Fixed a bug in CoordConv layer implementation, where coordinates were not scaled correctly along each dimensions.

Added CoordConv layer implementation, sourced from [https://raw.githubusercontent.com/uber-research/CoordConv].

Added utility functions to generate multilabel azimuth encoding from azimuth values, and corresponding test.

Added some code to do some class balancing, repeating 'center' class inputs, when training the 'left_center_right' model.

Updated some experiment configuration. Modified log directory name to include the name of the model trained.

Removed non-linear activation function 'sigmoid' in models, as my brother told me that it is not appropriate for regression. It effectively improves the training of the 'default' model, while it doesn't change the behavior of 'one_sided' model.

Added tensorboard to requirements, and added journalization of LOSS when training a model. Updated default experiment configuration.

Added a 'one_sided' model, that localize azimuth in [0,180] only, and added code to be able to train this model.

Renamed 'train2.py' script to 'train.py', and lowered the value of random_shift parameter from 100 to 10 for faster model training.

Removed randomization process in generate_augmented_labelled_dataset(), at this process is now done inside the model.

Updated train2.py script to do left and right zero padding when using convolution with mode 'valid'. Stimulus truncature with n_samples, which was incremented for each training iteration, has also been disabled. It should be removed in the future.

Added parameters in default_model_creator() for on-line data augmentation, random scale and random shift.

Refactoring of generate_augmented_labelled_dataset(). Added a parameter to select the convolution mode ('full', 'valid', by default, or 'same').

Added a test to visually check that scipy.signal.convolve modes {'full', 'valid', 'same'} have the correct behavior.

Added GtfLayer(), a Conv1D layer with non trainable kernels who's values are loaded from time-reversed Gammatone filters impulses responses, which are either passed directly or loaded from a directory containing the Gammatone filters. A test is used to verify that the layer can process some input, and some lines can be uncommented to check visually that the filters are correctly applied to the input (a dirac with some silence before and after).

Input of the model has been standardized to BWHF (Batch, Width, Height, Features) shape. RandomScale2DLayer and RandomShift2DLayer has been incorporated into the model in order to apply these two steps of data augmentation on-the-fly during the training of the model, but not during the test of the model.

Fixed bug in RandomScale2DLayer, where the shape must input tensor shape must be evaluated using a dynamic pointer scalar in the case where the batch size is dynamically allocated.

Added two layers to apply random scaling or random shift for tensors of standardized shape BWHF (Batch, Width, Height, Features).
When training is False, the layers do not apply random scaling or shifting. However, the random shift layer still apply a right zero padding to ensure that the shape of the layer output is identical during training and testing.

Added a bank of 20 Gammetone filters, of 512 coefficients and sampling rate 44100hz, covering the frequency range 100-20000hz. Filters are windowed with the right half of a Hann window (of size 1024).

Added a script to generate Gammatone FIR filters in .wav format, by using the library from Bingo Todd.
Because a rectangular windowing produces ringing in the magnitude of the frequency response at low frequencies when the filter have ~512 coefficients, we apply a windowing of the filter impulse response in order to improve the filter quality. After windowing, a normalization is applied to restore the correct magnitude peak in the frequency response domain.

changed split of filename in order to work on windows and unix file systems

Added set_gpus function in utils to support CPU or GPU configuration. Changed script using previous method to use the new function

IRCAM database splitted into raw hrirs, diffuse field compensated hrirs, and max amplitude compensated hrirs.