex :

@inproceedings{Labbe2022,
    title        = {Is my Automatic Audio Captioning System so Bad? SPIDEr-max: A Metric to Consider Several Caption Candidates},
    author       = {Labb\'{e}, Etienne and Pellegrini, Thomas and Pinquier, Julien},
    year         = 2022,
    month        = {November},
    booktitle    = {Proceedings of the 7th Detection and Classification of Acoustic Scenes and Events 2022 Workshop (DCASE2022)},
    address      = {Nancy, France},
    url          = {https://dcase.community/documents/workshop2022/proceedings/DCASE2022Workshop_Labbe_46.pdf}
}

If you use this software, please consider cite it as below :

@software{
    Labbe_aac-metrics_2023,
    author = {Labbé, Etienne},
    license = {MIT},
    month = {8},
    title = {{aac-metrics}},
    url = {https://github.com/Labbeti/aac-metrics/},
    version = {0.4.4},
    year = {2023},
}

Contact

Maintainer:

• Etienne Labbé "Labbeti": [e-mail]