使用paddlejsconverter转换paddleorc的中文文字识别模型后，模型不能正常加载。

[jprorikon]

模型使用 https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.6/doc/doc_en/models_list_en.md#21-chinese-recognition-model 的ch_PP-OCRv3_det的inference model

paddlejsconverter版本1.0.8 转换模型使用以下命令 python 'Paddle.js-release-v2.2.5/packages/paddlejs-converter/convertToPaddleJSModel.py' --modelPath=inference.pdmodel --paramPath=inference.pdiparams --outputDir=../ --useGPUOpt=True --disableOptimize=1

使用的web demo为 https://github.com/PaddlePaddle/FastDeploy/tree/develop/examples/application/js/web_demo中的TextDetection

将模型放置到本地，用inport“?url”的方式传递url给orc.loader

模型不能正常被加载，浏览器控制台报 index.js:1 webgl createProgram: conv2d_depthwise -- Error: Error: compile: ERROR: 0:188: 'undefined' : undeclared identifier ERROR: 0:188: '' : boolean expression expected 错误。

[jprorikon]

@JingyuanZhang 能帮忙解决一下么？

[im-ajaymeena]

I am also having similar issues when running OCR demo with converted model:

for converting I used: python convertToPaddleJSModel.py --modelPath=../../../inference.pdmodel --paramPath=../../../inference.pdiparams --outputDir=. --disableOptimize=1 --useGPUOpt=1

I used japan_PP-OCRv3_rec model

[allenliuvip]

+1 基本上没有模型能成功

[VuHuy-cse-9]

I received errors from paddles like this, conv2d_tranpose and conv2d had a similar notification: webgl createProgram: conv2d_depthwise -- Error: Error: compile: ERROR: 0:188: 'undefined' : undeclared identifier ERROR: 0:188: '' : boolean expression expected I was working around pp-OCRv3-det. I realized something:

• Conv2d-depthwise and Conv2d Tranpose are not supported by paddlejs (I used @paddle-js-models/ocrdet==4.1.1). Conv2d Depthwise is used in MobileNetv3, and Conv2d Tranpose is used in DBHead. You may (somehow) change these operations to make it work.
• Problem of conv2d can be solved by changing paddlepaddle-gpu from 2.6.0.post112 to 2.3.2.post112. I hope suggestions may be helpful.

For paddlejs-converter, I use: paddlepaddle==2.5.1 paddlejsconverter==1.0.7 paddlelite==2.8