TensorFlowLite for Unikraft

This is the port of tensorflowlite as external library Please refer to the README.md as well as the documentation in the doc/ subdirectory of the main unikraft repository.

Build

TensorFlowLite interpreter depends on the following libraries, that need to be added to Makefile in this order:

pthreads, e.g. pthread-embedded
libcxx
libcxxabi
libc, e.g. newlib
libunwind
libcompilerrt
libgemmlowp
libflatbuffers
libfarmhash
libeigen
libfft2

Root filesystem

Creating the filesystem

TensorFlowLite needs a filesystem which should contain one or more tflite models. Therefore, the filesystem needs to be created before running the VM.

Using the filesystem

Mounting the filesystem is a transparent operation. All you have to do is to provide the right Qemu parameters in order for Unikraft to mount the filesystem. We will use the 9pfs support for filesystems and for this you will need to use the following parameters:

-fsdev local,id=myid,path=<some directory>,security_model=none \
-device virtio-9p-pci,fsdev=myid,mount_tag=rootfs,disable-modern=on,disable-legacy=off

You should also use vfs.rootdev=rootfs (set by default) to specify the 9pfs mounting tag to Unikraft. To enable 9pfs, you'll need to select the following menu options, all under Library Configuration (this should be already done by the tflite config file):

uk9p: 9p client
vfscore: VFS Core Interface → vfscore: Configuration → Automatically mount a root filesysytem → Default root filesystem → 9PFS

Alternatively, and perhaps easier, is to use the qemu-guest script here:

kvm-guest -k build/helloworld_kvm-x86_64 -e rootfs -a "vfs.rootdev=fs0 --" -m 1024

How to run

Currently, main.cpp contains a minimal example for loading a tflite model and printing the interpreter state. The sample program will try to load the model from mobilenet_v1_1.0_224.tflite (this model and other models from the same family can be downloaded from here)

skuenzer / lib-tflite