Update feature selection in to_tf_dataset (#21935)

* Update feature selection

* Check compatibility with datasets version

* Checkout from datasets main

Author: amyeroberts

docs/source/en/tasks/image_classification.mdx

@@ -385,12 +385,12 @@ Convert your datasets to the `tf.data.Dataset` format using the [`~datasets.Data
 ```py
 >>> # converting our train dataset to tf.data.Dataset
 >>> tf_train_dataset = food["train"].to_tf_dataset(
-...     columns=["pixel_values"], label_cols=["label"], shuffle=True, batch_size=batch_size, collate_fn=data_collator
+...     columns="pixel_values", label_cols="label", shuffle=True, batch_size=batch_size, collate_fn=data_collator
 ... )
 
 >>> # converting our test dataset to tf.data.Dataset
 >>> tf_eval_dataset = food["test"].to_tf_dataset(
-...     columns=["pixel_values"], label_cols=["label"], shuffle=True, batch_size=batch_size, collate_fn=data_collator
+...     columns="pixel_values", label_cols="label", shuffle=True, batch_size=batch_size, collate_fn=data_collator
 ... )
 ```
 

docs/source/es/training.mdx

@@ -173,15 +173,15 @@ A continuación, convierte los datasets tokenizados en datasets de TensorFlow co
 ```py
 >>> tf_train_dataset = small_train_dataset.to_tf_dataset(
 ...     columns=["attention_mask", "input_ids", "token_type_ids"],
-...     label_cols=["labels"],
+...     label_cols="labels",
 ...     shuffle=True,
 ...     collate_fn=data_collator,
 ...     batch_size=8,
 ... )
 
 >>> tf_validation_dataset = small_eval_dataset.to_tf_dataset(
 ...     columns=["attention_mask", "input_ids", "token_type_ids"],
-...     label_cols=["labels"],
+...     label_cols="labels",
 ...     shuffle=False,
 ...     collate_fn=data_collator,
 ...     batch_size=8,

docs/source/pt/training.mdx

@@ -205,15 +205,15 @@ Especifique suas entradas em `columns` e seu rótulo em `label_cols`:
 ```py
 >>> tf_train_dataset = small_train_dataset.to_tf_dataset(
 ...     columns=["attention_mask", "input_ids", "token_type_ids"],
-...     label_cols=["labels"],
+...     label_cols="labels",
 ...     shuffle=True,
 ...     collate_fn=data_collator,
 ...     batch_size=8,
 ... )
 
 >>> tf_validation_dataset = small_eval_dataset.to_tf_dataset(
 ...     columns=["attention_mask", "input_ids", "token_type_ids"],
-...     label_cols=["labels"],
+...     label_cols="labels",
 ...     shuffle=False,
 ...     collate_fn=data_collator,
 ...     batch_size=8,

src/transformers/modeling_tf_utils.py

@@ -1413,6 +1413,12 @@ def prepare_tf_dataset(
         feature_cols = [col for col in output_columns if col in model_inputs and col not in model_labels]
         label_cols = [col for col in output_columns if col in model_labels]
 
+        # Backwards compatibility for older versions of datasets. Previously, if `columns` or `label_cols`
+        # were a single element list, the returned element spec would be a single element. Now, passing [feature]
+        # will return a dict structure {"feature": feature}, and passing a single string will return a single element.
+        feature_cols = feature_cols[0] if len(feature_cols) == 1 else feature_cols
+        label_cols = label_cols[0] if len(label_cols) == 1 else label_cols
+
         if drop_remainder is None:
             drop_remainder = shuffle
         tf_dataset = dataset.to_tf_dataset(