Update the DRAGNN

syntaxnet/Dockerfile

 # Java baseimage, for Bazel.
-FROM java:8
+FROM openjdk:8
 
 ENV SYNTAXNETDIR=/opt/tensorflow PATH=$PATH:/root/bin
 
@@ -50,6 +50,8 @@ RUN python -m pip install \
     && python -m pip install pygraphviz \
           --install-option="--include-path=/usr/include/graphviz" \
           --install-option="--library-path=/usr/lib/graphviz/" \
+    && python -m jupyter_core.command nbextension enable \
+          --py --sys-prefix widgetsnbextension \
     && rm -rf /root/.cache/pip /tmp/pip*
 
 # Installs the latest version of Bazel.
@@ -86,6 +88,5 @@ EXPOSE 8888
 # This does not need to be compiled, only copied.
 COPY examples $SYNTAXNETDIR/syntaxnet/examples
 # Todo: Move this earlier in the file (don't want to invalidate caches for now).
-RUN jupyter nbextension enable --py --sys-prefix widgetsnbextension
 
 CMD /bin/bash -c "bazel-bin/dragnn/tools/oss_notebook_launcher notebook --debug --notebook-dir=/opt/tensorflow/syntaxnet/examples"

syntaxnet/README.md

@@ -20,12 +20,16 @@ This repository is largely divided into two sub-packages:
 
 1.  **DRAGNN:
     [code](https://github.com/tensorflow/models/tree/master/syntaxnet/dragnn),
-    [documentation](g3doc/DRAGNN.md)** implements Dynamic Recurrent Acyclic
-    Graphical Neural Networks (DRAGNN), a framework for building multi-task,
-    fully dynamic constructed computation graphs. Practically, we use DRAGNN to
-    extend our prior work from [Andor et al.
+    [documentation](g3doc/DRAGNN.md),
+    [paper](https://arxiv.org/pdf/1703.04474.pdf)** implements Dynamic Recurrent
+    Acyclic Graphical Neural Networks (DRAGNN), a framework for building
+    multi-task, fully dynamically constructed computation graphs. Practically, we
+    use DRAGNN to extend our prior work from [Andor et al.
     (2016)](http://arxiv.org/abs/1603.06042) with end-to-end, deep recurrent
-    models and to provide a much easier to use interface to SyntaxNet.
+    models and to provide a much easier to use interface to SyntaxNet. *DRAGNN
+    is designed first and foremost as a Python library, and therefore much
+    easier to use than the original SyntaxNet implementation.*
+
 1.  **SyntaxNet:
     [code](https://github.com/tensorflow/models/tree/master/syntaxnet/syntaxnet),
     [documentation](g3doc/syntaxnet-tutorial.md)** is a transition-based
@@ -42,7 +46,7 @@ There are three ways to use SyntaxNet:
     SyntaxNet/DRAGNN baseline for the CoNLL2017 Shared Task, and running the
     ParseySaurus models.
 *   You can use DRAGNN to train your NLP models for other tasks and dataset. See
-    "Getting started with DRAGNN below."
+    "Getting started with DRAGNN" below.
 *   You can continue to use the Parsey McParseface family of pre-trained
     SyntaxNet models. See "Pre-trained NLP models" below.
 
@@ -117,9 +121,13 @@ We have a few guides on this README, as well as more extensive
 
 ![DRAGNN](g3doc/unrolled-dragnn.png)
 
-An easy and visual way to get started with DRAGNN is to run [our Jupyter
-Notebook](examples/dragnn/basic_parser_tutorial.ipynb). Our tutorial
+An easy and visual way to get started with DRAGNN is to run our Jupyter
+notebooks for [interactive
+debugging](examples/dragnn/interactive_text_analyzer.ipynb) and [training a new
+model](examples/dragnn/basic_parser_tutorial.ipynb). Our tutorial
 [here](g3doc/CLOUD.md) explains how to start it up from the Docker container.
+Once you have DRAGNN installed and running, try out the
+[ParseySaurus](g3doc/conll2017) models.
 
 ### Using the Pre-trained NLP models
 
@@ -285,6 +293,7 @@ Original authors of the code in this package include (in alphabetical order):
 *   Aliaksei Severyn
 *   Andy Golding
 *   Bernd Bohnet
+*   Chayut Thanapirom
 *   Chris Alberti
 *   Daniel Andor
 *   David Weiss
@@ -294,6 +303,7 @@ Original authors of the code in this package include (in alphabetical order):
 *   Ji Ma
 *   Keith Hall
 *   Kuzman Ganchev
+*   Lingpeng Kong
 *   Livio Baldini Soares
 *   Mark Omernick
 *   Michael Collins

