[feature] normalize stft and remove complex dimension

convert_to_istft.py

@@ -7,6 +7,7 @@
 import sys
 import tensorflow_wav
 import numpy as np
+import math
 if(len(sys.argv)<2):
     print("You have to pick a file")
     no_file_Picked_Exception
@@ -22,9 +23,11 @@
     nframes = wav['nframes']
     print('shape', wav['data'].shape)
     time = 192
+    print(np.min(wav['data']), np.max(wav['data']), np.mean(wav['data']), np.std(wav['data']))
+    #wav['data'] = np.exp(wav['data'])
     data = istft(wav['data'],fs, time, hop)
-    print(wav)
-    wav['data']=data
-    print(np.min(data), np.max(data))
+    wav['data'] = data*3
+    #print(wav)
+    #wav['data'] = np.sign(wav['data'])*np.sqrt(wav['data'])
     res= tensorflow_wav.save_wav(wav, file+".istft")
     print(file+".istft"+" is written")

convert_to_stft.py

@@ -7,6 +7,7 @@
 import sys
 import tensorflow_wav
 import numpy as np
+import math
 if(len(sys.argv)<2):
     print("You have to pick a file")
     no_file_Picked_Exception
@@ -21,7 +22,9 @@
     wav= tensorflow_wav.get_wav(file)
     data = stft(wav['data'],fs, framesz, hop)
     print(wav)
-    wav['data']=data
-    print(np.min(data), np.max(data))
+    wav['data']=data.real
+    print(wav['data'])
+    #wav['data'] = np.sign(wav['data'])*np.power(wav['data'], 2)
+    print(np.min(wav['data']), np.max(wav['data']))
     res= tensorflow_wav.save_stft(wav, file+".stft")
     print(file+".stft"+" is written")

model.py

@@ -12,9 +12,9 @@
 BITRATE=4096
 class DCGAN(object):
     def __init__(self, sess, wav_size=WAV_SIZE, is_crop=True,
-                 batch_size=64, sample_size = 2, wav_shape=[WAV_SIZE, WAV_HEIGHT, 2],
+                 batch_size=64, sample_size = 2, wav_shape=[WAV_SIZE, WAV_HEIGHT, 1],
                  y_dim=None, z_dim=64, gf_dim=64, df_dim=64,
-                 gfc_dim=1024, dfc_dim=1024, c_dim=2, dataset_name='default',
+                 gfc_dim=1024, dfc_dim=1024, c_dim=1, dataset_name='default',
                  checkpoint_dir='checkpoint'):
         """
 
@@ -74,7 +74,7 @@ def build_model(self):
         if self.y_dim:
             self.y= tf.placeholder(tf.float32, [None, self.y_dim], name='y')
 
-        self.wavs = tf.placeholder(tf.complex64, [self.batch_size, BITRATE],
+        self.wavs = tf.placeholder(tf.float32, [self.batch_size, BITRATE],
                                     name='real_wavs')
 
         self.z = tf.placeholder(tf.float32, [None, self.z_dim],
@@ -91,7 +91,7 @@ def build_model(self):
         self.sampler = self.sampler(self.z)
         self.sampler = tf.reshape(self.sampler,[-1])
         #self.sampler = tensorflow_wav.decode(self.sampler)
-        encoded_G = tensorflow_wav.compose(self.G)#tensorflow_wav.encode(self.G)
+        encoded_G = self.G#tensorflow_wav.encode(self.G)
         self.D_ = self.discriminator(encoded_G, reuse=True)
 
 
@@ -202,14 +202,12 @@ def get_wav_content(files):
                     for repeat in range(errd_range):
                         #print("discrim ", errd_range)
                         # Update D network
+                        print("Running discriminator with min/max", batch_wavs.min(), batch_wavs.max())
                         _= self.sess.run([d_optim],
                             feed_dict={ self.wavs: batch_wavs, self.z: batch_z })
                         #self.writer.add_summary(summary_str, counter)
 
                     # Run g_optim twice to make sure that d_loss does not go to zero (different from paper)
-                    #if(errG > 8):
-                    #    errg_range = 2
-                    #else:
                     errg_range=1
                     for repeat in range(errg_range):
                         #print("generating ", errg_range)
@@ -227,8 +225,8 @@ def get_wav_content(files):
                         % (epoch, idx, batch_idxs,
                             time.time() - start_time, errD_fake, errD_real, errG))
 
