avoid running into end of sequence

Author: rvankoert

src/data/manager.py

@@ -519,7 +519,7 @@ def _create_dataset(self,
         dataset = dataset.prefetch(tf.data.AUTOTUNE)
 
         # Assert the cardinality of the dataset if training
-        if steps_per_epoch is None and is_training:
+        if steps_per_epoch is None:
             batches = int(np.ceil(len(self.raw_data[partition_name][0]) / self.config['batch_size']))
             dataset = dataset.apply(tf.data.experimental.assert_cardinality(batches))
 

src/utils/decoding.py

@@ -11,7 +11,7 @@
 from utils.text import Tokenizer
 
 
-@tf.function
+@tf.function(reduce_retracing=True)
 def ctc_decode(
     y_pred: np.ndarray,
     input_length: np.ndarray,

src/utils/threading.py

@@ -224,7 +224,8 @@ def _update_statistics(self, batch_counter: defaultdict, batch_size: int) -> Non
     def stop(self) -> None:
         """Stops the thread and waits for it to terminate."""
         self.running = False
-        self.join()
+        if self is not threading.current_thread():
+            self.join()
 
 
 class DecodingWorker(Thread):