Created by: cryoco

System information -PaddlePaddle version: v1.8.0, 0231f58e592ad9f673ac1832d8c495c8ed65d24f -CPU: Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz -GPU: V100 -OS Platform (eg.Mac OS 10.14): Ubuntu1404 -Docker image: hub.baidubce.com/paddlepaddle/paddle:latest-dev -Python version: Python3.7 -Cmake orders -C++version.txt: used python api -API information: inference configuration

config.switch_use_feed_fetch_ops(False)
config.disable_gpu()

To Reproduce

1. run test with conv_transpose_eltwiseadd_bn_fuse_pass

python3.7 test_ocr_det.py --model_file=det/model --params_file=det/params --float32

The head part of output：

2. run test without conv_transpose_eltwiseadd_bn_fuse_pass by adding config.delete_pass("conv_transpose_eltwiseadd_bn_fuse_pass") in AnalysisConfig.

python3.7 test_ocr_det.py --model_file=det/model --params_file=det/params --float32

The head part of output：

Describe your current behavior It has been verified that when using conv_transpose_eltwiseadd_bn_fuse_pass, the cpu inference output is wrong. It's easy to check the right output by using GPU to perform inference:

#config.disable_gpu()
config.enable_use_gpu(100, 0)

This issue can be reproduced with and without MKLDNN.

Code to reproduce the issue test_ocr_det.txt

Need to be renamed to test_ocr_det.py

Model to reproduce the issue

wget https://paddleocr.bj.bcebos.com/inference.tar
tar -xf inference.tar

The det model is the one which reproduces this issue.

Other info / logs