🚨🚨🚨Deprecate evaluation_strategy to eval_strategy🚨🚨🚨 (#30190)

* Alias

* Note alias

* Tests and src

* Rest

* Clean

* Change typing?

* Fix tests

* Deprecation versions

examples/research_projects/self-training-text-classification/README.md

@@ -51,7 +51,7 @@ parameters_dict = {
     'train_file': os.path.join(data_dir, 'train.csv'),
     'infer_file': os.path.join(data_dir, 'infer.csv'),
     'eval_file': os.path.join(data_dir, 'eval.csv'),
-    'evaluation_strategy': 'steps',
+    'eval_strategy': 'steps',
     'task_name': 'scitail',
     'label_list': ['entails', 'neutral'],
     'per_device_train_batch_size': 32,

examples/research_projects/self-training-text-classification/finetuning.py

@@ -190,7 +190,7 @@ class FTTrainingArguments:
             )
         },
     )
-    evaluation_strategy: Optional[str] = dataclasses.field(
+    eval_strategy: Optional[str] = dataclasses.field(
         default="no",
         metadata={
             "help": 'The evaluation strategy to adopt during training. Possible values are: ["no", "step", "epoch]'
@@ -198,7 +198,7 @@ class FTTrainingArguments:
     )
     eval_steps: Optional[int] = dataclasses.field(
         default=1,
-        metadata={"help": 'Number of update steps between two evaluations if `evaluation_strategy="steps"`.'},
+        metadata={"help": 'Number of update steps between two evaluations if `eval_strategy="steps"`.'},
     )
     eval_metric: Optional[str] = dataclasses.field(
         default="accuracy", metadata={"help": "The evaluation metric used for the task."}
@@ -265,7 +265,7 @@ def train(args, accelerator, model, tokenizer, train_dataloader, optimizer, lr_s
                 # Evaluate during training
                 if (
                     eval_dataloader is not None
-                    and args.evaluation_strategy == IntervalStrategy.STEPS.value
+                    and args.eval_strategy == IntervalStrategy.STEPS.value
                     and args.eval_steps > 0
                     and completed_steps % args.eval_steps == 0
                 ):
@@ -331,7 +331,7 @@ def train(args, accelerator, model, tokenizer, train_dataloader, optimizer, lr_s
                 break
 
         # Evaluate during training
-        if eval_dataloader is not None and args.evaluation_strategy == IntervalStrategy.EPOCH.value:
+        if eval_dataloader is not None and args.eval_strategy == IntervalStrategy.EPOCH.value:
             accelerator.wait_for_everyone()
             new_checkpoint = f"checkpoint-{IntervalStrategy.EPOCH.value}-{epoch}"
             new_eval_result = evaluate(args, accelerator, eval_dataloader, "eval", model, new_checkpoint)[
@@ -571,7 +571,7 @@ def finetune(accelerator, model_name_or_path, train_file, output_dir, **kwargs):
     assert args.train_file is not None
     data_files[Split.TRAIN.value] = args.train_file
 
-    if args.do_eval or args.evaluation_strategy != IntervalStrategy.NO.value:
+    if args.do_eval or args.eval_strategy != IntervalStrategy.NO.value:
         assert args.eval_file is not None
         data_files[Split.EVAL.value] = args.eval_file
 

examples/research_projects/self-training-text-classification/run.sh

@@ -60,7 +60,7 @@ parameters_dict = {
   'train_file': os.path.join(data_dir, '${TRAIN_FILE}'),
   'infer_file': os.path.join(data_dir, '${INFER_FILE}'),
   'eval_file': os.path.join(data_dir, '${EVAL_FILE}'),
-  'evaluation_strategy': 'steps',
+  'eval_strategy': 'steps',
   'task_name': 'scitail',
   'label_list': ['entails', 'neutral'],
   'per_device_train_batch_size': 32,

examples/research_projects/self-training-text-classification/selftraining.py

@@ -79,7 +79,7 @@ class STTrainingArguments:
     eval_metric: Optional[str] = dataclasses.field(
         default="accuracy", metadata={"help": "The evaluation metric used for the task."}
     )
-    evaluation_strategy: Optional[str] = dataclasses.field(
+    eval_strategy: Optional[str] = dataclasses.field(
         default="no",
         metadata={
             "help": 'The evaluation strategy to adopt during training. Possible values are: ["no", "step", "epoch]'
@@ -208,7 +208,7 @@ def selftrain(model_name_or_path, train_file, infer_file, output_dir, **kwargs):
     data_files["train"] = args.train_file
     data_files["infer"] = args.infer_file
 
-    if args.evaluation_strategy != IntervalStrategy.NO.value:
+    if args.eval_strategy != IntervalStrategy.NO.value:
         assert args.eval_file is not None
         data_files["eval"] = args.eval_file
 
@@ -267,7 +267,7 @@ def selftrain(model_name_or_path, train_file, infer_file, output_dir, **kwargs):
             "label_list": args.label_list,
             "output_dir": current_output_dir,
             "eval_metric": args.eval_metric,
-            "evaluation_strategy": args.evaluation_strategy,
+            "eval_strategy": args.eval_strategy,
             "early_stopping_patience": args.early_stopping_patience,
             "early_stopping_threshold": args.early_stopping_threshold,
             "seed": args.seed,
@@ -341,7 +341,7 @@ def selftrain(model_name_or_path, train_file, infer_file, output_dir, **kwargs):
 
         data_files["train_pseudo"] = os.path.join(next_data_dir, f"train_pseudo.{args.data_file_extension}")
 
-        if args.evaluation_strategy != IntervalStrategy.NO.value:
+        if args.eval_strategy != IntervalStrategy.NO.value:
             new_eval_result = eval_result
 
             if best_iteration is None: