Library size is not 16KB as advertised. What are ways to reduce size?

[aberriangw]

I have the same issue as in #1815, except the build size is 1.1MB even with BUILD_TYPE=release. That is, if I run the command

make -f tensorflow/lite/micro/tools/make/Makefile TARGET=cortex_m_generic TARGET_ARCH=cortex-m4+fp microlite -j16 BUILD_TYPE=release

then the resulting static library gen/cortex_m_generic_cortex-m4+fp_release_gcc/lib/libtensorflow-microlite.a is 1.1MB, not anywhere near the 16KB I expected.

Here are my questions:

• The documentation here claims that the "core runtime" fits into 16KB on a Cortex M3. Doing the same command as above but with TARGET_ARCH=cortex-m3 yields a 1.1MB file gen/cortex_m_generic_cortex-m3_release_gcc/lib/libtensorflow-microlite.a. Is there some "core" that is only 16KB, whereas the rest of what's included in the static library increases the size to 1.1MB? If so, how can I isolate this "core" or understand better what the core library is? Or, rather, has the size increased to 16KB over time as the library has been further developed, and the documentation is just out of date?

• Is it possible to compile a model into a binary using this repo, such that (1) the binary only contains the necessary operators to run the model, and not the entirety of the TFLM library; and (2) the binary does not have any external dependency on TFLM? I feel like this would solve any size issues. I tried to compile the Hello World program into a binary as prescribed here, but compilation failed.

Thanks for your help!
P.S. I would have reopened the earlier issue, but I don't have permission, and I wasn't sure that anyone would have seen if I left a comment.

[ddavis-2015]

Here are my questions:

• The documentation here claims that the "core runtime" fits into 16KB on a Cortex M3. Doing the same command as above but with TARGET_ARCH=cortex-m3 yields a 1.1MB file gen/cortex_m_generic_cortex-m3_release_gcc/lib/libtensorflow-microlite.a. Is there some "core" that is only 16KB, whereas the rest of what's included in the static library increases the size to 1.1MB? If so, how can I isolate this "core" or understand better what the core library is? Or, rather, has the size increased to 16KB over time as the library has been further developed, and the documentation is just out of date?

The size of libtensorflow-microlite.a will vary over time. The size of your particular application is dependent on you (the developer) and what portion of the TFLM library is used by your application, as per your use compiler/linker flags and/or scripts.

• Is it possible to compile a model into a binary using this repo, such that (1) the binary only contains the necessary operators to run the model, and not the entirety of the TFLM library; and (2) the binary does not have any external dependency on TFLM? I feel like this would solve any size issues. I tried to compile the Hello World program into a binary as prescribed here, but compilation failed.

The number of operators (kernels) that would be included by the linker into your application can be controlled by your application. See the example for hello_world. Also, the README for the hello_world example provides working build instructions.

This issue will now be closed. Please feel free to reopen this issue should you have further comments.