NOTE: This is still very much a work in progress, and not even fully operational. DO NOT DEPEND ON THIS PROJECT. I will post updates when it is working and stable.
This is a repository containing the source for Tensorflow (https://www.tensorflow.org/, pinned at version v0.12.0 currently) and slightly modified to be able to be compiled with Emscripten (https://kripken.github.io/emscripten-site/).
- Emscripten
- Protobufs (NOTE: Currently pinned at 3.1.0, TODO: Update to 3.2.0)
- Bazel
Clone the repository. And download js-working-dir into ./tensorflow/contrib/makefile_js/js-working-dir
.
Open the makefile, ./tensorflow/contrib/makefile_js/Makefile
, and you should see
three options at the top which you are free to configure:
FINAL_PRODUCT
(html, js, c): What to produce as final output.ALLOCATE_MEM_UPFRONT
(true, false): Emscripten allocates one large array to simulate the C memory interface in JS, you can choose to statically allocate this at the beginning or to dynamically grow it as need be (MUCH SLOWER).SINGLE_THREAD
(true, false): Whether to use only one thread (must be set to true if you intend to compile for html or js).
Once you have configured the make file, run the following from the root of the repository:
$ emconfigure ./configure
$ ./tensorflow/contrib/makefile_js/download_dependencies.sh
$ bazel build tensorflow/examples/label_image/...
$ ./tensorflow/contrib/makefile_js/gen_file_lists.sh
$ emmake make -f ./tensorflow/contrib/makefile_js/Makefile
NOTE: Must run configure before download dependencies because otherwise the configure script will see build files in downloads and assume it should run those. Problem documented here. The bazel command is required to be run so that tensorflow generates certain source files, there is probably a simpler build for bazel, but we found this one to work.
TIP: Considering adding the flag '-jX' to the emmake command, which will multithread compilation with X threads.
This will generate (the file extensions depend on the FINAL_PRODUCT
flag specified above):
./tensorflow/contrib/makefile_js/gen/bin/benchmark.{o,js,html}
: The standard benchmarking tool that ships with tensorflow../tensorflow/contrib/makefile_js/gen/bin/labeler.{o,js,html}
: A google inception net to classify images (from the label image tutorial)../tensorflow/contrib/makefile_js/gen/bin/mnist.{o,js,html}
: A net trained on mnist to classify images (from the deep mnist tutorial).
To run these files, use the run scripts; for example:
$ cd ./tensorflow/contrib/makefile_js/
$ ./run_scripts/run_labeler_c.sh 0
(That will call the labeler script on makefile_js/js-working-dir/labeler_data/0.jpg
).
Read through run_scripts/*.sh
, to figure out how to call the libraries on their own.
- To clean the makefile-generated objects, call
$ emmake make -f ./tensorflow/contrib/makefile_js/Makefile clean
from the root of the repository. YOU MUST DO THIS BEFORE REBUILDING THE SCRIPTS FOR ANOTHER TARGET (i.e. if you were previouly building for c and now you want to build for js). - We provide a
run_trials
script that can benchmark mnist and labeler, to use it:
$ cd ./tensorflow/contrib/makefile_js/
$ python run_trails.py {mnist,labeler} {browser,node,c}
(Which will benchmark mnist/labeler in the browser/node/c.)
- For some reason trying to include the protobuf library (libprotobuf.a, referred to as $(PROTOBUF_LIB) in the makefile) was causing me grief on Ubuntu. Something to do with the llvm error 'global referenced in another module!'. Was not tested thoroughly (I was debugging a host of other issues at the time so this may have been an emergent bug caused by those), but may be worth looking into. I suspect it's compiler tool chain related (or possibly flag related, noe how we use the flag
-all_load
on OSX)... Bug disappeared when I returned to developing on OSX. - The above error was observed when working with Emscripten 1.37.1 on MacOSX, FOR NOW: PLEASE BUILD WITH EMSCRIPTEN 1.36.5 (this may also be an issue caused if you build initially with one version of emscripten and then upgrade emscripten and continue building... need more investigation)
Tomas Reimers, September 2016 - February 2017