This repository contains code to evaluate translation metrics on audio captioning predictions.
The code from the Microsoft COCO caption evaluation repository, in the folder coco_caption, is used to evaluate the metrics. The code has been refactored to work with Python 3 and to also evaluate the SPIDEr metric. Image-specific names and comments in-code were also changed to be audio-specific.
Before evaluating metrics, the user must run coco_caption/get_stanford_models.sh (or follow the commands in the
bash script) to download the libraries necessary for evaluating the SPICE metric.
SPICE evaluation uses 8GB of RAM and METEOR uses 2GB (both use Java). To limit RAM usage go to
coco_caption/pycocoevalcap and meteor/meteor.py:18 or spice/spice.py:63 respectively and change the third argument
of the java command.
The evaluate_metrics() function inside eval_metrics.py takes as inputs the csv file with the predicted captions and
the csv file with the reference captions. The optional parameter nb_reference_captions determines how many
reference captions are used to evaluate the metrics (5 by default).
The input files can be given either as file
paths (string or pathlib.Path) or lists of dicts with a dict for each row, the dicts having the column headers as keys
(as given by csv.DictReader in Python). The prediction file must have the fields file_name and caption_predicted and the
reference file must have the fields file_name and caption_reference_XX with XX being the two-digit index of the
caption, e.g. caption_reference_01,...,caption_reference_05 with five reference captions.
The metric evaluation function outputs the evaluated metrics in a dict with the lower case metric names as keys. One score is evaluated for each audio file and its predicted caption. Additionally, a single score is evaluated for the whole dataset. The format of the output is the following:
{<metric name in lower case>: {
'score': <single score value>,
'scores': {
<audio file name>: <per-file score value>
}
}}
This code is maintained by lippings