[Trainer] Undo #29896 (#30129)

* Undo

* Use tokenizer

* Undo data collator

docs/source/en/tasks/image_classification.md

@@ -322,7 +322,7 @@ At this point, only three steps remain:
 ...     data_collator=data_collator,
 ...     train_dataset=food["train"],
 ...     eval_dataset=food["test"],
-...     image_processor=image_processor,
+...     tokenizer=image_processor,
 ...     compute_metrics=compute_metrics,
 ... )
 
@@ -418,7 +418,7 @@ and use the [PushToHubCallback](../main_classes/keras_callbacks#transformers.Pus
 >>> metric_callback = KerasMetricCallback(metric_fn=compute_metrics, eval_dataset=tf_eval_dataset)
 >>> push_to_hub_callback = PushToHubCallback(
 ...     output_dir="food_classifier",
-...     image_processor=image_processor,
+...     tokenizer=image_processor,
 ...     save_strategy="no",
 ... )
 >>> callbacks = [metric_callback, push_to_hub_callback]

docs/source/en/tasks/object_detection.md

@@ -384,7 +384,7 @@ Finally, bring everything together, and call [`~transformers.Trainer.train`]:
 ...     args=training_args,
 ...     data_collator=collate_fn,
 ...     train_dataset=cppe5["train"],
-...     image_processor=image_processor,
+...     tokenizer=image_processor,
 ... )
 
 >>> trainer.train()

docs/source/en/tasks/semantic_segmentation.md

@@ -642,7 +642,7 @@ and use the [`PushToHubCallback`] to upload the model:
 ...     metric_fn=compute_metrics, eval_dataset=tf_eval_dataset, batch_size=batch_size, label_cols=["labels"]
 ... )
 
->>> push_to_hub_callback = PushToHubCallback(output_dir="scene_segmentation", image_processor=image_processor)
+>>> push_to_hub_callback = PushToHubCallback(output_dir="scene_segmentation", tokenizer=image_processor)
 
 >>> callbacks = [metric_callback, push_to_hub_callback]
 ```

docs/source/en/tasks/video_classification.md

@@ -407,7 +407,7 @@ Then you just pass all of this along with the datasets to `Trainer`:
 ...     args,
 ...     train_dataset=train_dataset,
 ...     eval_dataset=val_dataset,
-...     image_processor=image_processor,
+...     tokenizer=image_processor,
 ...     compute_metrics=compute_metrics,
 ...     data_collator=collate_fn,
 ... )

docs/source/es/tasks/image_classification.md

@@ -160,7 +160,7 @@ Al llegar a este punto, solo quedan tres pasos:
 ...     data_collator=data_collator,
 ...     train_dataset=food["train"],
 ...     eval_dataset=food["test"],
-...     image_processor=image_processor,
+...     tokenizer=image_processor,
 ... )
 
 >>> trainer.train()

docs/source/ja/tasks/image_classification.md

@@ -328,7 +328,7 @@ food["test"].set_transform(preprocess_val)
 ...     data_collator=data_collator,
 ...     train_dataset=food["train"],
 ...     eval_dataset=food["test"],
-...     image_processor=image_processor,
+...     tokenizer=image_processor,
 ...     compute_metrics=compute_metrics,
 ... )
 
@@ -426,7 +426,7 @@ Convert your datasets to the `tf.data.Dataset` format using the [`~datasets.Data
 >>> metric_callback = KerasMetricCallback(metric_fn=compute_metrics, eval_dataset=tf_eval_dataset)
 >>> push_to_hub_callback = PushToHubCallback(
 ...     output_dir="food_classifier",
-...     image_processor=image_processor,
+...     tokenizer=image_processor,
 ...     save_strategy="no",
 ... )
 >>> callbacks = [metric_callback, push_to_hub_callback]

docs/source/ja/tasks/object_detection.md

@@ -376,7 +376,7 @@ DETR モデルをトレーニングできる「ラベル」。画像プロセッ
 ...     args=training_args,
 ...     data_collator=collate_fn,
 ...     train_dataset=cppe5["train"],
-...     image_processor=image_processor,
+...     tokenizer=image_processor,
 ... )
 
 >>> trainer.train()

docs/source/ja/tasks/semantic_segmentation.md

@@ -434,7 +434,7 @@ TensorFlow でモデルを微調整するには、次の手順に従います。
 ...     metric_fn=compute_metrics, eval_dataset=tf_eval_dataset, batch_size=batch_size, label_cols=["labels"]
 ... )
 
->>> push_to_hub_callback = PushToHubCallback(output_dir="scene_segmentation", image_processor=image_processor)
+>>> push_to_hub_callback = PushToHubCallback(output_dir="scene_segmentation", tokenizer=image_processor)
 
 >>> callbacks = [metric_callback, push_to_hub_callback]
 ```

