develop 版本的lite预测库和opt转化模型运行iOS demo显示 rogram.cc:187 RuntimeProgram] Check failed: op: no Op found for feed

PaddlePaddle / PaddleOCR

Awesome multilingual OCR toolkits based on PaddlePaddle (practical ultra lightweight OCR system, support 80+ languages recognition, provide data annotation and synthesis tools, support training and deployment among server, mobile, embedded and IoT devices)

Apache License 2.0

39.76k stars 7.39k forks source link

develop 版本的lite预测库和opt转化模型运行iOS demo显示 rogram.cc:187 RuntimeProgram] Check failed: op: no Op found for feed #904

Closed heXiangpeng closed 3 years ago

heXiangpeng commented 3 years ago

lite预测库编译 ./lite/tools/build_ios.sh --arch=armv8 --with_cv=ON --with_extra=ON

develop分之模型转化 ./opt --model_file=./ch_ppocr_mobile_v1.1_rec_quant_infer/model --param_file=./ch_ppocr_mobile_v1.1_rec_quant_infer/params --optimize_out=./ch_ppocr_mobile_v1.1_rec_quant_opt --valid_targets=arm

这样转化出来的模型和预测库替换iOS demo中老的预测库和模型，出现如下错误

.wnloads/Paddle-Lite/lite/core/program.cc:187 RuntimeProgram] Check failed: op: no Op found for feed

LDOUBLEV commented 3 years ago

预测库是哪个版本的。也是develop分支的吗，试试编译Paddle-Lite 2.6.3分支

heXiangpeng commented 3 years ago

切换到2.6.3之后，按照deploy下lite目录下的文档使用转化工具出现如下问题 /opt --model_file=./ch_ppocr_mobile_v1.1_rec_quant_infer/model --param_file=./ch_ppocr_mobile_v1.1_rec_quant_infer/params --optimize_out=./ch_ppocr_mobile_v1.1_rec_quant_opt --valid_targets=arm

报错如下： /Paddle-Lite/lite/core/optimizer.h:232 RunPasses] == Running pass: static_kernel_pick_pass [F 10/10 10:42:23.635 ...lite/core/mir/static_kernel_pick_pass.cc:65 Apply] Check failed: !instruct.kernels().empty(): No kernels found for fake_channel_wise_dequantize_max_abs [3] 56875 abort ./opt --model_file=./quant_model/model --param_file=./quant_model/params

LDOUBLEV commented 3 years ago

切换到2.6.3之后，按照deploy下lite目录下的文档使用转化工具出现如下问题 /opt --model_file=./ch_ppocr_mobile_v1.1_rec_quant_infer/model --param_file=./ch_ppocr_mobile_v1.1_rec_quant_infer/params --optimize_out=./ch_ppocr_mobile_v1.1_rec_quant_opt --valid_targets=arm

报错如下： /Paddle-Lite/lite/core/optimizer.h:232 RunPasses] == Running pass: static_kernel_pick_pass [F 10/10 10:42:23.635 ...lite/core/mir/static_kernel_pick_pass.cc:65 Apply] Check failed: !instruct.kernels().empty(): No kernels found for fake_channel_wise_dequantize_max_abs [3] 56875 abort ./opt --model_file=./quant_model/model --param_file=./quant_model/params

那还是用paddle-lite develop分支试一下吧

heXiangpeng commented 3 years ago

paddle-lite develop 分之的opt 和预测库在iOS 14中报错报错代码 net_ocr1 = CreatePaddlePredictor<MobileConfig>(config);

Check failed: op: no Op found for feed

直接用官网的nb文件，paddle-lite 2.6.3 预测库报错

LoadModelNaiveFromFile] warning: the version of opt that transformed this model is not consistent with current Paddle-Lite version. version of opt:b92e8c4 version of current Paddle-Lite:v2.6.3 paddle-lite-ocr(2524,0x101a83880) malloc: can't allocate region : mach_vm_map(size=68719493120, flags: 100) failed (error code=3) paddle-lite-ocr(2524,0x101a83880) malloc: set a breakpoint in malloc_error_break to debug libc++abi.dylib: terminating with uncaught exception of type std::bad_alloc: std::bad_alloc terminating with uncaught exception of type std::bad_alloc: std::bad_alloc (lldb)

