I tried to compile and run mxnet-ssd ssd_inceptionv3_512 model on intel graphics (Deeplens Atom E3930).
I used Feb 7th TVM version and Mar 7th version with PR 2747
It compiles without any errors
It gives correct result for target: opencl, target_host: llvm, but it is very slow (16,851 ms on Atom). Result:
time: 16851 ms
1 car [6. 0.97556907 0.599846 0.1386168 0.9068787 0.29633522]
2 bicycle [1. 0.7629443 0.18050349 0.2417408 0.75496024 0.76243854]
But it gives incorrect inference result on Intel Atom if I set target to
target: opencl -device=intel_graphics -model=unknown, target_host: llvm
Result:
time: 1987 ms
1 pottedplant [15. 0.7816069 0. 0.07468978 1. 0.962857 ]
Correct result should be car, dog and bicycle.
I prepared code which helps to reproduce incorrect output clone my mxnet-ssd-tvm repo
clone git@github.com:apivovarov/mxnet-ssd-tvm.git
# combine deploy_ssd_inceptionv3_512-0000.params parts
cd deploy_ssd_inceptionv3_512
cat deploy_ssd_inceptionv3_512-0000.params.a* > deploy_ssd_inceptionv3_512-0000.params
Compile ssd_inceptionv3_512 model on main computer (having tvm + nnvm + topi)
python3 compile.py
# it will generate 3 files for intel graphics device.
model.so
model.json
model.params
copy files to device with Atom Intel Graphics and run the model
python3 run-ssd.py
# it will print
time: 1999 ms
1 chair [8. 0.5014031 0. 0.067276 1. 0.93533254]
# or
time: 1987 ms
1 pottedplant [15. 0.7816069 0. 0.07468978 1. 0.962857 ]
# or
time: 2009 ms
1 pottedplant [15. 0.6881262 0. 0.03234622 0.9601488 0.97136605]
2 bird [2. 0.55163026 0.41290188 0.53170156 0.47892678 0.65033484]
I tried to run inference on Intel i7 i7-8700 using the same compiled model files and I got CORRECT result. So, the issue can be reproduced only on Intel Atom cpu/gpu I used Intel® Atom™ Processor E3930 @ 1.30GHz