repair train bug in multi gpu

PaddlePaddle · Jan 21, 2025 · 7aeb60c · 7aeb60c
1 parent 1049c66
commit 7aeb60c
Show file tree

Hide file tree

Showing 2 changed files with 5 additions and 1 deletion.
diff --git a/configs/rec/rec_latex_ocr.yml b/configs/rec/rec_latex_ocr.yml
@@ -19,6 +19,7 @@ Global:
   rec_char_dict_path:  ppocr/utils/dict/latex_ocr_tokenizer.json
   save_res_path: ./output/rec/predicts_latexocr.txt
   d2s_train_image_shape: [1,256,256]
+  find_unused_parameters: True
 
 Optimizer:
   name: AdamW

diff --git a/tools/train.py b/tools/train.py
@@ -217,7 +217,10 @@ def main(config, device, logger, vdl_writer, seed):
     )
 
     if config["Global"]["distributed"]:
-        model = paddle.DataParallel(model)
+        find_unused_parameters = config["Global"].get("find_unused_parameters", False)
+        model = paddle.DataParallel(
+            model, find_unused_parameters=find_unused_parameters
+        )
     # start train
     program.train(
         config,