Merge pull request #16 from ROCmSoftwarePlatform/adabeyta_update_hf_training

Removed hardcoded warmup steps.

Author: amathews-amd

src/transformers/trainer.py

@@ -1876,8 +1876,8 @@ def _inner_training_loop(
 
         metrics = speed_metrics("train", start_time, num_samples=num_train_samples, num_steps=self.state.max_steps)
 
-        total_samples = args.max_steps*total_train_batch_size if args.max_steps > 0  else num_examples*num_train_epochs
-        perf_samples = total_samples - 10*total_train_batch_size
+        total_samples = self.state.global_step*total_train_batch_size if args.max_steps > 0 else num_examples*num_train_epochs
+        perf_samples = total_samples - self.args.warmup_steps*total_train_batch_size
         stable_train_metrics = speed_metrics("stable_train", start_train_stable_time, perf_samples)
 
         self.store_flos()

src/transformers/training_args.py

@@ -568,7 +568,7 @@ class TrainingArguments:
     warmup_ratio: float = field(
         default=0.0, metadata={"help": "Linear warmup over warmup_ratio fraction of total steps."}
     )
-    warmup_steps: int = field(default=0, metadata={"help": "Linear warmup over warmup_steps."})
+    warmup_steps: int = field(default=10, metadata={"help": "Linear warmup over warmup_steps."})
 
     log_level: Optional[str] = field(
         default="passive",