syntaxnet/docker-devel/Dockerfile.min

+# You need to build wheels before building this image. Please consult
+# docker-devel/README.txt.
+
+# This is the base of the openjdk image.
+#
+# It might be more efficient to use a minimal distribution, like Alpine. But
+# the upside of this being popular is that people might already have it.
+FROM buildpack-deps:jessie-curl
+
+ENV SYNTAXNETDIR=/opt/tensorflow PATH=$PATH:/root/bin
+
+RUN apt-get update \
+    && apt-get install -y \
+          file \
+          git \
+          graphviz \
+          libcurl3 \
+          libfreetype6 \
+          libgraphviz-dev \
+          liblapack3 \
+          libopenblas-base \
+          libpng12-0 \
+          libxft2 \
+          python-dev \
+          python-mock \
+          python-pip \
+          python2.7 \
+          zlib1g-dev \
+    && apt-get clean \
+    && (rm -f /var/cache/apt/archives/*.deb \
+        /var/cache/apt/archives/partial/*.deb /var/cache/apt/*.bin || true)
+
+# Install common Python dependencies. Similar to above, remove caches
+# afterwards to help keep Docker images smaller.
+RUN pip install --ignore-installed pip \
+    && python -m pip install numpy \
+    && rm -rf /root/.cache/pip /tmp/pip*
+RUN python -m pip install \
+          asciitree \
+          ipykernel \
+          jupyter \
+          matplotlib \
+          pandas \
+          protobuf \
+          scipy \
+          sklearn \
+    && python -m ipykernel.kernelspec \
+    && python -m pip install pygraphviz \
+          --install-option="--include-path=/usr/include/graphviz" \
+          --install-option="--library-path=/usr/lib/graphviz/" \
+    && rm -rf /root/.cache/pip /tmp/pip*
+
+COPY syntaxnet_with_tensorflow-0.2-cp27-none-linux_x86_64.whl $SYNTAXNETDIR/
+RUN python -m pip install \
+        $SYNTAXNETDIR/syntaxnet_with_tensorflow-0.2-cp27-none-linux_x86_64.whl \
+    && rm -rf /root/.cache/pip /tmp/pip*
+
+# This makes the IP exposed actually "*"; we'll do host restrictions by passing
+# a hostname to the `docker run` command.
+COPY tensorflow/tensorflow/tools/docker/jupyter_notebook_config.py /root/.jupyter/
+EXPOSE 8888
+
+# This does not need to be compiled, only copied.
+COPY examples $SYNTAXNETDIR/syntaxnet/examples
+# For some reason, this works if we run it in a bash shell :/ :/ :/
+CMD /bin/bash -c "python -m jupyter_core.command notebook --debug --notebook-dir=/opt/tensorflow/syntaxnet/examples"

syntaxnet/docker-devel/README.txt

+Docker is used for packaging the SyntaxNet. There are three primary things we
+build with Docker,
+
+1. A development image, which contains all source built with Bazel.
+2. Python/pip wheels, built by running a command in the development container.
+3. A minified image, which only has the compiled version of TensorFlow and
+   SyntaxNet, by installing the wheel built by the above step.
+
+
+Important info (please read)
+------------------------------
+
+One thing to be wary of is that YOU CAN LOSE DATA IF YOU DEVELOP IN A DOCKER
+CONTAINER. Please be very careful to mount data you care about to Docker
+volumes, or use a volume mount so that it's mapped to your host filesystem.
+
+Another note, especially relevant to training models, is that Docker sends the
+whole source tree to the Docker daemon every time you try to build an image.
+This can take some time if you have large temporary model files lying around.
+You can exclude your model files by editing .dockerignore, or just don't store
+them in the base directory.
+
+
+Step 1: Building the development image
+------------------------------
+
+Simply run `docker build -t dragnn-oss .` in the base directory. Make sure you
+have all the source checked out correctly, including git submodules.
+
+
+Step 2: Building wheels
+------------------------------
+
+Please run,
+
+  bash ./docker-devel/build_wheels.sh
+
+This actually builds the image from Step 1 as well.
+
+
+Step 3: Building the development image
+------------------------------
+
+First, ensure you have the file
+
+  syntaxnet_with_tensorflow-0.2-cp27-none-linux_x86_64.whl
+
+in your working directory, from step 2. Then run,
+
+  docker build -t dragnn-oss:latest-minimal -f docker-devel/Dockerfile.min
+
+If the filename changes (e.g. you are on a different architecture), just update
+Dockerfile.min.
+
+
+Developing in Docker
+------------------------------
+
+We recommend developing in Docker by using the `./docker-devel/build_devel.sh`
+script; it will set up a few volume mounts, and port mappings automatically.
+You may want to add more port mappings on your own. If you want to drop into a
+shell instead of launching the notebook, simply run,
+
+  ./docker-devel/build_devel.sh /bin/bash

