Switch backend to TensorFlow

Author: koz4k

common.py

@@ -4,7 +4,7 @@
 
 import numpy as np
 import librosa as lbr
-import keras.backend as K
+import tensorflow.keras.backend as K
 
 GENRES = ['blues', 'classical', 'country', 'disco', 'hiphop', 'jazz', 'metal',
         'pop', 'reggae', 'rock']
@@ -20,8 +20,8 @@
 def get_layer_output_function(model, layer_name):
     input = model.get_layer('input').input
     output = model.get_layer(layer_name).output
-    f = K.function([input, K.learning_phase()], output)
-    return lambda x: f([x, 0]) # learning_phase = 0 means test
+    f = K.function([input, K.learning_phase()], [output])
+    return lambda x: f([x, 0])[0] # learning_phase = 0 means test
 
 def load_track(filename, enforce_shape=None):
     new_input, sample_rate = lbr.load(filename, mono=True)

create_data_pickle.py

@@ -2,7 +2,7 @@
 import sys
 import numpy as np
 from math import pi
-from cPickle import dump
+from pickle import dump
 import os
 from optparse import OptionParser
 
@@ -30,10 +30,10 @@ def collect_data(dataset_path):
     track_paths = {}
 
     for (genre_index, genre_name) in enumerate(GENRES):
-        for i in xrange(TRACK_COUNT // len(GENRES)):
+        for i in range(TRACK_COUNT // len(GENRES)):
             file_name = '{}/{}.000{}.au'.format(genre_name,
                     genre_name, str(i).zfill(2))
-            print 'Processing', file_name
+            print('Processing', file_name)
             path = os.path.join(dataset_path, file_name)
             track_index = genre_index  * (TRACK_COUNT // len(GENRES)) + i
             x[track_index], _ = load_track(path, default_shape)
@@ -55,5 +55,5 @@ def collect_data(dataset_path):
     (x, y, track_paths) = collect_data(options.dataset_path)
 
     data = {'x': x, 'y': y, 'track_paths': track_paths}
-    with open(options.output_pkl_path, 'w') as f:
+    with open(options.output_pkl_path, 'wb') as f:
         dump(data, f)

extract_filters.py

@@ -1,5 +1,5 @@
 from common import get_layer_output_function, WINDOW_SIZE, WINDOW_STRIDE
-from keras.models import model_from_yaml
+from tensorflow.keras.models import model_from_yaml
 import librosa as lbr
 import numpy as np
 from functools import partial
@@ -56,7 +56,7 @@ def extract_filters(model, data, filters_path, count0):
         if not os.path.exists(layer_path):
             os.makedirs(layer_path)
 
-        print 'Computing outputs for layer', conv_layer_names[layer_index]
+        print('Computing outputs for layer', conv_layer_names[layer_index])
         output = output_fun(x)
 
         # matrices of shape n_tracks x time x n_filters
@@ -70,9 +70,9 @@ def extract_filters(model, data, filters_path, count0):
 
         undoer = conv_layer_undoers[layer_index]
 
-        for filter_index in xrange(argmax_over_track.shape[1]):
-            print 'Processing layer', conv_layer_names[layer_index], \
-                    'filter', filter_index
+        for filter_index in range(argmax_over_track.shape[1]):
+            print('Processing layer', conv_layer_names[layer_index], \
+                    'filter', filter_index)
 
             track_indices = argmax_over_track[:, filter_index]
             time_indices = argmax_over_time[track_indices, filter_index]

genre_recognizer.py

@@ -1,8 +1,8 @@
 from common import load_track, get_layer_output_function
 import numpy as np
-from keras.layers import Input
-from keras.models import model_from_yaml, Model
-from keras import backend as K
+from tensorflow.keras.layers import Input
+from tensorflow.keras.models import model_from_yaml, Model
+from tensorflow.keras import backend as K
 
 class GenreRecognizer():
 
@@ -11,10 +11,10 @@ def __init__(self, model_path, weights_path):
             model = model_from_yaml(f.read())
         model.load_weights(weights_path)
         self.pred_fun = get_layer_output_function(model, 'output_realtime')
-        print 'Loaded model.'
+        print('Loaded model.')
 
     def recognize(self, track_path):
-        print 'Loading song', track_path
+        print('Loading song', track_path)
         (features, duration) = load_track(track_path)
         features = np.reshape(features, (1,) + features.shape)
         return (self.pred_fun(features), duration)