Closed Vatsalparsaniya closed 3 years ago
Check this out, I learned how to train the model here. https://github.com/Desklop/RNNoise_Wrapper/blob/master/TRAINING.md
I was wondering how were you able to tabulate that logs?
Thank you for your response, I have already checked that TRAINING.md
file and my training part is working good. I'm more concern about hyper parameter setting for new datasets.
I have added CSVLogger callback for tabular logs.
Ok, I will just paste the
csv_logger = CSVLogger('training.log') model.fit(X_train, Y_train, callbacks=[csv_logger])
onto the training python file?
Hi, if you don't mind can you please help me studying rnnoise? I have some questions about the process but I can't seem to understand well.
Yes, Your csvlogger callback will work.
I'm also looking into rnnoise. For the time being, I just understand rnnoise for training purposes; I have no idea about the 87 features of training. You are welcome to ask questions here, and if I know the solution, I will gladly respond.
The 87 features are here https://github.com/xiph/rnnoise/issues/152
My log file looks like this, how can I make it just like yours?
I'm wondering what are the contents of denoise_training file, I can't open it to check its contents and functions.
change CSVLogger('training.log')
to CSVLogger('training.csv')
Hi, I want to ask what is the use of
rnnoise-master / src / denoise_training datasets / test_training_set / all_clean.raw datasets / test_training_set / all_noise.raw 5000000 > train_logs / test_training_set / training_test_b_500k.f32
what does the denoise_training file contain? I can't view its contents.
Hi, I want to ask how to measure the accuracy of the model?
I don't recall any accuracy matrix they are using. Also, I'm not sure how we'll be able to determine how accurate this noise removal task will be.
Hi, I want to ask what does the 5000000 87 value mean? I think it pertains with the data set but I can't relate how.
hi @Vatsalparsaniya if you have already trained this model on MS-SNSD dataset, can you please provide link to the training files or weight files if possible ? thank you
I removed all data because this method did not work for me. However, I have listed down some training steps that may be help you.
- sudo apt-get install autoconf
- sudo apt-get install libtool
- sudo apt install ffmpeg
1) clone : https://github.com/Desklop/RNNoise_Wrapper.git
2) bash ./compile_rnnoise.sh
3) unzip rnnoise_master_20.11.2020.zip
4) add data into dataset folder
5) install req_train.txt
6) python3 training_utils/prepare_dataset_for_training.py -cf datasets/training_set/clean/ -nf datasets/training_set/noise/ -bca datasets/training_set/all_clean.raw -bna datasets/training_set/all_noise.raw
7) rnnoise-master/src/denoise_training datasets/training_set/all_clean.raw datasets/training_set/all_noise.raw 5000000 > train_logs/MS_SNSD/training_5000k.f32
8) python3 rnnoise-master/training/bin2hdf5.py train_logs/MS_SNSD/training_5000k.f32 5000000 87 train_logs/MS_SNSD/training_5000k.h5
9) run rnn_train_mod.py
10) python3 rnnoise-master/training/dump_rnn_mod.py train_logs/MS_SNSD/weights_5000k.hdf5 rnnoise-master/src/rnn_data.c rnnoise-master/src/rnn_data.h
11) cd rnnoise-master
12) make clean
13) ./autogen.sh
14) ./configure
15) make
16) cd ..
17) cp rnnoise-master/.libs/librnnoise.so.0.4.1 train_logs/MS_SNSD/librnnoise_MS_SNSD_5000k.so.0.4.1
hi @Vatsalparsaniya , so basically all i want to know is.. were you able to remove other kind of noises like background tv or any thumping sound or any ?
Hi,
I'm currently working on the MS-SNSD dataset. I'm concerned about training. With a newly trained model, it appears that the model is not effectively removing noise. If someone could provide any original model training logs, such as loss value, or any suggestions for new training data, that would be greatly appreciated. For your convenience, I've attached my training logs. Mentioned logs are related to
5000000 87
matrix size.Thank you.