exclude jit time from the speed metric calculation of evaluation and prediction (#20553)

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

examples/pytorch/question-answering/trainer_qa.py

@@ -51,10 +51,13 @@ class QuestionAnsweringTrainer(Trainer):
                 # self.args.prediction_loss_only
                 prediction_loss_only=True if compute_metrics is None else None,
                 ignore_keys=ignore_keys,
+                metric_key_prefix=metric_key_prefix,
             )
         finally:
             self.compute_metrics = compute_metrics
         total_batch_size = self.args.eval_batch_size * self.args.world_size
+        if f"{metric_key_prefix}_jit_compilation_time" in output.metrics:
+            start_time += output.metrics[f"{metric_key_prefix}_jit_compilation_time"]
         output.metrics.update(
             speed_metrics(
                 metric_key_prefix,
@@ -74,7 +77,7 @@ class QuestionAnsweringTrainer(Trainer):
                     metrics[f"{metric_key_prefix}_{key}"] = metrics.pop(key)
             metrics.update(output.metrics)
         else:
-            metrics = {}
+            metrics = output.metrics
 
         if self.args.should_log:
             # Only the main node log the results by default
@@ -103,10 +106,13 @@ class QuestionAnsweringTrainer(Trainer):
                 # self.args.prediction_loss_only
                 prediction_loss_only=True if compute_metrics is None else None,
                 ignore_keys=ignore_keys,
+                metric_key_prefix=metric_key_prefix,
             )
         finally:
             self.compute_metrics = compute_metrics
         total_batch_size = self.args.eval_batch_size * self.args.world_size
+        if f"{metric_key_prefix}_jit_compilation_time" in output.metrics:
+            start_time += output.metrics[f"{metric_key_prefix}_jit_compilation_time"]
         output.metrics.update(
             speed_metrics(
                 metric_key_prefix,

examples/pytorch/question-answering/trainer_seq2seq_qa.py

@@ -71,10 +71,13 @@ class QuestionAnsweringSeq2SeqTrainer(Seq2SeqTrainer):
                 # self.args.prediction_loss_only
                 prediction_loss_only=True if compute_metrics is None else None,
                 ignore_keys=ignore_keys,
+                metric_key_prefix=metric_key_prefix,
             )
         finally:
             self.compute_metrics = compute_metrics
         total_batch_size = self.args.eval_batch_size * self.args.world_size
+        if f"{metric_key_prefix}_jit_compilation_time" in output.metrics:
+            start_time += output.metrics[f"{metric_key_prefix}_jit_compilation_time"]
         output.metrics.update(
             speed_metrics(
                 metric_key_prefix,
@@ -94,9 +97,9 @@ class QuestionAnsweringSeq2SeqTrainer(Seq2SeqTrainer):
                 if not key.startswith(f"{metric_key_prefix}_"):
                     metrics[f"{metric_key_prefix}_{key}"] = metrics.pop(key)
 
-            output.metrics.update(metrics)
+            metrics.update(output.metrics)
         else:
-            metrics = {}
+            metrics = output.metrics
 
         if self.args.should_log:
             # Only the main node log the results by default
@@ -106,7 +109,7 @@ class QuestionAnsweringSeq2SeqTrainer(Seq2SeqTrainer):
             # tpu-comment: Logging debug metrics for PyTorch/XLA (compile, execute times, ops, etc.)
             xm.master_print(met.metrics_report())
 
-        self.control = self.callback_handler.on_evaluate(self.args, self.state, self.control, output.metrics)
+        self.control = self.callback_handler.on_evaluate(self.args, self.state, self.control, metrics)
         return metrics
 
     def predict(
@@ -119,6 +122,7 @@ class QuestionAnsweringSeq2SeqTrainer(Seq2SeqTrainer):
         # Temporarily disable metric computation, we will do it in the loop here.
         compute_metrics = self.compute_metrics
         self.compute_metrics = None
+        start_time = time.time()
         eval_loop = self.prediction_loop if self.args.use_legacy_prediction_loop else self.evaluation_loop
         try:
             output = eval_loop(
@@ -128,10 +132,22 @@ class QuestionAnsweringSeq2SeqTrainer(Seq2SeqTrainer):
                 # self.args.prediction_loss_only
                 prediction_loss_only=True if compute_metrics is None else None,
                 ignore_keys=ignore_keys,
+                metric_key_prefix=metric_key_prefix,
             )
         finally:
             self.compute_metrics = compute_metrics
 
