Fluency ENhanced Sentence-bert Evaluation (FENSE), metric for audio caption evaluation. And Benchmark dataset AudioCaps-Eval, Clotho-Eval.

Last update: Dec 23, 2022

Overview

FENSE

The metric, Fluency ENhanced Sentence-bert Evaluation (FENSE), for audio caption evaluation, proposed in the paper "Can Audio Captions Be Evaluated with Image Caption Metrics?"

The main branch contains an easy-to-use interface for fast evaluation of an audio captioning system.

Online demo avaliable at https://share.streamlit.io/blmoistawinde/fense/main/streamlit_demo/app.py .

To get the dataset (AudioCaps-Eval and Clotho-Eval) and the code to reproduce, please refer to the experiment-code branch.

Installation

Clone the repository and pip install it.

git clone https://github.com/blmoistawinde/fense.git
cd fense
pip install -e .

Usage

Single Sentence

To get the detailed scores of each component for a single sentence.

from fense.evaluator import Evaluator

print("----Using tiny models----")
evaluator = Evaluator(device='cpu', sbert_model='paraphrase-MiniLM-L6-v2', echecker_model='echecker_clotho_audiocaps_tiny')

eval_cap = "An engine in idling and a man is speaking and then"
ref_cap = "A machine makes stitching sounds while people are talking in the background"

score, error_prob, penalized_score = evaluator.sentence_score(eval_cap, [ref_cap], return_error_prob=True)

print("Cand:", eval_cap)
print("Ref:", ref_cap)
print(f"SBERT sim: {score:.4f}, Error Prob: {error_prob:.4f}, Penalized score: {penalized_score:.4f}")

System Score

To get a system's overall score on a dataset by averaging sentence-level FENSE, you can use eval_system.py, with your system outputs prepared in the format like test_data/audiocaps_cands.csv or test_data/clotho_cands.csv .

For AudioCaps test set:

python eval_system.py --device cuda --dataset audiocaps --cands_dir ./test_data/audiocaps_cands.csv

For Clotho Eval set:

python eval_system.py --device cuda --dataset clotho --cands_dir ./test_data/clotho_cands.csv

Performance Benchmark

We benchmark the performance of FENSE with different choices of SBERT model and Error Detector on the two benchmark dataset AudioCaps-Eval and Clotho-Eval. (*) is the combination reported in paper.

AudioCaps-Eval

SBERT	echecker	HC	HI	HM	MM	total
paraphrase-MiniLM-L6-v2	none	62.1	98.8	93.7	75.4	80.4
paraphrase-MiniLM-L6-v2	tiny	57.6	94.7	89.5	82.6	82.3
paraphrase-MiniLM-L6-v2	base	62.6	98	82.5	85.4	85.5
paraphrase-TinyBERT-L6-v2	none	64	99.2	92.5	73.6	79.6
paraphrase-TinyBERT-L6-v2	tiny	58.6	95.1	88.3	82.2	82.1
paraphrase-TinyBERT-L6-v2	base	64.5	98.4	91.6	84.6	85.3(*)
paraphrase-mpnet-base-v2	none	63.1	98.8	94.1	74.1	80.1
paraphrase-mpnet-base-v2	tiny	58.1	94.3	90	83.2	82.7
paraphrase-mpnet-base-v2	base	63.5	98	92.5	85.9	85.9

Clotho-Eval

SBERT	echecker	HC	HI	HM	MM	total
paraphrase-MiniLM-L6-v2	none	59.5	95.1	76.3	66.2	71.3
paraphrase-MiniLM-L6-v2	tiny	56.7	90.6	79.3	70.9	73.3
paraphrase-MiniLM-L6-v2	base	60	94.3	80.6	72.3	75.3
paraphrase-TinyBERT-L6-v2	none	60	95.5	75.9	66.9	71.8
paraphrase-TinyBERT-L6-v2	tiny	59	93	79.7	71.5	74.4
paraphrase-TinyBERT-L6-v2	base	60.5	94.7	80.2	72.8	75.7(*)
paraphrase-mpnet-base-v2	none	56.2	96.3	77.6	65.2	70.7
paraphrase-mpnet-base-v2	tiny	54.8	91.8	80.6	70.1	73
paraphrase-mpnet-base-v2	base	57.1	95.5	81.9	71.6	74.9

Reference

If you use FENSE in your research, please cite:

@misc{zhou2021audio,
      title={Can Audio Captions Be Evaluated with Image Caption Metrics?}, 
      author={Zelin Zhou and Zhiling Zhang and Xuenan Xu and Zeyu Xie and Mengyue Wu and Kenny Q. Zhu},
      year={2021},
      eprint={2110.04684},
      archivePrefix={arXiv},
      primaryClass={cs.SD}
}

You might also like...

I-BERT: Integer-only BERT Quantization

I-BERT: Integer-only BERT Quantization HuggingFace Implementation I-BERT is also available in the master branch of HuggingFace! Visit the following li

139 Dec 27, 2022

Source code for NAACL 2021 paper "TR-BERT: Dynamic Token Reduction for Accelerating BERT Inference"

TR-BERT Source code and dataset for "TR-BERT: Dynamic Token Reduction for Accelerating BERT Inference". The code is based on huggaface's transformers.

37 Oct 30, 2022

LV-BERT: Exploiting Layer Variety for BERT (Findings of ACL 2021)

LV-BERT Introduction In this repo, we introduce LV-BERT by exploiting layer variety for BERT. For detailed description and experimental results, pleas

14 Aug 24, 2022

The source codes for ACL 2021 paper 'BoB: BERT Over BERT for Training Persona-based Dialogue Models from Limited Personalized Data'

BoB: BERT Over BERT for Training Persona-based Dialogue Models from Limited Personalized Data This repository provides the implementation details for

124 Dec 27, 2022

Fluency ENhanced Sentence-bert Evaluation (FENSE), metric for audio caption evaluation. And Benchmark dataset AudioCaps-Eval, Clotho-Eval.

Related tags

Overview

FENSE

Installation

Usage

Single Sentence

System Score

Performance Benchmark

Reference

You might also like...

I-BERT: Integer-only BERT Quantization

Source code for NAACL 2021 paper "TR-BERT: Dynamic Token Reduction for Accelerating BERT Inference"

LV-BERT: Exploiting Layer Variety for BERT (Findings of ACL 2021)

The source codes for ACL 2021 paper 'BoB: BERT Over BERT for Training Persona-based Dialogue Models from Limited Personalized Data'

Pure python PEMDAS expression solver without using built-in eval function

Generating images from caption and vice versa via CLIP-Guided Generative Latent Space Search

TAP: Text-Aware Pre-training for Text-VQA and Text-Caption, CVPR 2021 (Oral)

Yet another video caption

Fine-grained Control of Image Caption Generation with Abstract Scene Graphs

Releases(V0.1)

V0.1(Oct 2, 2021)

Owner

Zhiling Zhang

Visual Tracking by TridenAlign and Context Embedding

Incremental Transformer Structure Enhanced Image Inpainting with Masking Positional Encoding (CVPR2022)

Blender add-on: Add to Cameras menu: View → Camera, View → Add Camera, Camera → View, Previous Camera, Next Camera

Display, filter and search log messages in your terminal

This repository contains the code to replicate the analysis from the paper "Moving On - Investigating Inventors' Ethnic Origins Using Supervised Learning"

Code for Universal Semi-Supervised Semantic Segmentation models paper accepted in ICCV 2019

🔥 Real-time Super Resolution enhancement (4x) with content loss and relativistic adversarial optimization 🔥

(to be released) [NeurIPS'21] Transformers Generalize DeepSets and Can be Extended to Graphs and Hypergraphs

Pytorch implementations of the paper Value Functions Factorization with Latent State Information Sharing in Decentralized Multi-Agent Policy Gradients

Keras-tensorflow implementation of Fully Convolutional Networks for Semantic Segmentation（Unfinished）

Source code for our paper "Improving Empathetic Response Generation by Recognizing Emotion Cause in Conversations"

Simple reference implementation of GraphSAGE.

Gesture Volume Control v.2

Pytorch-diffusion - A basic PyTorch implementation of 'Denoising Diffusion Probabilistic Models'

Using this you can control your PC/Laptop volume by Hand Gestures (pinch-in, pinch-out) created with Python.

A 10000+ hours dataset for Chinese speech recognition

Teaching end to end workflow of deep learning

A short and easy PyTorch implementation of E(n) Equivariant Graph Neural Networks

A texturizer that I just made. Nothing special here.

Synthetic LiDAR sequential point cloud dataset with point-wise annotations