dsindex
commented
7 years ago @bazhura
what about do not patch?
i means
$ git clone --recurse-submodules https://github.com/tensorflow/serving
# checkout proper version of serving
$ cd serving
$ git checkout 89e9dfbea055027bc31878ee8da66b54a701a746
$ git submodule update --init --recursive
# checkout proper version of tf_models
$ cd tf_models
$ git checkout a4b7bb9a5dd2c021edcd3d68d326255c734d0ef0
$ cd syntaxnet
$ cd tensorflow
$ ./configure
$ cd ..
# modify ./tensorflow/tensorflow/workspace.bzl for downloading 'zlib'
# native.new_http_archive(
# name = "zlib_archive",
# url = "http://zlib.net/fossils/zlib-1.2.8.tar.gz",
# sha256 = "36658cb768a54c1d4dec43c3116c27ed893e88b02ecfcb44f2166f9c0b7f2a0d",
# build_file = path_prefix + "zlib.BUILD",
# )
$ bazel test --linkopt=-headerpad_max_install_names syntaxnet/... util/utf8/...and then
$ git clone https://github.com/dsindex/syntaxnet.git work
$ cd
....
$ ./train.sh
....finally, export models and use it
(say, /path/to/models)
# goto pathed-version
$ cat /path/to/models/context.pbtxt.template | sed "s=OUTPATH=/path/to/models=" > /path/to/models/context.pbtxt
$ bazel-bin/tensorflow_serving/example/parsey_mcparseface --model_dir=/path/to/models --export_path=exported
# modify all path in exported/00000001/assets/context.pbtxt
# for example,
# from
# input {
# name: "tag-map"
# Part {
# file_pattern: "syntaxnet/models/parsey_mcparseface/tag-map"
# }
# }
# to
# input {
# name: "tag-map"
# Part {
# file_pattern: "tag-map"
# }
# }
# run parsey_api with exported model
$ ./bazel-bin/tensorflow_serving/example/parsey_api --port=9000 exported/00000001
bazhura
Hello @dsindex Thank you for bringing up this project, I believe it's magic to have it.
Struggling to put things together to serve Croatian model from parsey's cousins on tensorflow serving. (Let's call it "Different-Lang" model). I have reached the point where the trained model needs to be exported, but I cannot get my head around it. Both models_tf in serving and syntaxnet for training was checkout at a4b7bb9a5dd2c021edcd3d68d326255c734d0ef0 .
First things first, I built TF serving as described.
Next, built SyntaxNet for CPU on a different machine, like a separate project
Test Parser Trainer
Download UD and train Different-Lang model
add the context.pbtxt and update file location value, file names “...-ud-...” + record-format to "different-lang-text"
./train.sh -v -v
$ cp ../api/parsey_mcparseface.py tensorflow_serving/example $ bazel --output_user_root=bazel_root build --nocheck_visibility -c opt -s //tensorflow_serving/example:parsey_mcparseface --genrule_strategy=standalone --spawn_strategy=standalone --verbose_failures $ ls bazel-bin/tensorflow_serving/example/parsey_mcparseface
$ cat ../models/context.pbtxt.template | sed "s=OUTPATH=/home/alina/work/UD_Different-Lang/tmp/syntaxnet-output/brain_pos/greedy/64-0.08-3600-0.9-0=" > ../models/context.pbtxt $ bazel-bin/tensorflow_serving/example/parsey_mcparseface --model_dir=../models --export_path=exported