1""" 2Utility functions that are commonly used 3""" 4 5import numpy as np 6from scipy.signal import windows, lfilter 7from prettytable import PrettyTable 8 9 10# Source: https://gist.github.com/thongonary/026210fc186eb5056f2b6f1ca362d912 11def count_parameters(model): 12 table = PrettyTable(["Modules", "Parameters"]) 13 total_params = 0 14 for name, parameter in model.named_parameters(): 15 if not parameter.requires_grad: continue 16 param = parameter.numel() 17 table.add_row([name, param]) 18 total_params+=param 19 print(table) 20 print(f"Total Trainable Params: {total_params}") 21 return total_params 22 23def stft(x, w = 'boxcar', N = 320, H = 160): 24 x = np.concatenate([x,np.zeros(N)]) 25 # win_custom = np.concatenate([windows.hann(80)[:40],np.ones(240),windows.hann(80)[40:]]) 26 return np.stack([np.fft.rfft(x[i:i + N]*windows.get_window(w,N)) for i in np.arange(0,x.shape[0]-N,H)]) 27 28def random_filter(x): 29 # Randomly filter x with second order IIR filter with coefficients in between -3/8,3/8 30 filter_coeff = np.random.uniform(low = -3.0/8, high = 3.0/8, size = 4) 31 b = [1,filter_coeff[0],filter_coeff[1]] 32 a = [1,filter_coeff[2],filter_coeff[3]] 33 return lfilter(b,a,x) 34 35def feature_xform(feature): 36 """ 37 Take as input the (N * 256) xcorr features output by LPCNet and perform the following 38 1. Downsample and Upsample by 2 (followed by smoothing) 39 2. Append positional embeddings (of dim k) coresponding to each xcorr lag 40 """ 41 42 from scipy.signal import resample_poly, lfilter 43 44 45 feature_US = lfilter([0.25,0.5,0.25],[1],resample_poly(feature,2,1,axis = 1),axis = 1)[:,:feature.shape[1]] 46 feature_DS = lfilter([0.5,0.5],[1],resample_poly(feature,1,2,axis = 1),axis = 1) 47 Z_append = np.zeros((feature.shape[0],feature.shape[1] - feature_DS.shape[1])) 48 feature_DS = np.concatenate([feature_DS,Z_append],axis = -1) 49 50 # pos_embedding = [] 51 # for i in range(k): 52 # pos_embedding.append(np.cos((2**i)*np.pi*((np.repeat(np.arange(feature.shape[1]).reshape(feature.shape[1],1),feature.shape[0],axis = 1)).T/(2*feature.shape[1])))) 53 54 # pos_embedding = np.stack(pos_embedding,axis = -1) 55 56 feature = np.stack((feature_DS,feature,feature_US),axis = -1) 57 # feature = np.concatenate((feature,pos_embedding),axis = -1) 58 59 return feature 60