This package compiles Accelerate code to a TensorFlow graph. For details on Accelerate, refer to the main repository.
Contributions and bug reports are welcome!
Please feel free to contact me through GitHub or gitter.im.
To build a (local) copy of tensorflow-lite:
mkdir build
cd build
cmake ../extra-deps/tensorflow/tensorflow/lite
cmake --build . -j
The TensorFlow C-bindings are required to build this project. In order to
install them, follow the instructions provided by
TensorFlow. Make sure to install
the TensorFlow 2.3.0
CPU
or
GPU
bindings, and not the latest version, since this is the version required by the
TensorFlow-Haskell dependency. As such, we recommend not installing in the
default location (/usr/local on Linux or MacOS systems), but to a different
location. To make sure the build succeeds, you need to tell stack where to find
these files, using the extra-lib-dirs
and extra-include-dirs
fields. Make
sure to set the LIBRARY_PATH
and LD_LIBRARY_PATH
as described in the
installation instructions as well.
To build the required TensorFlow and TensorFlow-haskell packages, you need to have protoc installed. If you do not have it installed, follow the directions on this webpage.
TODO: Make sure everything in this section is correct; at the moment, the list of what to install might be incomplete. TODO: non-debian Linux instructions. Follow the instructions from Coral to get access to their debian packages through apt(-get). Then, install the following libraries:
- libedgetpu-dev (TODO: check necessity, probably required)
- edgetpu_compiler
- libedgetpu1-std (recommended unless the higher frequency is required)
Other dependencies have to be installed manually before running stack build
.
Among these are cpuinfo, farmhash. (TODO: find out what exactly is on this
list.) These exist in the Ubuntu package management system and can be installed
through apt:
sudo apt install libcpuinfo-dev libfarmhash-dev