build successed, but run failed. about retinaface-tvm HOT 10 OPEN

you can see your compiled_model.json is different from mine, the last shape in the list_shape is [1, 4, 80, 60, 5], which has 5 dimensions, it should be wrong according to the retinaface model structure. So you can check your model conversion script again.

you may try the new tvm.relay module to complier the model.

from retinaface-tvm.

Hi, my OS is ubuntu 18.04, use cmd 'ldd --version' and it indicates that the GLIBC version is 2.27.
The llvm toolchain optimize the ./models/480P/mnet.25.x86.cpu.so which depends on GLIBC_2.27, and if your OS is ubuntu 16.04, the GLIBC version is 2.23 and it;s not compatible for my compliled ./models/480P/mnet.25.x86.cpu.so.

So you need to compile your own tvm relay/nnvm models based on your system.

from retinaface-tvm.

Thank you very much. I run your project after compile retinaface model, but I encounter error again, and my compile code is here. may you pleasure to help me correct this problem?

[ 20%] Building CXX object CMakeFiles/retinaface_tvm.dir/src/anchor_generator.cpp.o
[ 40%] Building CXX object CMakeFiles/retinaface_tvm.dir/src/config.cpp.o
[ 60%] Building CXX object CMakeFiles/retinaface_tvm.dir/src/detect.cpp.o
[ 80%] Building CXX object CMakeFiles/retinaface_tvm.dir/src/tools.cpp.o
[100%] Linking CXX executable ../bin/retinaface_tvm
[100%] Built target retinaface_tvm
Pre Forward Time: 16.0252ms
terminate called after throwing an instance of 'dmlc::Error'
what(): [09:21:40] /home/rongy/ML_frame/tvm/src/runtime/graph/graph_runtime.cc:140: Check failed: data->ndim == data_out->ndim (5 vs. 4) :
Stack trace:
[bt] (0) /usr/local/lib/libtvm_runtime.so(+0x19192) [0x7fd2298f2192]
[bt] (1) /usr/local/lib/libtvm_runtime.so(+0x84b0a) [0x7fd22995db0a]
[bt] (2) /usr/local/lib/libtvm_runtime.so(+0x85ae2) [0x7fd22995eae2]
[bt] (3) ./bin/retinaface_tvm(+0x5bde) [0x563878299bde]
[bt] (4) /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf0) [0x7fd228c8e830]
[bt] (5) ./bin/retinaface_tvm(+0x7499) [0x56387829b499]

Aborted (core dumped)

from retinaface-tvm.

Hi, did you replace the ./lib/libtvm_runtime.so? it shows your tvm base code is not as newer as mine.
the error means your input data shape dimensions of one layer is not the same as defined in your ./models/480P/mnet.25.x86.cpu.json (list_shape dict).
you can check your ./models/480P/mnet.25.x86.cpu.json file and replace the ./lib/libtvm_runtime.so.

from retinaface-tvm.

I recompiled the tvm and replace the libtvm_runtime.so before run your project. And sorry, I didn't link the model and its compile code successfully, here is it

from retinaface-tvm.

you can see your compiled_model.json is different from mine, the last shape in the list_shape is [1, 4, 80, 60, 5], which has 5 dimensions, it should be wrong according to the retinaface model structure. So you can check your model conversion script again.

from retinaface-tvm.

yes, there are many difference between our compiled models, I just want to know where is your raw retinaface model, mine is here RetinaFace-MobileNet0.25

from retinaface-tvm.

yes, mine is downloaded from googledrive, which is different from https://github.com/Charrin/RetinaFace-Cpp/blob/master/convert_models/mnet/mnet.25-symbol.json

from retinaface-tvm.

it

HI Royzon, I m facing the same issue, could I see your compile model ?

Many thanks

from retinaface-tvm.

@Howave the model you downloaded is 640x640 input(from the caffe prototxt, but I think the input should same with mxnet one), however, in your source code, it is 480(width)x640(height), have you get the correct result?

from retinaface-tvm.