def load_data(self):
print('Extracting features for train set:')
if os.path.exists(cfg.DATASET_PATH + 'features.pckl'):
with open(cfg.DATASET_PATH + 'features.pckl', 'rb') as f:
X = pickle.load(f)
with open(cfg.DATASET_PATH + 'targets.pckl', 'rb') as f:
y = pickle.load(f)
else:
X, y = extract_melgrams(cfg.LABELS_PATH)
pickle.dump(X, open(cfg.DATASET_PATH + 'features.pckl', "wb"), protocol=4)
pickle.dump(y, open(cfg.DATASET_PATH + 'targets.pckl', "wb"), protocol=4)
self.X_train, self.X_test, y_train, y_test = train_test_split(X, y, stratify=y, test_size=0.33, random_state=23)
print('X_train shape:', self.X_train.shape)
self.X_val, self.X_test, y_val, y_test = train_test_split(self.X_test, y_test, stratify=y_test, test_size=0.5, random_state=23)
print('X_val shape:', self.X_val.shape)
print('X_test shape:', self.X_test.shape)
self.Y_train = np_utils.to_categorical(y_train, self.nb_classes)
self.Y_val = np_utils.to_categorical(y_val, self.nb_classes)
self.Y_test = np_utils.to_categorical(y_test, self.nb_classes)
test_songs_list = 'list_example.txt'
# Initialize model
model = MusicTaggerCRNN(weights=None, input_tensor=(1, 96, 1366))
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
if LOAD_WEIGHTS:
model.load_weights(weights_path + 'custom_model_best.h5')
#model.summary()
X_test, num_frames_test = extract_melgrams(test_songs_list,
MULTIFRAMES,
process_all_song=False,
num_songs_genre='')
num_frames_test = np.array(num_frames_test)
t0 = time.time()
print('\n--------- Predicting ---------', '\n')
results = np.zeros((X_test.shape[0], tags.shape[0]))
predicted_labels_mean = np.zeros((num_frames_test.shape[0], 1))
predicted_labels_frames = np.zeros((X_test.shape[0], 1))
song_paths = open(test_songs_list, 'r').read().splitlines()
previous_numFrames = 0
train_gt_list = 'lists/train_gt_list.txt'
test_gt_list = 'lists/test_gt_list.txt'
# Data Loading
if LOAD_DB:
if MULTIFRAMES:
print 'Loading dataset multiframe...'
X_train, y_train, num_frames_train = load_dataset('')
X_test, y_test, num_frames_test = load_dataset('')
else:
X_train, X_test, y_train, y_test = load_dataset('')
# Compute mel-spectogram for all the frames
else:
X_train, y_train, num_frames_train = extract_melgrams(train_songs_list, MULTIFRAMES, process_all_song=False, num_songs_genre=20)
print('X_train shape:', X_train.shape)
X_test, y_test, num_frames_test = extract_melgrams(test_songs_list, MULTIFRAMES, process_all_song=False, num_songs_genre=10)
print(X_train.shape, 'train samples')
print(X_test.shape, 'test samples')
y_train = np.array(y_train)
y_test = np.array(y_test)
if SAVE_DB:
if MULTIFRAMES:
save_dataset('music_dataset/music_dataset_multiframe_train.h5', X_train, y_train,num_frames_train)
save_dataset('music_dataset/music_dataset_multiframe_test.h5', X_test,y_test,num_frames_test)
def run(path):
# Parameters to set
TEST = 1
LOAD_MODEL = 0
LOAD_WEIGHTS = 1
MULTIFRAMES = 1
time_elapsed = 0
# GTZAN Dataset Tags
tags = [
'blues', 'classical', 'country', 'disco', 'hiphop', 'jazz', 'metal',
'pop', 'reggae', 'rock'
]
tags = np.array(tags)
# Paths to set
model_name = "example_model"
model_path = "models_trained/" + model_name + "/"
weights_path = "models_trained/" + model_name + "/weights/"
test_songs_list = path
#if(vid_id == None):
# test_songs_list = "./music"
#else:
# test_songs_list = "./music2"
# Initialize model
model = MusicTaggerCRNN(weights=None, input_tensor=(1, 96, 1366))
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
if LOAD_WEIGHTS:
