RNN.rar_u net 分离 人声_voice separation_人声_人声分离_盲源分离 深度学习

import tensorflow as tf #from model import Model import os import shutil from data import Data #from preprocess import to_spectrogram, get_magnitude #from utils import Diff from config import TrainConfig import librosa import numpy as np from tensorflow.contrib.rnn import GRUCell, MultiRNNCell import soundfile as sf from config import EvalConfig, ModelConfig from mir_eval.separation import bss_eval_sources import datetime def get_phase(stft_maxtrixes): return np.angle(stft_maxtrixes) def spec_to_batch(src): freq, n_frames = src.shape ModelConfigSEQ_LEN=128 # Padding pad_len = 0 if n_frames % ModelConfigSEQ_LEN > 0: pad_len = (ModelConfigSEQ_LEN - (n_frames % ModelConfigSEQ_LEN)) pad_width = ((0, 0), (0, pad_len)) padded_src = np.pad(src, pad_width=pad_width, mode='constant', constant_values=0) #print("padded_src shape:",padded_src.shape) assert(padded_src.shape[-1] % ModelConfigSEQ_LEN == 0) batch=padded_src.transpose(1,0)[:,:,np.newaxis] batch=np.reshape(batch,(-1,ModelConfigSEQ_LEN,freq,1)) batch=batch.transpose(0,2,1,3) #batch = np.reshape(padded_src.transpose(1,0), (-1, freq, ModelConfigSEQ_LEN,1)) return batch, padded_src def batch_to_spec(src): a=src.transpose(0,2,1,3) b=np.reshape(a,(-1,512)).transpose(1,0) return b def get_stft_matrix(magnitudes, phases): return magnitudes * np.exp(1.j * phases) def griffin_lim(mag, len_frame=1024, len_hop=256, num_iters=50, phase_angle=None, length=None): assert(num_iters > 0) if phase_angle is None: phase_angle = np.pi * np.random.rand(*mag.shape) spec = get_stft_matrix(mag, phase_angle) for i in range(num_iters): wav = librosa.istft(spec, win_length=len_frame, hop_length=len_hop, length=length) if i != num_iters - 1: spec = librosa.stft(wav, n_fft=len_frame, win_length=len_frame, hop_length=len_hop) _, phase = librosa.magphase(spec) phase_angle = np.angle(phase) spec = get_stft_matrix(mag, phase_angle) return wav def to_wav_mag_only(mag, init_phase, len_frame=ModelConfig.L_FRAME, len_hop=ModelConfig.L_HOP, num_iters=50): return np.array(list(map(lambda m, p: griffin_lim(m, len_frame, len_hop, num_iters=num_iters, phase_angle=p), list(zip(mag, init_phase))))) def to_wav(mag, phase, len_hop=ModelConfig.L_HOP): stft_maxrix = get_stft_matrix(mag, phase) return np.array([librosa.istft(s, hop_length=len_hop) for s in stft_maxrix]) def fuyuan(mag, pha): wav1=np.ones(pha.shape,dtype=np.complex128) for h in range(wav1.shape[0]): for j in range(wav1.shape[1]): wav1[h][j]=np.complex(np.cos(pha[h][j])*mag[h][j],np.sin(pha[h][j])*mag[h][j]) return wav1 def bss_eval_global(mixed_wav, src1_wav, src2_wav, pred_src1_wav, pred_src2_wav): len_cropped = pred_src1_wav.shape[-1] src1_wav = src1_wav[:, :len_cropped] src2_wav = src2_wav[:, :len_cropped] mixed_wav = mixed_wav[:, :len_cropped] gnsdr = gsir = gsar = np.zeros(2) total_len = 0 for i in range(EvalConfig.NUM_EVAL): sdr, sir, sar, _ = bss_eval_sources(np.array([src1_wav[i], src2_wav[i]]), np.array([pred_src1_wav[i], pred_src2_wav[i]]), False) sdr_mixed, _, _, _ = bss_eval_sources(np.array([src1_wav[i], src2_wav[i]]), np.array([mixed_wav[i], mixed_wav[i]]), False) nsdr = sdr - sdr_mixed gnsdr += len_cropped * nsdr gsir += len_cropped * sir gsar += len_cropped * sar total_len += len_cropped gnsdr = gnsdr / total_len gsir = gsir / total_len gsar = gsar / total_len print("gnsdr:",gnsdr,"gsir:",gsir,"gsar:",gsar) #return gnsdr, gsir, gsar def train(): global_step = tf.Variable(0, dtype=tf.int32, trainable=False) is_train=tf.placeholder(tf.bool) batchsize=tf.placeholder(tf.float32) x_mixed = tf.placeholder(tf.float32, [None, 512, 128,1], name='x_mixed') y_src1= tf.placeholder(tf.float32, [None, 512, 128,1], name='y_pre1') y_src2= tf.placeholder(tf.float32, [None, 512, 128,1], name='y_pre2') yconv1=tf.placeholder(tf.float32, [None, 256, 64,16]) yconv2=tf.placeholder(tf.float32, [None, 128, 32,32]) yconv3=tf.placeholder(tf.float32, [None, 64, 16,64]) yconv4=tf.placeholder(tf.float32, [None, 32, 8,128]) yconv5=tf.placeholder(tf.float32, [None, 16, 4,256]) ayconv1=tf.placeholder(tf.float32, [None, 256, 64,16]) ayconv2=tf.placeholder(tf.float32, [None, 128, 32,32]) ayconv3=tf.placeholder(tf.float32, [None, 64, 16,64]) ayconv4=tf.placeholder(tf.float32, [None, 32, 8,128]) ayconv5=tf.placeholder(tf.float32, [None, 16, 4,256]) #第1个卷积层 filter1=tf.Variable(tf.random_normal([5,5,1,16])) conv_out1=tf.nn.conv2d(x_mixed, filter1, strides=[1, 2, 2, 1], padding='SAME') batch_mean1, batch_var1 = tf.nn.moments(conv_out1, [0, 1, 2], keep_dims=True) shift1 = tf.Variable(tf.zeros([16])) scale1 = tf.Variable(tf.ones([16])) epsilon = 1e-3 batnor1 = tf.nn.batch_normalization(conv_out1, batch_mean1, batch_var1, shift1, scale1, epsilon) out1=tf.nn.leaky_relu(batnor1) yconv1=out1 afilter1=tf.Variable(tf.random_normal([5,5,1,16])) aconv_out1=tf.nn.conv2d(x_mixed, afilter1, strides=[1, 2, 2, 1], padding='SAME') abatch_mean1,abatch_var1 = tf.nn.moments(aconv_out1, [0, 1, 2], keep_dims=True) ashift1 = tf.Variable(tf.zeros([16])) ascale1 = tf.Variable(tf.ones([16])) epsilon = 1e-3 abatnor1 = tf.nn.batch_normalization(aconv_out1, abatch_mean1, abatch_var1, ashift1, ascale1, epsilon) aout1=tf.nn.leaky_relu(abatnor1) ayconv1=aout1 #第2个卷积层 filter2=tf.Variable(tf.random_normal([5,5,16,32])) conv_out2=tf.nn.conv2d(out1, filter2, strides=[1, 2, 2, 1], padding='SAME') batch_mean2, batch_var2 = tf.nn.moments(conv_out2, [0, 1, 2], keep_dims=True) shift2 = tf.Variable(tf.zeros([32])) scale2 = tf.Variable(tf.ones([32])) batnor2 = tf.nn.batch_normalization(conv_out2, batch_mean2, batch_var2, shift2, scale2, epsilon) out2=tf.nn.leaky_relu(batnor2) yconv2=out2 afilter2=tf.Variable(tf.random_normal([5,5,16,32])) aconv_out2=tf.nn.conv2d(aout1, afilter2, strides=[1, 2, 2, 1], padding='SAME') abatch_mean2, abatch_var2 = tf.nn.moments(aconv_out2, [0, 1, 2], keep_dims=True) ashift2 = tf.Variable(tf.zeros([32])) ascale2 = tf.Variable(tf.ones([32])) abatnor2 = tf.nn.batch_normalization(aconv_out2, abatch_mean2, abatch_var2, ashift2, ascale2, epsilon) aout2=tf.nn.leaky_relu(abatnor2) ayconv2=aout2 #第3个卷积层 filter3=tf.Variable(tf.random_normal([5,5,32,64])) conv_out3=tf.nn.conv2d(out2, filter3, strides=[1, 2, 2, 1], padding='SAME') batch_mean3, batch_var3 = tf.nn.moments(conv_out3, [0, 1, 2], keep_dims=True) shift3 = tf.Variable(tf.zeros([64])) scale3 = tf.Variable(tf.ones([64])) batnor3 = tf.nn.batch_normalization(conv_out3, batch_mean3, batch_var3, shift3, scale3, epsilon) out3=tf.nn.leaky_relu(batnor3) yconv3=out3 afilter3=tf.Variable(tf.random_normal([5,5,32,64])) aconv_out3=tf.nn.conv2d(aout2, afilter3, strides=[1, 2, 2, 1], padding='SAME') abatch_mean3, abatch_var3 = tf.nn.moments(aconv_out3, [0, 1, 2], keep_dims=True) ashift3 = tf.Variable(tf.zeros([64])) ascale3 = tf.Variable(tf.ones([64])) abatnor3 = tf.nn.batch_normalization(aconv_out3, abatch_mean3, abatch_var3, ashift3, ascale3, epsilon) aout3=tf.nn.leaky_relu(abatnor3) ayconv3=aout3 #第4个卷积层 filter4=tf.Variable(tf.random_normal([5,5,64,128])) conv_out4=tf.nn.conv2d(out3, filter4, strides=[1, 2, 2, 1], padding='SAME')