The solution to the inconsistency between the PaddleX deployment inference model and the GUI interface test results

Problem description:

Recently, when PaddleX is used for target detection and FasterRCNN is used, the exported inference_model is deployed to windows and the result calculated locally is wrong.

The following is the code for model deployment

model = pdx.deploy.Predictor('F:\panyutong\PaddleX\model\model5\inference_model\inference_model')image_name = 'F:\panyutong\LG\\test\\test\IMAGES\\1404.jpg'predicts = model.predict(image_name)

The predicts test and output are incorrect:

1. The coordinates are negative

2. The confidence level is greater than 1

3. The predicted category exceeds the number of labels

Workaround:

The prediction result using GPU is correct, modify the code as

model = pdx.deploy.Predictor('F:\panyutong\PaddleX\model\model5\inference_model\inference_model',use_gpu=True)image_name = 'F:\panyutong\LG\\test\\test\IMAGES\\1404.jpg'predicts = model.predict(image_name)

---

New question:

Error after using GPU: Process has ended, exit code -1073740791 (0xC0000409)

Workaround:

I am using pycharm, select edit configuration, check the analog output console (screenshots from other blogs are quoted here)

This will output the error message:

Could not locate zlibwapi.dll. Please make sure it is in your library path

It turns out that some resources are missing

Link: https://pan.baidu.com/s/1JE4D5NQ0MfRkZdAEVQZHVQ
Extraction code: t4uv

The lib file is placed in C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.1\lib
Put the dll file in C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.1\bin

Reference blog: Resource acquisition

Run the program again after adding, the result is correct!