Fix clear gradient of OFA-bert (PaddlePaddle#77)

* fix clear gradient * update readme
wangxicoding · Mar 8, 2021 · a5116de · a5116de
1 parent 2edaba0
commit a5116de
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 2 deletions.
diff --git a/examples/model_compression/ofa/README.md b/examples/model_compression/ofa/README.md
@@ -26,7 +26,7 @@ BERT-base模型是一个迁移能力很强的通用语义表示模型，但是
 首先需要对Pretrain-Model在实际的下游任务上进行Finetuning，得到需要压缩的模型。
 
 ```shell
-cd ../../glue/
+cd ../../benchmark/glue/
 ```
 
 ```python

diff --git a/examples/model_compression/ofa/run_glue_ofa.py b/examples/model_compression/ofa/run_glue_ofa.py
@@ -439,7 +439,7 @@ def do_train(args):
                 loss.backward()
             optimizer.step()
             lr_scheduler.step()
-            ofa_model.model.clear_grad()
+            optimizer.clear_grad()
 
             if global_step % args.logging_steps == 0:
                 if (not args.n_gpu > 1) or paddle.distributed.get_rank() == 0: