Changed dataset handling

src/__init__.py

+from utils import trainer
+from utils import data_loader
+from utils import evaluation
+from utils import summary_utils

src/models/discriminator.py

@@ -36,7 +36,7 @@ class Discriminator(tf.keras.Model):
                                                  kernel_initializer = self.kernel_initializer,
                                                  use_bias = False,
                                                  name = self.model_name +"_dense",
-                                                 activation = tf.nn.relu))
+                                                 activation = None))
     
     
     def call(self,input,training=False):

src/models/eval_functions/separate_classifier_wgan_eval_fns.py

@@ -46,7 +46,7 @@ class EvalFunctions(object):
                 axis=-1,
             ))
 
-        gen_loss = tf.reduce_mean((fake_features - x[1]) ** 2)
+        gen_loss = 0#tf.reduce_mean((fake_features - x[1]) ** 2)
         # Wasserstein losses
         gen_loss += tf.reduce_mean(false)
         discr_loss = tf.reduce_mean(true) - tf.reduce_mean(false)

src/models/generator.py

@@ -66,4 +66,4 @@ class Generator(tf.keras.Model):
             elif layer[1]==1:
                 x = layer[0](x,training)
         x = x[:,:input.shape[1],:input.shape[2],:]
-        return x+input
+        return x#+input

src/scripts/birdsong_simple_main.py

@@ -18,7 +18,7 @@ import tensorflow as tf  # pylint: disable=g-bad-import-order
 
 BINS = 1025
 N_FRAMES = 216
-N_CHANNELS = 2
+N_CHANNELS = 1
 
 
 def augment_input(sample,n_classes,training):
@@ -224,7 +224,6 @@ def main(argv):
                     save_dir = model_save_dir,
                     input_keys = ["input_features","false_sample"],
                     label_keys = ["labels"],
-                    init_data = tf.random.normal([batch_size,BINS,N_FRAMES,N_CHANNELS]),
                     start_epoch = 0)
     
     trainer.train()

src/scripts/eval_classifier.py

@@ -136,7 +136,7 @@ def main(argv):
                  init_ch = 64,
                  init_ksize = 7,
                  init_stride = 2,
-                 use_max_pool = True,
+                 use_max_pool = True,cd 
                  kernel_regularizer = tf.keras.regularizers.l2(2e-4),
                  kernel_initializer = tf.keras.initializers.he_normal(),
                  name = "classifier",

src/utils/data_loader.py

@@ -144,6 +144,7 @@ class DataGenerator(object):
                  preload_samples = False,
                  training_percentage = 90,
                  save_created_features = True,
+                 replicate_to_stereo = False,
                  force_mono = True,
                  max_time = 5,
                  max_samples_per_audio = 6,
@@ -167,6 +168,7 @@ class DataGenerator(object):
         self.preload_false_samples = preload_false_samples
         self.hop_length = hop_length
         self.max_time = max_time
+        self.replicate_to_stereo = replicate_to_stereo
         self.max_samples_per_audio = max_samples_per_audio
         self.force_feature_recalc = force_feature_recalc
         self.save_created_features = save_created_features
@@ -274,7 +276,6 @@ class DataGenerator(object):
         """Creates the features by doing a STFT"""
         
         filename = sample['filename']
-        #filename = str(sample['audio_id']) +  ".mp3"
         if "channels" in sample.keys() and not(self.force_mono):
             try:
                 channels_str = sample['channels']
@@ -356,6 +357,7 @@ class DataGenerator(object):
         n = len(samples)
         
         pool = multiprocessing.Pool(os.cpu_count())
+        print(n)
         for i, _ in enumerate(pool.imap_unordered(self.create_feature, samples), 1):
             sys.stderr.write('\rdone {0:%}'.format(max(0,i/n)))
         