docs/source/ja/tasks/video_classification.md

@@ -414,7 +414,7 @@ def compute_metrics(eval_pred):
 ...     args,
 ...     train_dataset=train_dataset,
 ...     eval_dataset=val_dataset,
-...     image_processor=image_processor,
+...     tokenizer=image_processor,
 ...     compute_metrics=compute_metrics,
 ...     data_collator=collate_fn,
 ... )

docs/source/ko/tasks/image_classification.md

@@ -321,7 +321,7 @@ food["test"].set_transform(preprocess_val)
 ...     data_collator=data_collator,
 ...     train_dataset=food["train"],
 ...     eval_dataset=food["test"],
-...     image_processor=image_processor,
+...     tokenizer=image_processor,
 ...     compute_metrics=compute_metrics,
 ... )
 
@@ -417,7 +417,7 @@ TensorFlow에서 모델을 미세 조정하려면 다음 단계를 따르세요:
 >>> metric_callback = KerasMetricCallback(metric_fn=compute_metrics, eval_dataset=tf_eval_dataset)
 >>> push_to_hub_callback = PushToHubCallback(
 ...     output_dir="food_classifier",
-...     image_processor=image_processor,
+...     tokenizer=image_processor,
 ...     save_strategy="no",
 ... )
 >>> callbacks = [metric_callback, push_to_hub_callback]

docs/source/ko/tasks/object_detection.md

@@ -366,7 +366,7 @@ DatasetDict({
 ...     args=training_args,
 ...     data_collator=collate_fn,
 ...     train_dataset=cppe5["train"],
-...     image_processor=image_processor,
+...     tokenizer=image_processor,
 ... )
 
 >>> trainer.train()

docs/source/ko/tasks/semantic_segmentation.md

@@ -424,7 +424,7 @@ TensorFlow에서 모델을 미세 조정하려면 다음 단계를 따르세요:
 ...     metric_fn=compute_metrics, eval_dataset=tf_eval_dataset, batch_size=batch_size, label_cols=["labels"]
 ... )
 
->>> push_to_hub_callback = PushToHubCallback(output_dir="scene_segmentation", image_processor=image_processor)
+>>> push_to_hub_callback = PushToHubCallback(output_dir="scene_segmentation", tokenizer=image_processor)
 
