Which sha256 number for Tflite and libedgetpu

[timocafe]

Description

Hello,

[Short] I have some issue with bazel specially for the sha number I should indicate in the WORKSPACE. With my current number my binary crash into InterpreterBuilder. Similar problem here.

[Long] my python code works with 8/16 bits quantize model. Due to performance, I wanted to test the cpp version. To obtain a binary, I have crossed compile with bazel (it works for hello world), however the binary crashes in libedgetpu. I build my bazel on this example.

Finally to still have something, I did a local compilation of tflite-sha d855adfc5a0195788bf5f92c3c7352e638aa1109, it works but the release is "bit" old (jan. 2020, and it does not support 16bit quantization.

I provide a simple hello word bazel/docker. My dev board is also built before april 2019, and it has been flash mendel 5.3. I think several SHA numbers are not up to date. Or I may did a mistake with my bazel, I am not familiar with this packet manager.

bugfolder.zip

make DOCKER_TARGETS=qinait DOCKER_CPUS=aarch64 docker-shell
make
./bug whatever8b.tflite

I can not provide my model due to IP.

Thank you very much

Issue Type

Support

Operating System

Mac OS

Coral Device

Dev Board

Other Devices

No response

Programming Language

C++

Relevant Log Output

// the log comes from the hello word code I provide in the description.

mendel@tuned-jet:~$ ./bug model/model_quantize_8b.tflite 
tpu: 0
type: Apex (PCIe)
path: /dev/apex_0
Building model: model/model_quantize_8b.tflite
Segmentation fault

[hjonnala]

Hello @timocafe

Please try the below steps:
STEP1:

1. git clone --recurse-submodules https://github.com/google-coral/libcoral
2. add bug.cpp file to examples folder: https://github.com/google-coral/libcoral/tree/master/coral/examples
3. add the foolowing code to build file: https://github.com/google-coral/libcoral/blob/master/coral/examples/BUILD

cc_binary(
    name = "bug",
    srcs = ["bug.cpp"],
    deps = [
        "@libedgetpu//tflite/public:oss_edgetpu_direct_all",
        "@org_tensorflow//tensorflow/lite:builtin_op_data",
        "@org_tensorflow//tensorflow/lite:framework",
        "@org_tensorflow//tensorflow/lite/kernels:builtin_ops",
    ],
    linkopts = ["-pthread","-lstdc++fs"],
)

4. add the below line at: https://github.com/google-coral/libcoral/blob/master/Makefile#L102
   //coral/examples:bug \
5. add the below line at: https://github.com/google-coral/libcoral/blob/master/Makefile#L109
   $(BAZEL_OUT_DIR)/coral/examples/bug \

STEP2:

cd libcoral
make docker-shell
make CPU=aarch64 examples

STEP3: (copy the binary to dev board)
mdt push ../../libcoral/out/aarch64/examples/bug

STEP4: (On Dev Board)

mendel@arid-calf:~$ sudo chmod 770 bug
mendel@arid-calf:~$ ./bug efficientdet-lite-salad_edgetpu.tflite 
tpu: 0
type: Apex (PCIe)
path: /dev/apex_0
Building model: efficientdet-lite-salad_edgetpu.tflite
Built interpreter success!
Tensors Allocated!
mendel@arid-calf:~$ 

Here is the pre complied binary: https://github.com/hjonnala/snippets/blob/main/edgetpu_600/bug

[timocafe]

I succeed the cross compilation and it run, thank you very much. Just a last question currently on the dock env the GCC version is gcc 6.3, on the coral it is 8.3. It is bit painful for the c++17 and std::filesystem which is experimental for GCC 6.3. Is it possible to modify the WORKSPACE and update to 8.3 ? Cheers.

[google-coral-bot]

Are you satisfied with the resolution of your issue?
Yes
No