Revert frozen training arguments (#25903)

* Revert frozen training arguments

* TODO

examples/pytorch/image-pretraining/run_mae.py

@@ -163,15 +163,6 @@ class CustomTrainingArguments(TrainingArguments):
         default=1e-3, metadata={"help": "Base learning rate: absolute_lr = base_lr * total_batch_size / 256."}
     )
 
-    def __post_init__(self):
-        # Compute absolute learning rate while args are mutable
-        super().__post_init__()
-        if self.base_learning_rate is not None:
-            total_train_batch_size = self.train_batch_size * self.gradient_accumulation_steps * self.world_size
-            delattr(self, "_frozen")
-            self.learning_rate = self.base_learning_rate * total_train_batch_size / 256
-            setattr(self, "_frozen", True)
-
 
 def collate_fn(examples):
     pixel_values = torch.stack([example["pixel_values"] for example in examples])
@@ -362,6 +353,13 @@ def main():
         # Set the validation transforms
         ds["validation"].set_transform(preprocess_images)
 
+    # Compute absolute learning rate
+    total_train_batch_size = (
+        training_args.train_batch_size * training_args.gradient_accumulation_steps * training_args.world_size
+    )
+    if training_args.base_learning_rate is not None:
+        training_args.learning_rate = training_args.base_learning_rate * total_train_batch_size / 256
+
     # Initialize our trainer
     trainer = Trainer(
         model=model,

examples/pytorch/summarization/run_summarization.py

@@ -18,7 +18,6 @@ Fine-tuning the library models for sequence to sequence.
 """
 # You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.
 
-import dataclasses
 import logging
 import os
 import sys
@@ -675,10 +674,14 @@ def main():
         return result
 
     # Override the decoding parameters of Seq2SeqTrainer
-    if training_args.generation_max_length is None:
-        training_args = dataclasses.replace(training_args, generation_max_length=data_args.val_max_target_length)
-    if training_args.generation_num_beams is None:
-        training_args = dataclasses.replace(training_args, generation_num_beams=data_args.num_beams)
+    training_args.generation_max_length = (
+        training_args.generation_max_length
+        if training_args.generation_max_length is not None
+        else data_args.val_max_target_length
+    )
+    training_args.generation_num_beams = (
+        data_args.num_beams if data_args.num_beams is not None else training_args.generation_num_beams
+    )
 
     # Initialize our Trainer
     trainer = Seq2SeqTrainer(