fixing learning rate schedule when using gradient_accumulation_steps

2018-11-10 16:11:14 +01:00
parent ea85cca8ab
commit a81a1ef8e9
2 changed files with 18 additions and 2 deletions
--- a/run_classifier.py
+++ b/run_classifier.py
@@ -464,7 +464,7 @@ def main():
    if args.do_train:
        train_examples = processor.get_train_examples(args.data_dir)
        num_train_steps = int(
-            len(train_examples) / args.train_batch_size * args.num_train_epochs)
+            len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs)

    model = BertForSequenceClassification(bert_config, len(label_list))
    if args.init_checkpoint is not None: