✍️ update training script

Author: nglehuy

examples/conformer/train_tpu_keras_subword_conformer.py

@@ -97,7 +97,7 @@
     global_batch_size *= strategy.num_replicas_in_sync
     # build model
     conformer = Conformer(**config.model_config, vocabulary_size=text_featurizer.num_classes)
-    conformer._build(speech_featurizer.shape, prediction_shape=text_featurizer.prepand_shape, batch_size=batch_size)
+    conformer._build(speech_featurizer.shape, prediction_shape=text_featurizer.prepand_shape, batch_size=global_batch_size)
     conformer.summary(line_length=120)
 
     optimizer = tf.keras.optimizers.Adam(

tensorflow_asr/featurizers/speech_featurizers.py

@@ -27,6 +27,14 @@
 tpu = has_tpu()
 
 
+# def tf_resample(signal, rate_in, rate_out):
+#     if rate_in == rate_out: return signal
+#     rate_in = tf.cast(rate_in, dtype=tf.float32)
+#     rate_out = tf.cast(rate_out, dtype=tf.float32)
+#     ratio = rate_out / rate_in
+#     nsamples = tf.math.ceil(tf.shape(signal)[0] * ratio)
+
+
 def load_and_convert_to_wav(path: str) -> tf.Tensor:
     wave, rate = librosa.load(os.path.expanduser(path), sr=None, mono=True)
     return tf.audio.encode_wav(tf.expand_dims(wave, axis=-1), sample_rate=rate)