model.load_weights(weights_path + 'crnn_net_gru_adam_ours_epoch_40.h5')
#model.summary()
X_test, num_frames_test = extract_melgrams(test_songs_list,
MULTIFRAMES,
process_all_song=False,
num_songs_genre='')
num_frames_test = np.array(num_frames_test)
t0 = time.time()
print '\n--------- Predicting ---------', '\n'
results = np.zeros((X_test.shape[0], tags.shape[0]))
predicted_labels_mean = np.zeros((num_frames_test.shape[0], 1))
predicted_labels_frames = np.zeros((X_test.shape[0], 1))
song_paths = os.listdir(test_songs_list)
previous_numFrames = 0
n = 0
for i in range(0, num_frames_test.shape[0]):
print 'Song number' + str(i) + ': ' + song_paths[i]
num_frames = num_frames_test[i]
print 'Num_frames of 30s: ', str(num_frames), '\n'
results[previous_numFrames:previous_numFrames +
num_frames] = model.predict(
X_test[previous_numFrames:previous_numFrames +
num_frames, :, :, :])
s_counter = 0
for j in range(previous_numFrames, previous_numFrames + num_frames):
#normalize the results
total = results[j, :].sum()
results[j, :] = results[j, :] / total
print 'Percentage of genre prediction for seconds '+ str(20+s_counter*30) + ' to ' \
+ str(20+(s_counter+1)*30) + ': '
sort_result(tags, results[j, :].tolist())
predicted_label_frames = predict_label(results[j, :])
predicted_labels_frames[n] = predicted_label_frames
s_counter += 1
n += 1
print '\n', 'Mean genre of the song: '
results_song = results[previous_numFrames:previous_numFrames +
num_frames]
mean = results_song.mean(0)
sort_result(tags, mean.tolist())
predicted_label_mean = predict_label(mean)
predicted_labels_mean[i] = predicted_label_mean
print '\n', 'The predicted music genre for the song is', str(
tags[predicted_label_mean]), '!\n'
previous_numFrames = previous_numFrames + num_frames
print '************************************************************************************************'
# colors = ['b','g','c','r','m','k','y','#ff1122','#5511ff','#44ff22']
# fig, ax = plt.subplots()
# index = np.arange(tags.shape[0])
# opacity = 1
# bar_width = 0.2
# #print mean.tolist()
# #for g in range(0, tags.shape[0]):
# plt.bar(left=index, height=mean, width=bar_width, alpha=opacity, color=colors)
#
# plt.xlabel('Genres')
# plt.ylabel('Percentage')
# plt.title('Scores by genre')
# plt.xticks(index + bar_width / 2, tags)
# plt.tight_layout()
# fig.autofmt_xdate()
# plt.savefig('genres_prediction.png')
return tags[predicted_label_mean], song_paths[0], mean.tolist()
train_songs_list = 'lists/train_songs_list.txt'
test_songs_list = 'lists/test_songs_list.txt'
# Indicate the name of train and test songs
train_gt_list = 'lists/train_gt_list.txt'
test_gt_list = 'lists/test_gt_list.txt'
# Data Loading or computing the Mel-spectogram for each song
if LOAD:
X_train, y_train, num_frames_train = load_dataset(
'music_dataset/music_dataset_train.h5')
X_test, y_test, num_frames_test = load_dataset(
'music_dataset/music_dataset_test.h5')
else:
print('Computing melgrams for training dataset')
X_train, y_train, num_frames_train = extract_melgrams(
train_songs_list, process_all_song=False, num_songs_genre=70)
print('X_train shape:', X_train.shape)
print('Computing melgrams for testing dataset')
X_test, y_test, num_frames_test = extract_melgrams(test_songs_list,
process_all_song=False,
num_songs_genre=30)
print('X_train shape:', X_train.shape)
print(X_train.shape, 'train samples')
print(X_test.shape, 'test samples')
y_train = np.array(y_train)
y_test = np.array(y_test)
Y_train = np_utils.to_categorical(y_train, 10)
Y_test = np_utils.to_categorical(y_test, 10)