Finished up the gan specific functions, and most of the spectogram normalization and resizing functions (lifted straight from specgan). Implementing an inverse spectogram function is yet to be done. Debugging and optimization after that.
Added a save function to the generator model every 20 epochs. Also added a function in input.py to convert the audio arrays into melspectograms. Added a display function if we want to see the generated spectograms. TODO: Discuss how to normalize (to (0, 1)) and denormalize the spectograms. Didn't really get how they did it in specgan
Added code to train.py to train a VAE model. TODO: Figure out the input_shape.
Added gan.py with the model definitions and the train function. SpecGAN architecture, with a couple differences. Used the basic loss function, can try out others later. TODO: A couple preprocessing and display functions, as well as a save function for the model at points. Will work on that.
###18-Aug-2021 : updated visualize_data in input.py, Arvind Modified code in visualize_data to plot a graph for a given .wav file using librosa. TODO: Decide what to pass into conv layers.
Added code to the functions load_audio to input a filepath containing .wav files and return a list of tensors containing decoded float32 np arrays. Added functionality to play_sample using PyAudio that takes in a filename and plays the corresponding .wav audio. Made functional temporarily visualize_waveform using tensorflow(to be updated) and modified parameters of extract_data to include an option of instrument of choice. ToDo Add a readme file listing allowed inputs for the same.
Added code to 'extract_data' to store files of a certain instrument into a list, and shuffle them. Added a new function load_data_mono in order to work with mono music instead of stereo - a decision we will have to make. The code is still incomplete, have to decide on what exactly needs to be done - in what format Wavenet takes input.
Added the input.py and train.py files with a rudimentary structure of functions to be implemented. More functions can be added if necessary. Defined classes for two models - VAE and WaveNet, to be implemented later. Download the IRMAS dataset from its mirror at zenodo (only the training set).
{information about the commit, what's new how it affects other code (if it does)}