+        total_batch_size = self.args.eval_batch_size * self.args.world_size
+        if f"{metric_key_prefix}_jit_compilation_time" in output.metrics:
+            start_time += output.metrics[f"{metric_key_prefix}_jit_compilation_time"]
+        output.metrics.update(
+            speed_metrics(
+                metric_key_prefix,
+                start_time,
+                num_samples=output.num_samples,
+                num_steps=math.ceil(output.num_samples / total_batch_size),
+            )
+        )
         if self.post_process_function is None or self.compute_metrics is None:
             return output
 
@@ -142,5 +158,5 @@ class QuestionAnsweringSeq2SeqTrainer(Seq2SeqTrainer):
         for key in list(metrics.keys()):
             if not key.startswith(f"{metric_key_prefix}_"):
                 metrics[f"{metric_key_prefix}_{key}"] = metrics.pop(key)
-
+        metrics.update(output.metrics)
         return PredictionOutput(predictions=predictions.predictions, label_ids=predictions.label_ids, metrics=metrics)

src/transformers/modelcard.py

@@ -766,6 +766,7 @@ def parse_log_history(log_history):
             _ = metrics.pop("eval_runtime", None)
             _ = metrics.pop("eval_samples_per_second", None)
             _ = metrics.pop("eval_steps_per_second", None)
+            _ = metrics.pop("eval_jit_compilation_time", None)
             values = {"Training Loss": training_loss, "Epoch": epoch, "Step": step}
             for k, v in metrics.items():
                 if k == "eval_loss":

src/transformers/trainer.py

@@ -1345,7 +1345,9 @@ class Trainer:
             model = nn.DataParallel(model)
 
         if self.args.jit_mode_eval:
+            start_time = time.time()
             model = self.torch_jit_model_eval(model, dataloader, training)
+            self.jit_compilation_time = round(time.time() - start_time, 4)
 
         # Note: in torch.distributed mode, there's no point in wrapping the model
         # inside a DistributedDataParallel as we'll be under `no_grad` anyways.
@@ -2819,6 +2821,8 @@ class Trainer:
         )
 
         total_batch_size = self.args.eval_batch_size * self.args.world_size
+        if f"{metric_key_prefix}_jit_compilation_time" in output.metrics:
+            start_time += output.metrics[f"{metric_key_prefix}_jit_compilation_time"]
         output.metrics.update(
             speed_metrics(
                 metric_key_prefix,
@@ -2886,6 +2890,8 @@ class Trainer:
             test_dataloader, description="Prediction", ignore_keys=ignore_keys, metric_key_prefix=metric_key_prefix
         )
         total_batch_size = self.args.eval_batch_size * self.args.world_size
+        if f"{metric_key_prefix}_jit_compilation_time" in output.metrics:
+            start_time += output.metrics[f"{metric_key_prefix}_jit_compilation_time"]
         output.metrics.update(
             speed_metrics(
                 metric_key_prefix,
@@ -3102,6 +3108,8 @@ class Trainer:
 
         if all_losses is not None:
             metrics[f"{metric_key_prefix}_loss"] = all_losses.mean().item()
+        if hasattr(self, "jit_compilation_time"):
+            metrics[f"{metric_key_prefix}_jit_compilation_time"] = self.jit_compilation_time
 
         # Prefix all keys with metric_key_prefix + '_'
         for key in list(metrics.keys()):

src/transformers/trainer_utils.py

@@ -224,7 +224,11 @@ def default_compute_objective(metrics: Dict[str, float]) -> float:
     loss = metrics.pop("eval_loss", None)
     _ = metrics.pop("epoch", None)
     # Remove speed metrics
-    speed_metrics = [m for m in metrics.keys() if m.endswith("_runtime") or m.endswith("_per_second")]
+    speed_metrics = [
+        m
+        for m in metrics.keys()
+        if m.endswith("_runtime") or m.endswith("_per_second") or m.endswith("_compilation_time")
+    ]
     for sm in speed_metrics:
         _ = metrics.pop(sm, None)
     return loss if len(metrics) == 0 else sum(metrics.values())

src/transformers/utils/notebook.py

@@ -339,6 +339,7 @@ class NotebookProgressCallback(TrainerCallback):
             _ = metrics.pop(f"{metric_key_prefix}_runtime", None)
             _ = metrics.pop(f"{metric_key_prefix}_samples_per_second", None)
             _ = metrics.pop(f"{metric_key_prefix}_steps_per_second", None)
+            _ = metrics.pop(f"{metric_key_prefix}_jit_compilation_time", None)
             for k, v in metrics.items():
                 if k == f"{metric_key_prefix}_loss":
                     values["Validation Loss"] = v