faster augmentation

src/audio/augment.py

@@ -9,7 +9,8 @@ This package is published under GNU GPL Version 3.
 
 import nlpaug.flow as flow
 import nlpaug.augmenter.audio as naa
-
+from audio.utils import pad_with_silence
+from audio.features import normalize  
 
 class AudioAugmenter(object):
     def __init__(self, fs):
@@ -24,18 +25,21 @@ class AudioAugmenter(object):
         self._aug_flow = flow.Sequential([
             shift,
             crop,
-            vltp,
+            # vltp,
             # speed,
             # pitch,
             noise,
         ])
 
-    def augment_audio_array(self, signal, fs):
-        assert fs == self._fs
-        data = signal.astype(dtype="float32")
-        augmented_data = self._aug_flow.augment(data)
-        return augmented_data.astype(dtype="float32")
-
+    def augment_audio_array(self, signals, fs):
+        # assert fs == self._fs
+        # data = signal.astype(dtype="float32")
+        augmented_data = self._aug_flow.augment(signals, num_thread=8)
+        data = []
+        for x in augmented_data:
+            x = pad_with_silence(x, 5 *fs)
+            data.append(x)
+        return data
 
 if __name__ == "__main__":
 

src/audio/features.py

@@ -22,6 +22,8 @@ def normalize(signal):
     normalize a float signal to have a maximum absolute value of 1.0
     """
     maximum = max(abs(signal.max()), abs(signal.min()))
+    if maximum == 0.0:
+        return signal
     return signal / float(maximum)
 
 

src/config_gen.yaml

 
-languages: ["english","farsi","french","german","mandarin","spanish","polish","russian"]
+#languages: ["__noise","english","farsi","french","german","mandarin","spanish","polish","russian"]
+languages: ["__noise","english","french","german","spanish"]
 
 # Training
-train_dir: "/data/common_voice_filtered/fiveSec_validated/wav/train/"
-val_dir: "/data/common_voice_filtered/fiveSec_validated/wav/dev/"
+train_dir: "/data/common_voice_filtered/fiveSec/wav/train/"
+val_dir: "/data/common_voice_filtered/fiveSec/wav/test/"
 audio_length_s: 5
 sample_rate: 16000
 feature_nu: 80
-batch_size: 128
+batch_size: 64
 learning_rate: 0.001
 num_epochs: 20
 model: "attRnnSTFT"
-augment: False
+augment: True

src/models/attRnnSTFT.py

@@ -75,7 +75,9 @@ def create_model(config):
 	attVector = Dot(axes=[1, 1])([attScores, x])  # [b_s, vec_dim]
 
 	x = Dense(64, activation='relu')(attVector)
+	x = Dropout(0.25)(x)
 	x = Dense(32)(x)
+	x = Dropout(0.25)(x)
 
 	output = Dense(len(config["languages"]), activation='softmax', name='output')(x)
 

src/train_gen.py

@@ -31,14 +31,14 @@ def batch_gen(generator, batch_size, augmenter=None, fs=None, desired_audio_leng
 	while True:
 		try:
 			x,y = next(generator)
-			if augmenter:
-				x = augmenter.augment_audio_array(x, fs)
-				x = pad_with_silence(x, desired_audio_length_s *fs)
-			x = normalize(x)
 			x_batch.append(x)
 			y_batch.append(y)
 			i += 1
 			if i == batch_size:
+				if augmenter:
+					x_batch = augmenter.augment_audio_array(x_batch, fs)
+				for i in range(len(x_batch)):
+					x_batch[i] = normalize(x_batch[i])
 				x_arr = np.asarray(x_batch)
 				y_arr = np.asarray(y_batch)
 				yield x_arr, y_arr
@@ -148,7 +148,11 @@ def train(config_path, log_dir, model_path):
 		logs = {'train_acc': train_acc, 'train_loss': train_loss, 
 		 'val_acc': val_acc, 'val_loss': val_loss}
 
-		model.save(os.path.join(log_dir, 'model' + "_" + str(epoch + 1)))
+		@tf.function(input_signature=[tf.TensorSpec((1, audio_length_s*fs, 1), dtype=tf.float32)])
+		def model_predict(input_1):
+			return {'outputs': model(input_1, training=False)}
+
+		model.save(os.path.join(log_dir, 'model' + "_" + str(epoch + 1)), signatures={'serving_default': model_predict})
 		write_csv(os.path.join(log_dir, 'log.csv'), optimizer, epoch, logs)