DynamicInput.py

####################################################################################################
#                       FUNCTIONS TO HANDLE DYNAMIC INPUT SPECTOGRAM SIZE                          #
####################################################################################################
# Author: Rajesh R (Git: its-rajesh)
# Date: 14/Januray/2023

import numpy as np


class DynamicInput:

    def __init__(self):
        pass


    '''
    FUNCTION: ENCODES (SPLITS & MERGE) THE INPUT SPECTOGRAM INTO VARIOUS BINS (See Ref. Figure)
    USE: To handle dynamic inputs to neural netwok.
    Parameters:
    (1) array: Input spectogram of size m*n
    (2) frame: No of bins to be grouped together. Default is 3, If frame = no of columns in input spect/3, 
        and skip=3, then output is same as input.
    (3) skip: Overlap within the bins (1<=skip<=3), default is 1 to maintain loss of info.
        either use skip =1 or frame = 1 for no loss of info.

    Note: Batches/Grouping is always 3.
    '''

    def encode(self, array, frame=3, skip=1):
        
        data = []
        for i in range(0, array.shape[1]-(3*frame-1), skip):
            y_ = array[:, i:i+frame]
            y = array[:, i+frame:i+(2*frame)]
            y__ = array[:, i+(2*frame):i+(3*frame)]
            concat = np.concatenate((y_, y, y__), axis=0, casting="same_kind")
            data.append(concat)
            
        self.encoded = np.array(data)
        return self.encoded


    '''
    FUNCTION: DECODES (REVERSE OF ENCODE) THE SPLIT BINS INTO ORIGINAL SPECTOGRAM (See Ref. Figure)
    USE: To handle dynamic inputs to neural netwok.
    Parameters:
    (1) array: Input spectogram bins of size l*m*n
    (2) frame: Same value as given in encode function
    (3) skip: Same value as given in encode function

    Note: Batches/Grouping is always 3.
    '''

    def decode(self, array, frame=3, skip=1):
        reconst = []
        l, m, n = array.shape
        flag = False
        for i in array:
            
            y_ = i[:int(m/3)]
            y = i[int(m/3):int(2*(m/3))]
            y__ = i[int(2*(m/3)):m]

            if skip == 1:
                if flag:
                    t = np.array([y_[:, 2]]).T
                    reconst = np.concatenate((reconst, t), axis=1, casting="same_kind")
                else:
                    reconst = y_
                    
            if skip == 2:
                if flag:
                    t = np.array([y_[:, 1:2]]).T
                    reconst = np.concatenate((reconst, t), axis=1, casting="same_kind")
                else:
                    reconst = y_
                    
            if skip == 3:
                if flag:
                    reconst = np.concatenate((reconst, y_), axis=1, casting="same_kind")
                else:
                    reconst = y_
                    
            flag = True
            
        if skip == 1:
            reconst = np.concatenate((reconst, y, y__), axis=1, casting="same_kind")

        self.decoded = np.array(reconst)
        return self.decoded