furkanyesiler / acoss Goto Github PK

Hi,

according to the README, everything seems to be end-to-end when running benchmark across the whole covers80 dataset.

Is there a way to simply compare two audio using any of the algorithms, and determine whether they are indeed cover versions of each other?

• get the small batch of metadata
• write initial script for cross-referencing
• make sure the intersection is large enough on full metadata (+ subsets)

Hi! In acoss/util.py, line 101, if the work_id or track_id is all composed of integers, then Pandas will infer that the data type is numpy.int64, and there will be an error when joining int and string together. Maybe a datatype convertion is needed here.

Use the audio novelty functions from madmom as input to the algorithm

Hi, i am trying to run the benchmark.

benchmark(dataset_csv=COVERS_80_CSV, 
    feature_dir=feature_dir,
    algorithm="Serra09", 
    parallel=True,
    n_workers=2)

This error is given.

Traceback (most recent call last):
  File "/data3/lootiangkuan/AudioHashing/QbH/acoss/acoss/coverid.py", line 65, in benchmark
    serra09.all_pairwise(parallel, n_cores=n_workers, symmetric=True)
  File "/data3/lootiangkuan/AudioHashing/QbH/acoss/acoss/algorithms/algorithm_template.py", line 191, in all_pairwise
    progressbar.next()
UnboundLocalError: local variable 'progressbar' referenced before assignment

Hi, I git cloned this on 9th April 2020.

I was running

batch_feature_extractor(dataset_csv=COVERS_80_CSV, 
                        audio_dir=audio_dir, 
                        feature_dir=feature_dir,
                        n_workers=4,
                        mode="parallel", 
                        params=extractor_profile)

And the following error was returned

Traceback (most recent call last):
  File "./debug.py", line 56, in <module>
    batch_feature_extractor(dataset_csv=COVERS_80_CSV,
  File "/data3/lootiangkuan/AudioHashing/QbH/acoss/acoss/extractors.py", line 150, in batch_feature_extractor
    Parallel(n_jobs=n_threads, verbose=1)(delayed(compute_features_from_list_file)\
NameError: name 'n_threads' is not defined

• Add a python script which compute all the necessary mid-level audio features from an input audio file or from a collection of audio files

The following 'code-refactoring-todos' were identified along with @furkanyesiler

The changes are made in the packaging dev branch of the repo.

Preprocess -

Extractors

• Proper docstrings for all the functions
• Logging error audio files in the batch extraction process
• Specify a list of features as cmd args.
• Change exception in compute features from list file
• Unify string formatting using {}.format
• Add cmd-line option for the feature extraction of other datasets (eg. from a given path)
• Refactor cmd-line args for running non-parallel jobs (eg. cluster, cpu thing).
• Make feature extractor unison for any file paths.

Features

• Wrapper to call feature methods on a specific python version. If not available throw an exception.
• Clean the file

Local config

• Single local config file for the whole repo.

Requirements

• Separate requirements.txt file for various python versions

README

• Proper description and toy examples.

Benchmarking and CoverStats (@ctralie can you give me a hand on this part? )

• Adhere code to pep-8 standards
• Make getEvalStatistics clean
• Follow conventions for writing docstrings
• Standardize argparse of cover similarity algorithms
• Clean all functions in statistics/ with pep8, etc
• Make a single command line interface to call all benchmarking algorithms with a consistent set of parameters

Main repo and packaging

• Add CI tools and unit tests.
• Package acoss as a python library and build pip wheels
• Logo for the repo and dataset
• Scripts for building automated docs.
• Proper README with setup instructions and example files
• Check file storage formats for distributing datasets.