tensorflow: Unable to cross compile TFLite for Raspberry Pi Zero using Docker image nightly-develop

System information

• OS Platform and Distribution: Linux Mint 19:03 (based on Ubuntu 18:04)
• Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device: N/A ? Attempting to crosscompile tensorflowlite for Raspbery Pi Zero (Armv6)
• TensorFlow installed from (source or binary): nightly-develop docker image
• TensorFlow version: 2.2
• Python version: 2.7
• Installed using virtualenv? pip? conda?: docker file from latest nightly build
• Bazel version (if compiling from source): bazel 0:15:0
• GCC/Compiler version (if compiling from source): gcc (Ubuntu 7.3.0-27ubuntu1~18.04) 7.3.0
• CUDA/cuDNN version: n/a
• GPU model and memory: none

When cross compiling for raspberry pi zero (ARMv6), as per instructions on following page https://www.tensorflow.org/lite/guide/build_rpi and running the command indicated to build for the ARMv6 architecture, ./tensorflow/lite/tools/make/build_rpi_lib.sh TARGET_ARCH=armv6

there are errors indicating that the compiler options set are incorrect

/usr/arm-linux-gnueabihf/include/bits/stdio.h: In function ‘int getchar()’: /usr/arm-linux-gnueabihf/include/bits/stdio.h:44:14: sorry, unimplemented: Thumb-1 hard-float VFP ABI getchar (void)

Any other info / logs

if I examine the file

./tensorflow/lite/tools/make/build_rpi_lib.sh I found that the TARGET_ARCH parameter was hard coded for armv7 in the final line of the file

If I edit the file so that this reads

CC_PREFIX=arm-linux-gnueabihf- make -j 3 -f tensorflow/lite/tools/make/Makefile TARGET=rpi TARGET_ARCH=armv6 (the hardcoding is not ideal for obvious reasons ) and then change the armv6 compiler options in

tensorflow/lite/tools/make/targets/rpi_makefile.inc to the following

ifeq ($(TARGET_ARCH), armv6)
CXXFLAGS +=
-marm
-mfpu=vfp
-mlong-calls
-mthumb-interwork
-mfloat-abi=hard

CCFLAGS += \
  -marm \
  -mfpu=vfp \
  -mlong-calls \
  -mthumb-interwork \
  -mfloat-abi=hard

LDFLAGS := \
  -Wl,--no-export-dynamic \
  -Wl,--exclude-libs,ALL \
  -Wl,--gc-sections \
  -Wl,--as-needed \
  -mfpu=vfp \
  -mlong-calls \
  -mfloat-abi=hard \
  -mthumb-interwork 
endif

I can compile without errors (and just a couple of warnings).

Given that it seems to be difficult (or just very slow) to compile bazel directly on a raspberry pi zero due to the lack of memory and that increasing the memory available (even via machine virtualisation such as QEMU appears impossible), I think that for deployment on a PI Zero, cross compiling is the only way forward. At the moment I can inference, but only by using a full tensorflow implementation, which is really too slow for my needs)

Additionally it would be nice to have a way to generate a python wheel as well. The main ci_build process does this for the full fat tensorflow package, (and i can run one of those on the pi zero), but I’d like to just use tensorflow-lite for my inferencing, as i think any performance improvements that I can get would be beneficial. I’m using a Pi Zero for it’s low power consumption for energy efficiency reasons, so swapping out to a PI 3 or Pi 4 would not really be viable for my project.

thanks for your help