@@ -385,6 +387,7 @@ class DataGenerator(object):
 
         pool = multiprocessing.Pool(os.cpu_count())
         for i, _ in enumerate(pool.imap_unordered(self.create_feature, samples), 1):
+            print(i)
             sys.stderr.write('\rdone {0:%}'.format(max(0,i/n)))
             
     def generate_sample_for_file_and_time(self):
@@ -396,9 +399,10 @@ class DataGenerator(object):
             if site == "site_3":
                 spectra = self.create_feature(sample)
                 for key in spectra.keys():
-                    #If only mono --> duplicate
-                    if spectra[key].shape[0] == 1:
-                        spectrum = np.tile(spectra[key],[2,1,1])
+                    if self.replicate_to_stereo:
+                        #If only mono --> duplicate
+                        if spectra[key].shape[0] == 1:
+                            spectrum = np.tile(spectra[key],[2,1,1])
 
                     #Transpose spectrogramms for "channels_last"
                     spectrum = tf.transpose(spectrum,perm=[1,2,0])
@@ -408,9 +412,10 @@ class DataGenerator(object):
                 start = int(float(seconds))-5
                 end = int(float(seconds))
                 spectrum = self.create_feature_for_time(filename,start,end,mono=True)
-                #If only mono --> duplicate
-                if spectrum.shape[0] == 1:
-                    spectrum = np.tile(spectrum,[2,1,1])
+                if self.replicate_to_stereo:
+                    #If only mono --> duplicate
+                    if spectrum.shape[0] == 1:
+                        spectrum = np.tile(spectrum,[2,1,1])
 
                 #Transpose spectrogramms for "channels_last"
                 spectrum = tf.transpose(spectrum,perm=[1,2,0])
@@ -433,10 +438,11 @@ class DataGenerator(object):
 
                 if np.any(spectrum) == None or spectrum.shape[-1] != stft_len:
                     continue
-            
-                #If only mono --> duplicate
-                if spectrum.shape[0] == 1:
-                    spectrum = np.tile(spectra[spec_key],[2,1,1])
+                
+                if self.replicate_to_stereo:
+                    #If only mono --> duplicate
+                    if spectrum.shape[0] == 1:
+                        spectrum = np.tile(spectra[spec_key],[2,1,1])
                 
                 #Transpose spectrogramms for "channels_last"
                 spectrum = tf.transpose(spectrum,perm=[1,2,0])
@@ -509,20 +515,22 @@ class DataGenerator(object):
             false_rnd_key = false_spec_keys[np.random.randint(0,len(false_spec_keys))]
             false_spectra = false_spectra_npz.item()[false_rnd_key]
 
-            #If only mono --> duplicate
-            if spectra.shape[0] == 1:
-                spectra = np.tile(spectra,[2,1,1])
-                
+          
             
-            #If false only mono --> duplicate
-            if false_spectra.shape[0] == 1:
-                false_spectra = np.tile(false_spectra,[2,1,1])
+            if self.replicate_to_stereo:
+                #If only mono --> duplicate
+                if spectra.shape[0] == 1:
+                    spectra = np.tile(spectra,[2,1,1])
+                    
+                #If false only mono --> duplicate
+                if false_spectra.shape[0] == 1:
+                    false_spectra = np.tile(false_spectra,[2,1,1])
 
             #Transpose spectrogramms for "channels_last"
             spectra = tf.transpose(spectra,perm=[1,2,0])
             false_spectra = tf.transpose(false_spectra,perm=[1,2,0])
             
-            if self.is_training:
+            if self.is_training or self.is_validation:
                 sample = {'input_features':spectra,
                     'labels':tf.one_hot(sample['bird_id'],self.dataset.n_classes+1),
                     'false_sample':false_spectra}
@@ -535,5 +543,9 @@ class DataGenerator(object):
 
 if __name__ == "__main__":
     ds = Dataset("/srv/TUG/datasets/cornell_birdcall_recognition",is_training_set=True)
-    dg = DataGenerator(ds,None,force_feature_recalc=True,is_training=True,preload_samples=False)
-    dg.create_all_features()
+    dg = DataGenerator(ds,None,
+                       force_feature_recalc=True,
+                       is_training=True,
+                       preload_samples=False,
+                       preload_false_samples=False)
+    dg.create_false_features_multi_cpu()