qizhen816 commented 3 years ago

同样的问题 2.6.3模型初始化时报错bad_alloc；develop分支读取模型文件时报错no Op

heXiangpeng commented 3 years ago

同样的问题 2.6.3模型初始化时报错bad_alloc；develop分支读取模型文件时报错no Op

你解决了吗？

qizhen816 commented 3 years ago

同样的问题 2.6.3模型初始化时报错bad_alloc；develop分支读取模型文件时报错no Op

你解决了吗？

没有上周一直在换预测库和模型搞，一直有问题两个git也没有更新只能等等看了

heXiangpeng commented 3 years ago

同样的问题 2.6.3模型初始化时报错bad_alloc；develop分支读取模型文件时报错no Op

你解决了吗？

没有上周一直在换预测库和模型搞，一直有问题两个git也没有更新只能等等看了

我也是天天看他们更新的代码

shangzhizhou commented 3 years ago

@LDOUBLEV 请看下当前进展

qizhen816 commented 3 years ago

@heXiangpeng 你的可以跑了吗

LDOUBLEV commented 3 years ago

Update：

Paddle-Lite 编译和模型转换的问题已经解决，可以切换到Paddle-lite2.7分支，编译预测库和opt模型转换工具；
Paddle-Lite 2.7分支的预测库，测试在android上可以使用，ios当前没有测试机，尚未测试；

qizhen816 commented 3 years ago

@LDOUBLEV 我IOS这边用2.7的转换工具和预测库，转换工具在服务器上编译并转换，预测库在mac上编译，purne和原始的east都会报错：np Op found for feed

CoderMaurice commented 3 years ago

遇到了同样的问题，请问有新的进展吗? @qizhen816 @LDOUBLEV

qizhen816 commented 3 years ago

没有进展，好像ios停止维护了

heXiangpeng commented 3 years ago

@qizhen816 @CoderMaurice iOS 修复了吗？

GreatTreeZX commented 3 years ago

@qizhen816 @heXiangpeng @CoderMaurice @LDOUBLEV @shangzhizhou 有进展吗？同样的问题

DannyIsFunny commented 3 years ago

使用的是静态库吗，静态库需要增加使用两个头文件：

/////////////////////////////////////////////////////////////////////////
// If this demo is linked to static library:libpaddle_api_light_bundled.a
// , you should include `paddle_use_ops.h` and `paddle_use_kernels.h` to
// avoid linking errors such as `unsupport ops or kernels`.
/////////////////////////////////////////////////////////////////////////
// #include "paddle_use_kernels.h"  // NOLINT
// #include "paddle_use_ops.h"      // NOLINT

heXiangpeng commented 3 years ago

@DannyIsFunny 好的，已经测试过lite 2.7 加了这个就运行正常了，感谢

qizhen816 commented 3 years ago

@DannyIsFunny 多谢搞定了

CoderMaurice commented 3 years ago

我放开了所有出现这两个头文件的地方编译了paddle lite 2.7 用的模型是 ch_ppocr_mobile_v1.1_det_prune_opt.nb ch_ppocr_mobile_v1.1_rec_quant_opt.nb 还是报no Op found for feed

@heXiangpeng @qizhen816 你们是怎么操作的？

qizhen816 commented 3 years ago

@CoderMaurice 在你的IOS工程里引用这两个头文件了吗

CoderMaurice commented 3 years ago

@qizhen816 是这几个吗我引用之后报错 no Op found for hard_swish

qizhen816 commented 3 years ago

@CoderMaurice 先确认一下模型转换工具和预测库的版本是不是一致2.7，工程里include的头文件是不是2.7编译后自动生成的，不行的话试试转换工具用develop分支？我也不是很懂，这个hard_swish激活函数的问题貌似有人遇到过。

CoderMaurice commented 3 years ago

@qizhen816

重新转换了一遍模型还是报错

能麻烦把你的模型和paddlelite静态库发给我吗 maurice_mjn@foxmail.com

多谢

qizhen816 commented 3 years ago

@CoderMaurice 我是用的east检测发给你了

CoderMaurice commented 3 years ago

@qizhen816 谢谢我刚搞定原来是我编译的paddle lite有问题使用了默认的参数没有加 extra 和 cv 参数现在一切正常