[Trainer] Add nan/inf logging filter (#13619)

* finish

* add test

* push

* remove unnecessary code

* up

* correct test

* Update src/transformers/training_args.py

src/transformers/trainer.py

@@ -1297,9 +1297,16 @@ class Trainer:
                 ):
                     # Avoid unnecessary DDP synchronization since there will be no backward pass on this example.
                     with model.no_sync():
-                        tr_loss += self.training_step(model, inputs)
+                        tr_loss_step = self.training_step(model, inputs)
                 else:
-                    tr_loss += self.training_step(model, inputs)
+                    tr_loss_step = self.training_step(model, inputs)
+
+                if args.logging_nan_inf_filter and (torch.isnan(tr_loss_step) or torch.isinf(tr_loss_step)):
+                    # if loss is nan or inf simply add the average of previous logged losses
+                    tr_loss += tr_loss / 1 + (self.state.global_step - self._globalstep_last_logged)
+                else:
+                    tr_loss += tr_loss_step
+
                 self.current_flos += float(self.floating_point_ops(inputs))
 
                 # Optimizer step for deepspeed must be called on every step regardless of the value of gradient_accumulation_steps

src/transformers/training_args.py

@@ -173,6 +173,16 @@ class TrainingArguments:
             Whether to log and evaluate the first :obj:`global_step` or not.
         logging_steps (:obj:`int`, `optional`, defaults to 500):
             Number of update steps between two logs if :obj:`logging_strategy="steps"`.
+        logging_nan_inf_filter (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether to filter :obj:`nan` and :obj:`inf` losses for logging. If set to obj:`True` the loss of every step
+            that is :obj:`nan` or :obj:`inf` is filtered and the average loss of the current logging window is taken
+            instead.
+
+            .. note::
+
+                :obj:`logging_nan_inf_filter` only influences the logging of loss values, it does not change the
+                behavior the gradient is computed or applied to the model.
+
         save_strategy (:obj:`str` or :class:`~transformers.trainer_utils.IntervalStrategy`, `optional`, defaults to :obj:`"steps"`):
             The checkpoint save strategy to adopt during training. Possible values are:
 
@@ -468,6 +478,7 @@ class TrainingArguments:
     )
     logging_first_step: bool = field(default=False, metadata={"help": "Log the first global_step"})
     logging_steps: int = field(default=500, metadata={"help": "Log every X updates steps."})
+    logging_nan_inf_filter: str = field(default=True, metadata={"help": "Filter nan and inf losses for logging."})
     save_strategy: IntervalStrategy = field(
         default="steps",
         metadata={"help": "The checkpoint save strategy to use."},

tests/test_trainer.py

@@ -15,6 +15,7 @@
 
 import dataclasses
 import gc
+import math
 import os
 import random
 import re
@@ -528,6 +529,31 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
         train_output = trainer.train()
         self.assertEqual(train_output.global_step, 10)
 
+    def test_logging_inf_nan_filter(self):
+        config = GPT2Config(vocab_size=100, n_positions=128, n_ctx=128, n_embd=32, n_layer=3, n_head=4)
+        tiny_gpt2 = GPT2LMHeadModel(config)
+        x = torch.randint(0, 100, (128,))
+        train_dataset = RepeatDataset(x)
+
+        # Trainer without inf/nan filter
+        args = TrainingArguments("./test", learning_rate=1e9, logging_steps=5, logging_nan_inf_filter=False)
+        trainer = Trainer(tiny_gpt2, args, train_dataset=train_dataset)
+        trainer.train()
+        log_history_no_filter = trainer.state.log_history
+
+        # Trainer with inf/nan filter
+        args = TrainingArguments("./test", learning_rate=1e9, logging_steps=5, logging_nan_inf_filter=True)
+        trainer = Trainer(tiny_gpt2, args, train_dataset=train_dataset)
+        trainer.train()
+        log_history_filter = trainer.state.log_history
+
+        def is_any_loss_nan_or_inf(log_history):
+            losses = [l["loss"] for l in log_history[:-1]]
+            return any(math.isnan(x) for x in losses) or any(math.isinf(x) for x in losses)
+
+        self.assertTrue(is_any_loss_nan_or_inf(log_history_no_filter))
+        self.assertFalse(is_any_loss_nan_or_inf(log_history_filter))
+
     def test_train_and_eval_dataloaders(self):
         n_gpu = max(1, torch.cuda.device_count())
         trainer = get_regression_trainer(learning_rate=0.1, per_device_train_batch_size=16)