syntaxnet/docker-devel/build_devel.sh

@@ -23,5 +23,6 @@ syntaxnet_base="/opt/tensorflow/syntaxnet"
 docker run --rm -ti \
   -v "${root_path}"/syntaxnet:"${syntaxnet_base}"/syntaxnet \
   -v "${root_path}"/dragnn:"${syntaxnet_base}"/dragnn \
+  -v "${root_path}"/examples:"${syntaxnet_base}"/examples \
   -p 127.0.0.1:8888:8888 \
   dragnn-oss "$@"

syntaxnet/dragnn/protos/spec.proto

@@ -13,14 +13,10 @@ package syntaxnet.dragnn;
 message MasterSpec {
   repeated ComponentSpec component = 1;
 
-  // DEPRECATED: Use the "batch_size" param of DragnnTensorFlowTrainer instead.
-  optional int32 deprecated_batch_size = 2 [default = 1, deprecated = true];
-
-  // DEPRECATED: Use ComponentSpec.*_beam_size instead.
-  optional int32 deprecated_beam_size = 3 [default = 1, deprecated = true];
-
   // Whether to extract debug traces.
   optional bool debug_tracing = 4 [default = false];
+
+  reserved 2, 3, 5;
 }
 
 // Complete specification for a single task.
@@ -221,10 +217,6 @@ message GridPoint {
   // problems for updates at the start of training.
   optional double gradient_clip_norm = 11 [default = 0.0];
 
-  // DEPRECATED: Use TrainTarget instead.
-  repeated double component_weights = 5;
-  repeated bool unroll_using_oracle = 6;
-
   // A spec for using multiple optimization methods.
   message CompositeOptimizerSpec {
     // First optimizer.
@@ -254,6 +246,8 @@ message GridPoint {
   // should be restricted. If left empty, no filtering will take
   // place. Typically a single component.
   optional string self_norm_components_filter = 21;
+
+  reserved 5, 6;
 }
 
 // Training target to be built into the graph.

syntaxnet/dragnn/python/BUILD

@@ -154,6 +154,7 @@ py_test(
     srcs = ["visualization_test.py"],
     deps = [
         ":visualization",
+        "//dragnn/protos:spec_py_pb2",
         "//dragnn/protos:trace_py_pb2",
         "@org_tensorflow//tensorflow:tensorflow_py",
     ],

syntaxnet/dragnn/python/visualization.py

@@ -54,6 +54,15 @@ def parse_trace_json(trace):
   return as_json
 
 
+def _optional_master_spec_json(master_spec):
+  """Helper function to return 'null' or a master spec JSON string."""
+  if master_spec is None:
+    return 'null'
+  else:
+    return json_format.MessageToJson(
+        master_spec, preserving_proto_field_name=True)
+
+
 def _container_div(height='700px', contents=''):
   elt_id = str(uuid.uuid4())
   html = """
@@ -64,7 +73,11 @@ def _container_div(height='700px', contents=''):
   return elt_id, html
 
 
-def trace_html(trace, convert_to_unicode=True, height='700px', script=None):
+def trace_html(trace,
+               convert_to_unicode=True,
+               height='700px',
+               script=None,
+               master_spec=None):
   """Generates HTML that will render a master trace.
 
   This will result in a self-contained "div" element.
@@ -76,6 +89,8 @@ def trace_html(trace, convert_to_unicode=True, height='700px', script=None):
       often pass the output of this function to IPython.display.HTML.
     height: CSS string representing the height of the element, default '700px'.
     script: Visualization script contents, if the defaults are unacceptable.
+    master_spec: Master spec proto (parsed), which can improve the layout. May
+      be required in future versions.
 
   Returns:
     unicode or str with HTML contents.
@@ -89,10 +104,14 @@ def trace_html(trace, convert_to_unicode=True, height='700px', script=None):
   {div_html}
   <script type='text/javascript'>
   {script}
-  visualizeToDiv({json}, "{elt_id}");
+  visualizeToDiv({json}, "{elt_id}", {master_spec_json});
   </script>
   """.format(
-      script=script, json=json_trace, elt_id=elt_id, div_html=div_html)
+      script=script,
+      json=json_trace,
+      master_spec_json=_optional_master_spec_json(master_spec),
+      elt_id=elt_id,
+      div_html=div_html)
   return unicode(as_str, 'utf-8') if convert_to_unicode else as_str
 
 
@@ -174,11 +193,13 @@ class InteractiveVisualization(object):
         script=script, div_html=div_html)
     return unicode(html, 'utf-8')  # IPython expects unicode.
 
-  def show_trace(self, trace):
+  def show_trace(self, trace, master_spec=None):
     """Returns a JS script HTML fragment, which will populate the container.
 
