legal
legal
I have the same problem. Have you solved it?
你这个训练的很好啊，我用的train的模式，我的semantic loss在val上都只能到0.001，detail loss是0.03左右，和你的结果差了一个量级呢
top_pool.cpython-37m-x86_64-linux-gnu.so: undefined symbol: _ZN3c105ErrorC1ENS_14SourceLocationERKSs
Have you compiled sucessfully?
In CornerNet.json,you should change both batch_size and chunk_sizes.For example, the batch_size is 30, the chunk_sizes is [15, 15]
How many GPU do you have? If you have just one, I think you can try set chunk_sizes=[1].
https://github.com/PaddlePaddle/Paddle-Lite/issues/10469 希望提供opencl上全局限制Image2D的接口，通过config文件配置若想避免layout算子强制执行buffer2Image，需要把一连串的算子都指定到cpu，但推理就会很慢了
> 麻烦把你说的命令发一下吧 ./lite/tools/build_linux.sh --arch=armv8 --toolchain=gcc --with_opencl=ON --with_log=ON --with_profile=ON --with_extra=ON --with_python=ON
> 我试了下，可以找到 build.lite.linux.armv8.gcc.opencl/inference_lite_lib.armlinux.armv8.opencl 目录啊，我猜测你是改代码后希望能增量编译吧，你可以直接进入 build.lite.linux.armv8.gcc.opencl 执行： > > ``` > $ cd build.lite.linux.armv8.gcc.opencl > $ make -j publish_inference > ``` 好的，非常感谢~