qinggangsun / unknown_number_source_separation

Source separation of underwater acoustic radiated noise signals from ships with unknown numbers of signals. Using keras 2.2.4 with tensorflow-gpu 1.12.0 backend.

unknown_number_source_separation

Underwater single-channel acoustic source separation with unknown numbers using autoencoder nerual networks. Using keras-gpu 2.2.4 with tensorflow-gpu 1.12.0 backend.

How to cite this work

This is the official code of the blow article. Please cite this work in your publications as :

@misc{sun2022source,
      title={Source Separation of Unknown Numbers of Single-Channel Underwater Acoustic Signals Based on Autoencoders}, 
      author={Qinggang Sun and Kejun Wang},
      year={2022},
      eprint={2207.11749},
      archivePrefix={arXiv},
      primaryClass={cs.SD}
} 

How to use

1. Download and organize data files as "data_dir_tree.md".
    The official website of the shipsEar database is https://atlanttic.uvigo.es/underwaternoise/ .
    Users may contact the author David Santos Domínguez through email [email protected] to get the database.
2. Install dependent packages in the "requirements.txt" file.
3. Run
	prepare_data_shipsear_recognition_mix_s0tos3.py
	recognition_mix_shipsear_s0tos3_preprocess.py
	prepare_data_shipsear_separation_mix_s0tos3.py
	separation_mix_shipsear_s0tos3_preprocess.py
   to preprocess datas.
4. Experiments:
	(0) Known numbers of source separation with multiple autoencoders.
		train_separation_multiple_autoencoder.py
	(1) Known numbers of source separation with multiple output autoencoders.
		train_separation_known_number.py
	(2) Algorithm 1.
		train_separation_one_autoencoder.py
	(3) Algorithm 2.
		train_single_source_autoencoder_ns.py
		train_single_source_autoencoder_ns_search_encoded_dense.py
	(4) Algorithm 3.
		train_single_source_autoencoder_ns.py
		train_separation_one_autoencoder_freeze_decoder.py

Reference

Please cite the original work as :
Generative sourceseparation with GANs
Two-Step Sound Source Separation: Training on Learned Latent Targets
Dual-Path RNN for Single-Channel Speech Separation
Dual-Path RNN for Single-Channel Speech Separation (in Keras-Tensorflow2)
Conv-TasNet: Surpassing Ideal Time–Frequency Magnitude Masking for Speech Separation
conv-tasnet
Deep Learning for Blind Source Separation
Conv-TasNet
Conv-TasNet
Wave-U-Net
Wave-U-net-TF2
Sound Separation
bsseval

unknown_number_source_separation's Issues

Publish some fatal problems affecting parts of the results of the experiments.

Fatal problems:

• DPRNN models cannot save weights caused by a Lambda layer.
• A Lambda layer that slices specific channel of data doesnot run correctly.

Small bugs:

• A sample of s3 is mislabeled in file "data_dir_tree.md". Fortunately, this mistake did not affect the results of previous experiments.
• A default path of data in file "data_dir_tree.md" is wrong.
• A bug about decoder input in model 8.
• Some models are not compatible with specific batch size.

New features:

• No "environment.yml" for linux os.

These problems will be fixed in next published version.

Z_4_ns_train.hdf5?

Hello, I've been trying to try out your code for research purposes, but ran into a problem. I managed to get past the small bugs such as the default data path and such, but when I try to run the train_separation_multiple_autoencoder.py file, I get the following problem:
OSError: Unable to open file (unable to open file: name = '../data/shipsEar/mix_separation\10547_10547\s0tos3\mix_1to3sep\wavmat\max_one_rand\Z_4_ns_train.hdf5', errno = 2, error message = 'No such file or directory', flags = 0, o_flags = 0)

The problem is, Z_4_ns_train.hdf5 is not generated in any of the previous codes. It did generate other .hdf5 files, but never Z_4_ns_train or the like, and I cannot seem to figure out how to make it do that.