Code to create networks that localize sounds sources in 3D environments
-
Main training/testing python script is
call_model_training_valid_pad_francl.py.- This script is responsible for processing the experiment parameters, validating the model folder, saving a copy of the experiment parameters there and ensuring the same folder isn't used for two different training rounds.
- An example set of parameeters can be found for testing in
slurm_CNN_spherical_testing_francl.shand inslurm_CNN_spherical_training_francl.sh.
-
Networks weights can be downloaded at: https://www.dropbox.com/sh/af6vaotxt41i7pe/AACfTzMxMLfv-Edmn33S4gTpa?dl=0
-
The model input should be nervegrams with their associated metadata saved into tensorflow records. The cochlear model we use is the PyCochleagram package . We have a wrapper to transform stereo
.wavfiles into the proper input available here: https://github.com/afrancl/BinauralDataGen- The precomputed training data that were used to train the original model can be downloaded here: https://drive.google.com/drive/folders/1pUyMmkurxEWFeofbLKy5-QVWrcMbM8Wz?usp=sharing
Note: Before running, please change the model save folder to point to your directory with the model architecture config file and data folder to point to your data. Both of these are in the associated shell scripts. The code itself contains no absolute paths.
To aid reproducibility and decrease setup time we provide a Singularity Image that contains all packages necessary to run the code without any further setup. The image is available on dropbox here: https://www.dropbox.com/s/ey74fiw4uquww0n/tfv1.13_tcmalloc.simg?dl=0