Fixes to TF collators (#21143)

* Add num_workers for prepare_tf_dataset

* Bugfix in the default collator and change default tensor type

* Remove the "num_workers" arg and move it to a new PR

src/transformers/data/data_collator.py

@@ -159,7 +159,7 @@ def tf_default_data_collator(features: List[InputDataClass]) -> Dict[str, Any]:
         label_col_name = None
     if label_col_name is not None:
         if isinstance(first[label_col_name], tf.Tensor):
-            dtype = tf.int64 if first[label_col_name].dtype.is_integer() else tf.float32
+            dtype = tf.int64 if first[label_col_name].dtype.is_integer else tf.float32
         elif isinstance(first[label_col_name], np.ndarray) or isinstance(first[label_col_name], np.generic):
             dtype = tf.int64 if np.issubdtype(first[label_col_name].dtype, np.integer) else tf.float32
         elif isinstance(first[label_col_name], (tuple, list)):

src/transformers/modeling_tf_utils.py

@@ -1345,9 +1345,9 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
 
         if collate_fn is None:
             if tokenizer is None:
-                collate_fn = DefaultDataCollator(return_tensors="tf")
+                collate_fn = DefaultDataCollator(return_tensors="np")
             else:
-                collate_fn = DataCollatorWithPadding(tokenizer=tokenizer, return_tensors="tf")
+                collate_fn = DataCollatorWithPadding(tokenizer=tokenizer, return_tensors="np")
         if collate_fn_args is None:
             collate_fn_args = dict()