 >>> callbacks = [metric_callback, push_to_hub_callback]
 ```

docs/source/ko/tasks/video_classification.md

@@ -411,7 +411,7 @@ def compute_metrics(eval_pred):
 ...     args,
 ...     train_dataset=train_dataset,
 ...     eval_dataset=val_dataset,
-...     image_processor=image_processor,
+...     tokenizer=image_processor,
 ...     compute_metrics=compute_metrics,
 ...     data_collator=collate_fn,
 ... )

examples/pytorch/image-classification/run_image_classification.py

@@ -411,7 +411,7 @@ def val_transforms(example_batch):
         train_dataset=dataset["train"] if training_args.do_train else None,
         eval_dataset=dataset["validation"] if training_args.do_eval else None,
         compute_metrics=compute_metrics,
-        image_processor=image_processor,
+        tokenizer=image_processor,
         data_collator=collate_fn,
     )
 

examples/pytorch/image-pretraining/run_mae.py

@@ -369,7 +369,7 @@ def preprocess_images(examples):
         args=training_args,
         train_dataset=ds["train"] if training_args.do_train else None,
         eval_dataset=ds["validation"] if training_args.do_eval else None,
-        image_processor=image_processor,
+        tokenizer=image_processor,
         data_collator=collate_fn,
     )
 

examples/pytorch/image-pretraining/run_mim.py

@@ -458,7 +458,7 @@ def preprocess_images(examples):
         args=training_args,
         train_dataset=ds["train"] if training_args.do_train else None,
         eval_dataset=ds["validation"] if training_args.do_eval else None,
-        image_processor=image_processor,
+        tokenizer=image_processor,
         data_collator=collate_fn,
     )
 

examples/pytorch/semantic-segmentation/run_semantic_segmentation.py

@@ -510,7 +510,7 @@ def preprocess_val(example_batch):
         train_dataset=dataset["train"] if training_args.do_train else None,
         eval_dataset=dataset["validation"] if training_args.do_eval else None,
         compute_metrics=compute_metrics,
-        image_processor=image_processor,
+        tokenizer=image_processor,
         data_collator=default_data_collator,
     )
 

examples/tensorflow/image-classification/run_image_classification.py

@@ -552,7 +552,7 @@ def compute_metrics(p):
                     output_dir=training_args.output_dir,
                     hub_model_id=push_to_hub_model_id,
                     hub_token=training_args.push_to_hub_token,
-                    image_processor=image_processor,
+                    tokenizer=image_processor,
                     **model_card_kwargs,
                 )
             )

src/transformers/trainer.py

@@ -60,7 +60,6 @@
 from .debug_utils import DebugOption, DebugUnderflowOverflow
 from .feature_extraction_sequence_utils import SequenceFeatureExtractor
 from .hyperparameter_search import ALL_HYPERPARAMETER_SEARCH_BACKENDS, default_hp_search_backend
-from .image_processing_utils import BaseImageProcessor
 from .integrations.deepspeed import deepspeed_init, deepspeed_load_checkpoint, is_deepspeed_available
 from .integrations.tpu import tpu_spmd_dataloader
 from .modelcard import TrainingSummary
@@ -331,9 +330,6 @@ class Trainer:
             by this function will be reflected in the predictions received by `compute_metrics`.
 
             Note that the labels (second parameter) will be `None` if the dataset does not have them.
-        image_processor ([`BaseImageProcessor`], *optional*):
-            The image processor used to preprocess the data. If provided, it will be saved along the model to make it easier
-            to rerun an interrupted training or reuse the fine-tuned model.
 
     Important attributes:
 
@@ -369,7 +365,6 @@ def __init__(
         callbacks: Optional[List[TrainerCallback]] = None,
         optimizers: Tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR] = (None, None),
         preprocess_logits_for_metrics: Optional[Callable[[torch.Tensor, torch.Tensor], torch.Tensor]] = None,
-        image_processor: Optional["BaseImageProcessor"] = None,
     ):
         if args is None:
             output_dir = "tmp_trainer"
@@ -502,7 +497,6 @@ def __init__(
         self.train_dataset = train_dataset
         self.eval_dataset = eval_dataset
         self.tokenizer = tokenizer
