some errors in nvdla_compiler parsing Alexnet.caffemodel as nv_small configuration.

[LpengPC]

(1)the new nvdla_compiler can parse lenet as nv_small , and it works well on FPGA.

(2)while in parsing the Alexnet as nv_small, the compiler reported some errors as follows:

./nvdla_compiler --prototxt ./models/deploy.prototxt --caffemodel ./models/bvlc_alexnet.caffemodel --cprecision int8 --configtarget nv_small

creating new wisdom context... opening wisdom context... parsing caffe network... libnvdla<3> mark prob Marking total 1 outputs initialize all tensors with const scaling factors of 127... attaching parsed network to the wisdom... compiling profile "fast-math"... config "nv_small"... libnvdla<2> Overriding output multibatch size from 1 to 10 (DLA) Error 0x0000000b: Couldn't narrow down to suitable SF for e-2/tsd-27 (propagating from engine-ast/EngineEdge.cpp, function determineSurfaceFormat(), line 260) (DLA) Error 0x0000000b: (propagating from engine-ast/EngineGraph.cpp, function registerAllSurfaces(), line 948) (DLA) Error 0x0000000b: (propagating from Compiler.cpp, function registerBuffers(), line 1568) (DLA) Error 0x00000008: failed compilation phase: registerBuffers (propagating from Compiler.cpp, function compileInternal(), line 488) (DLA) Error 0x00000008: (propagating from Compiler.cpp, function compileInternal(), line 405) (DLA) Error 0x00000008: (propagating from Compiler.cpp, function compile(), line 354) (DLA) Error 0x00000008: (propagating from CompileTest.cpp, function compileProfile(), line 48) (DLA) Error 0x00000008: (propagating from ParseTest.cpp, function parseAndCompile(), line 235) (DLA) Error 0x00000008: (propagating from main.cpp, function launchTest(), line 96)

(3)while in parsing the Alexnet as nv_full, the compiler is ok.

[prasshantg]

@LpengPC can you please try with batch size 1?

[prasshantg]

I am able to reproduce this issue, checking.

[LpengPC]

@prasshantg, the compiler can parse the Alexnet as nv_small well now?

[shazib-summar]

Any update on this, guys?

[shazib-summar]

can you please try with batch size 1?

How do I do that?

[prasshantg]

@LpengPC @killerzula AlexNet will not work with this compiler release as int8 group conv and LRN is not supported.

Please use resnet-50 which is verified network with this compiler release.

[AmritaMathuriya]

@prasshantg Could you please give me the link to the tested Resnet-50 workload? Is it included in the Git repo?

[prasshantg]

@AmritaMathuriya https://github.com/KaimingHe/deep-residual-networks

[lkr856]

@LpengPC my error is samilar : sudo ./nvdla_compiler --prototxt /home/server/Downloads/nn/caffe-master/models/bvlc_alexnet/deploy.prototxt --caffemodel /home/server/Downloads/nn/bvlc_alexnet.caffemodel --configtarget nv_fullcreating new wisdom context... opening wisdom context... parsing caffe network... libnvdla<3> mark prob Marking total 1 outputs initialize all tensors with const scaling factors of 127... attaching parsed network to the wisdom... compiling profile "fast-math"... config "nv_full"... (DLA) Error 0x0000000b: Couldn't narrow down to suitable SF for e-2/tsd-27 (propagating from engine-ast/EngineEdge.cpp, function determineSurfaceFormat(), line 278) (DLA) Error 0x0000000b: (propagating from engine-ast/EngineGraph.cpp, function registerAllSurfaces(), line 966) (DLA) Error 0x0000000b: (propagating from Compiler.cpp, function registerBuffers(), line 1586) (DLA) Error 0x00000008: failed compilation phase: registerBuffers (propagating from Compiler.cpp, function compileInternal(), line 506) (DLA) Error 0x00000008: (propagating from Compiler.cpp, function compileInternal(), line 423) (DLA) Error 0x00000008: (propagating from Compiler.cpp, function compile(), line 372) (DLA) Error 0x00000008: (propagating from CompileTest.cpp, function compileProfile(), line 66) (DLA) Error 0x00000008: (propagating from ParseTest.cpp, function parseAndCompile(), line 252) (DLA) Error 0x00000008: (propagating from main.cpp, function launchTest(), line 111)

so how to deal with this? thanks

[lkr856]

And can someone explain the error "Couldn't narrow down to suitable SF for e-2/tsd-27" to me

[linkongyuan]

@LpengPC @lkr856 Hi, guys, I have the same problems like below:

[nvdla@ubuntu: /home/nvdla/Project/vp/sw/prebuilt/x86-ubuntu/] $ ./nvdla_compiler --prototxt bvlc_alexnet/deploy.prototxt --caffemodel bvlc_alexnet/bvlc_alexnet.caffemodel creating new wisdom context... opening wisdom context... parsing caffe network... libnvdla<3> mark prob Marking total 1 outputs initialize all tensors with const scaling factors of 127... attaching parsed network to the wisdom... compiling profile "fast-math"... config "nv_full"... (DLA) Error 0x0000000b: Couldn't narrow down to suitable SF for e-2/tsd-27 (propagating from engine-ast/EngineEdge.cpp, function determineSurfaceFormat(), line 278) (DLA) Error 0x0000000b: (propagating from engine-ast/EngineGraph.cpp, function registerAllSurfaces(), line 966) (DLA) Error 0x0000000b: (propagating from Compiler.cpp, function registerBuffers(), line 1586) (DLA) Error 0x00000008: failed compilation phase: registerBuffers (propagating from Compiler.cpp, function compileInternal(), line 506) (DLA) Error 0x00000008: (propagating from Compiler.cpp, function compileInternal(), line 423) (DLA) Error 0x00000008: (propagating from Compiler.cpp, function compile(), line 372) (DLA) Error 0x00000008: (propagating from CompileTest.cpp, function compileProfile(), line 66) (DLA) Error 0x00000008: (propagating from ParseTest.cpp, function parseAndCompile(), line 252) (DLA) Error 0x00000008: (propagating from main.cpp, function launchTest(), line 111)

Have you solved your problem? And can you kindly help me with these problems? Thank you very much!