Fixing feature encoder for tf.string variable length features.

PiperOrigin-RevId: 314208560

Author: eli7

tensor2tensor/data_generators/problem.py

@@ -718,10 +718,14 @@ def decode_example(self, serialized_example):
           [1], tf.int64, getattr(self._hparams, "sampling_keep_top_k", -1))
 
     if data_items_to_decoders is None:
-      data_items_to_decoders = {
-          field: contrib.slim().tfexample_decoder.Tensor(field)
-          for field in data_fields
-      }
+      data_items_to_decoders = {}
+      for field in data_fields:
+        if data_fields[field].dtype is tf.string:
+          default_value = b""
+        else:
+          default_value = 0
+        data_items_to_decoders[field] = contrib.slim().tfexample_decoder.Tensor(
+            field, default_value=default_value)
 
     decoder = contrib.slim().tfexample_decoder.TFExampleDecoder(
         data_fields, data_items_to_decoders)