-        self.image_processor = image_processor
 
         # Bnb Quantized models doesn't support `.to` operation.
         if (
@@ -554,7 +548,7 @@ def __init__(
         default_callbacks = DEFAULT_CALLBACKS + get_reporting_integration_callbacks(self.args.report_to)
         callbacks = default_callbacks if callbacks is None else default_callbacks + callbacks
         self.callback_handler = CallbackHandler(
-            callbacks, self.model, self.tokenizer, self.image_processor, self.optimizer, self.lr_scheduler
+            callbacks, self.model, self.tokenizer, self.optimizer, self.lr_scheduler
         )
         self.add_callback(PrinterCallback if self.args.disable_tqdm else DEFAULT_PROGRESS_CALLBACK)
 
@@ -3289,8 +3283,6 @@ def _save_tpu(self, output_dir: Optional[str] = None):
             )
         if self.tokenizer is not None and self.args.should_save:
             self.tokenizer.save_pretrained(output_dir)
-        if self.image_processor is not None and self.args.should_save:
-            self.image_processor.save_pretrained(output_dir)
 
         # We moved the model from TPU -> CPU for saving the weights.
         # Now we should move it back to subsequent compute still works.
@@ -3328,8 +3320,6 @@ def _save(self, output_dir: Optional[str] = None, state_dict=None):
 
         if self.tokenizer is not None:
             self.tokenizer.save_pretrained(output_dir)
-        if self.image_processor is not None:
-            self.image_processor.save_pretrained(output_dir)
 
         # Good practice: save your training arguments together with the trained model
         torch.save(self.args, os.path.join(output_dir, TRAINING_ARGS_NAME))
@@ -4027,9 +4017,6 @@ def _push_from_checkpoint(self, checkpoint_folder):
         # Saving the tokenizer is fast and we don't know how many files it may have spawned, so we resave it to be sure.
         if self.tokenizer is not None:
             self.tokenizer.save_pretrained(output_dir)
-        # Same for the image processor
-        if self.image_processor is not None:
-            self.image_processor.save_pretrained(output_dir)
         # Same for the training arguments
         torch.save(self.args, os.path.join(output_dir, TRAINING_ARGS_NAME))
 
@@ -4083,7 +4070,7 @@ def push_to_hub(
         **kwargs,
     ) -> str:
         """
-        Upload `self.model` and `self.tokenizer` or `self.image_processor` to the 🤗 model hub on the repo `self.args.hub_model_id`.
+        Upload `self.model` and `self.tokenizer` to the 🤗 model hub on the repo `self.args.hub_model_id`.
 
         Parameters:
             commit_message (`str`, *optional*, defaults to `"End of training"`):

src/transformers/trainer_callback.py

@@ -189,8 +189,6 @@ class TrainerCallback:
             The model being trained.
         tokenizer ([`PreTrainedTokenizer`]):
             The tokenizer used for encoding the data.
-        image_processor ([`BaseImageProcessor`]):
-            The image processor used for encoding the images.
         optimizer (`torch.optim.Optimizer`):
             The optimizer used for the training steps.
         lr_scheduler (`torch.optim.lr_scheduler.LambdaLR`):
@@ -309,13 +307,12 @@ def on_prediction_step(self, args: TrainingArguments, state: TrainerState, contr
 class CallbackHandler(TrainerCallback):
     """Internal class that just calls the list of callbacks in order."""
 
-    def __init__(self, callbacks, model, tokenizer, image_processor, optimizer, lr_scheduler):
+    def __init__(self, callbacks, model, tokenizer, optimizer, lr_scheduler):
         self.callbacks = []
         for cb in callbacks:
             self.add_callback(cb)
         self.model = model
         self.tokenizer = tokenizer
-        self.image_processor = image_processor
         self.optimizer = optimizer
         self.lr_scheduler = lr_scheduler
         self.train_dataloader = None
@@ -420,7 +417,6 @@ def call_event(self, event, args, state, control, **kwargs):
                 control,
                 model=self.model,
                 tokenizer=self.tokenizer,
-                image_processor=self.image_processor,
                 optimizer=self.optimizer,
                 lr_scheduler=self.lr_scheduler,
                 train_dataloader=self.train_dataloader,