The prediction accuracy of the pcpvt_base model after mkldnn is turned on is different from that of pure CPU, and the difference is greater than 0.01.

[baoachun]

Commit id：0be7157185138a5e92732225f56737d34fc532f8

PR: #34192

[lidanqing-intel]

@baoachun Where is this screen shot from? With this we could not reproduce. Could you please provide command? And Is this issue must be fixed for Paddle 2.2

@jczaja Please help watch this issue. Thanks

[baoachun]

@lidanqing-intel @jczaja
Hi, there is the test code, https://github.com/PaddlePaddle/PaddleTest/blob/develop/inference/python_api_test/test_class_model/test_pcpvt_base.py#L64.

python3.8 -m pytest --disable-warnings -v test_pcpvt_base.py::test_mkldnn

[lidanqing-intel]

@jczaja This is not urgent for 2.2RC

[lidanqing-intel]

@baoachun @jakpiase Baoachun updated this issue caused other 4 issues. So I revert your PR first. If you make fix PR before 17th September, please @baoachun to let him test his 5 cards first and then merge the fix PR. Thanks.
Update: This issue need to be fixed for v2.2

[lidanqing-intel]

@jakpiase Hi, I can not revert for you, because there are too many file conflicts and I m not sure if I could remove or not. Please revert it or fix it before 17th September (which is like night 16th September Poland). Keep updating under this issue. Thanks.

[baoachun]

First step:
git clone https://github.com/PaddlePaddle/PaddleClas.git
git checkout develop

Second step:
Download models and images from here: https://github.com/PaddlePaddle/PaddleClas/blob/develop/docs/en/tutorials/quick_start_recognition_en.md

Problem 1:
cd PaddleClas/deploy; python python/predict_system.py -c configs/inference_product.yaml -o Global.infer_imgs=./recognition_demo_data_v1.1/test_product/daoxiangcunjinzhubing_6.jpg -o Global.use_gpu=False -o Global.use_tensorrt=False -o Global.use_fp16=False -o Global.enable_mkldnn=True

Problem 2:
cd PaddleClas/deploy; python python/predict_system.py -c configs/inference_vehicle.yaml -o Global.infer_imgs=./recognition_demo_data_v1.1/test_vehicle/audia4-102.jpeg -o Global.use_gpu=False -o Global.use_tensorrt=False -o Global.use_fp16=False -o Global.enable_mkldnn=True

Problem 3:
cd PaddleClas/deploy; python python/predict_system.py -c configs/inference_logo.yaml -o Global.infer_imgs=./recognition_demo_data_v1.1/test_logo/benz-001.jpeg -o Global.use_gpu=False -o Global.use_tensorrt=False -o Global.use_fp16=False -o Global.enable_mkldnn=True

Problem 4:
cd PaddleClas/deploy; python python/predict_cls.py -c configs/inference_cls.yaml -o Global.inference_model_dir=../inference/ShuffleNetV2 -o batch_size=1 -o Global.use_gpu=False -o Global.use_tensorrt=False -o Global.enable_mkldnn=True

Problem 5:
cd PaddleClas/deploy; python python/predict_system.py -c configs/inference_cartoon.yaml -o Global.infer_imgs=./recognition_demo_data_v1.1/test_cartoon/aisidesi-001.jpeg -o Global.use_gpu=False -o Global.use_tensorrt=False -o Global.use_fp16=False -o Global.enable_mkldnn=True

[baoachun]

#35794
The same question.

[jakpiase]

@baoachun I couldn't find a way how to download "..inference/ShuffleNetV2" model, but I have found

https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/whole_chain/ShuffleNetV2_x1_0_inference.tar

And I have ran this model. It gives same result with and without oneDNN enabled. If I have tested wrong model could you please tell me how to download the correct one?
And pcpvt_base model will be fixed by: #35781(disabling reshape and squeeze ops)

[baoachun]

@jakpiase
Hi, the model that you used is correct, and the prediction result is wrong compare with CPU.

[jakpiase]

Hi, @baoachun, sorry for my mistake, you are right. These results are incorrect now, but I have checked and they will be fixed after #35781 is merged