-                    SAVE_COUNT=100
-                    SAMPLE_COUNT=20
+                    SAVE_COUNT=500
+                    SAMPLE_COUNT=100
 
                     print("Batch ", counter)
                     if np.mod(counter, SAVE_COUNT) == SAVE_COUNT-3:
@@ -309,7 +307,7 @@ def generator(self, z, y=None):
             print('h3',h3.get_shape())
 
             h4= deconv2d(h3,
-                    [self.batch_size, WAV_SIZE, WAV_HEIGHT, 2], name='g_h4', with_w=False, no_bias=False)
+                    [self.batch_size, WAV_SIZE, WAV_HEIGHT, 1], name='g_h4', with_w=False, no_bias=False)
 
             print('h4',h4.get_shape())
             tanh = tf.nn.tanh(h4)
@@ -338,7 +336,7 @@ def sampler(self, z, y=None):
             h3 = tf.nn.relu(self.g_bn3(h3, train=False))
             print('h3', h3.get_shape())
 
-            h4 = deconv2d(h3, [self.batch_size, 64, 64, 2], name='g_h4', no_bias=False)
+            h4 = deconv2d(h3, [self.batch_size, 64, 64, 1], name='g_h4', no_bias=False)
             print('h4', h4.get_shape())
 
             #tanh = tf.nn.tanh(h4)

tensorflow_wav.py

@@ -42,7 +42,6 @@ def save_wav(in_wav, path):
     wav.writeframes(processed)
 
 def save_stft(in_wav, path):
-    in_wav['data'] = in_wav['data']/(1e7+1e7j)
     f = open(path, "wb")
     try:
         pickle.dump(in_wav, f, pickle.HIGHEST_PROTOCOL)
@@ -52,16 +51,12 @@ def save_stft(in_wav, path):
 def get_stft(filename):
     f = open(filename, "rb")
     data = pickle.load(f)
-    data['data'] = data['data']*(1e7+1e7j)
     f.close()
     return data
 
 
-def decompose(input, rank=3):
-    real, imag = tf.split(rank, 2, input)
-    complex = tf.complex(real, imag)
-    return complex
 def compose(input, rank=3):
+    return input
     real = tf.real(input)
     imag = tf.imag(input)
     return tf.concat(rank, [real, imag])
@@ -74,6 +69,6 @@ def encode(input,bitrate=4096):
     return output
 
 def scale_up(input):
-    output = tf.nn.tanh(input)
-    return decompose(input)#*2e4
+    output = tf.nn.tanh(input)*(65535)
+    return output
 

test_stft_cpu.py

@@ -1,19 +1,27 @@
 import scipy, pylab
 import numpy as np
 import tensorflow_wav
+from math import sqrt
+
+def fft(x):
+    n = x.shape[0]
+    return scipy.fft(x)*1/sqrt(n)
 def stft(x, fs, framesz, hop):
     #print("STFT got", x, fs, framesz, hop)
     framesamp = int(framesz*fs)
     hopsamp = int(hop*fs)
     w = scipy.hanning(framesamp)
     def do_fft(w,x,i,framesamp):
         #print("Running FFT for ", i, framesamp)
-        return scipy.fft(w*x[i:i+framesamp])
+        return fft(w*x[i:i+framesamp])
     X = scipy.array([do_fft(w,x,i,framesamp) 
                      for i in range(0, len(x)-framesamp, hopsamp)])
     #print("X SHAPE IS", len(X), len(X[0]))
     return X
 
+def ifft(x):
+    n = x.shape[0]
+    return scipy.ifft(x)*1/sqrt(n)
 def istft(X, fs, T, hop):
     x = scipy.zeros(T*fs)
     framesamp = X.shape[1]
@@ -23,7 +31,7 @@ def istft(X, fs, T, hop):
         if(n>=X.shape[0]): 
             break
         #print("setting i to i+framesamp", n, i, framesamp, i+framesamp, len(X), len(x))
-        x[i:i+framesamp] += scipy.real(scipy.ifft(X[n]))
+        x[i:i+framesamp] += scipy.real(ifft(X[n]))
     #print(x.shape)
     return x