     Args:
       trace: binary-encoded MasterTrace string.
+      master_spec: Master spec proto (parsed), which can improve the layout. May
+        be required in future versions.
 
     Returns:
       unicode with HTML contents.
@@ -187,8 +208,10 @@ class InteractiveVisualization(object):
     <meta charset="utf-8"/>
     <script type='text/javascript'>
     document.getElementById("{elt_id}").innerHTML = "";  // Clear previous.
-    visualizeToDiv({json}, "{elt_id}");
+    visualizeToDiv({json}, "{elt_id}", {master_spec_json});
     </script>
     """.format(
-        json=parse_trace_json(trace), elt_id=self.elt_id)
+        json=parse_trace_json(trace),
+        master_spec_json=_optional_master_spec_json(master_spec),
+        elt_id=self.elt_id)
     return unicode(html, 'utf-8')  # IPython expects unicode.

syntaxnet/dragnn/python/visualization_test.py

+# -*- coding: utf-8 -*-
 """Tests for dragnn.python.visualization."""
 
 from __future__ import absolute_import
@@ -5,6 +6,7 @@ from __future__ import division
 from __future__ import print_function
 
 from tensorflow.python.platform import googletest
+from dragnn.protos import spec_pb2
 from dragnn.protos import trace_pb2
 from dragnn.python import visualization
 
@@ -15,10 +17,16 @@ def _get_trace_proto_string():
       step_trace=[
           trace_pb2.ComponentStepTrace(fixed_feature_trace=[]),
       ],
-      name='test_component',)
+      # Google Translate says this is "component" in Chinese. (To test UTF-8).
+      name='零件',)
   return trace.SerializeToString()
 
 
+def _get_master_spec():
+  return spec_pb2.MasterSpec(
+      component=[spec_pb2.ComponentSpec(name='jalapeño')])
+
+
 class VisualizationTest(googletest.TestCase):
 
   def testCanFindScript(self):
@@ -37,6 +45,13 @@ class VisualizationTest(googletest.TestCase):
     widget.initial_html()
     widget.show_trace(_get_trace_proto_string())
 
+  def testMasterSpecJson(self):
+    visualization.trace_html(
+        _get_trace_proto_string(), master_spec=_get_master_spec())
+    widget = visualization.InteractiveVisualization()
+    widget.initial_html()
+    widget.show_trace(_get_trace_proto_string(), master_spec=_get_master_spec())
+
 
 if __name__ == '__main__':
   googletest.main()

syntaxnet/dragnn/viz/dragnn_layout.js

@@ -189,6 +189,25 @@ DragnnLayout.prototype.finalLayout = function(partition, stepPartition, cy) {
         Math.sign(slope) * Math.min(300, Math.max(100, Math.abs(slope)));
   });
 
+  // Reset ordering of components based on whether they are actually
+  // left-to-right. In the future, we may want to do the whole layout based on
+  // the master spec (what remains is slope magnitude and component order); then
+  // we can also skip the initial layout and CoSE intermediate layout.
+  if (this.options.masterSpec) {
+    _.each(this.options.masterSpec.component, (component) => {
+      const name = component.name;
+      const transitionParams = component.transition_system.parameters || {};
+      // null/undefined should default to true.
+      const leftToRight = transitionParams.left_to_right != 'false';
+
+      // If the slope isn't going in the direction it should, according to the
+      // master spec, reverse it.
+      if ((leftToRight ? 1 : -1) != Math.sign(stepSlope[name])) {
+        stepSlope[name] = -stepSlope[name];
+      }
+    });
+  }
+
   // Set new node positions. As before, component nodes auto-size to fit.
   _.each(stepPartition, (stepNodes) => {
     const component = _.head(stepNodes).data('parent');

syntaxnet/dragnn/viz/visualize.js

@@ -154,10 +154,13 @@ class InteractiveDragnnGraph {
    *
    * @param {!Object} masterTrace Master trace proto from DRAGNN.
    * @param {!Object} element Container DOM element to populate.
+   * @param {?Object} masterSpec Master spec proto from DRAGNN; if provided,
+   *     used to improve the layout.
    */
-  constructor(masterTrace, element) {
+  constructor(masterTrace, element, masterSpec) {
     this.masterTrace = masterTrace;
     this.element = element;
+    this.masterSpec = masterSpec || null;
   }
 
   /**
@@ -196,7 +199,7 @@ class InteractiveDragnnGraph {
     sel.abscomp().nodes().hide();
 
     // Redo layout.
-    this.cy.layout({name: 'dragnn'});
+    this.cy.layout({name: 'dragnn', masterSpec: this.masterSpec});
   }
 
   /**
@@ -211,7 +214,7 @@ class InteractiveDragnnGraph {
       boxSelectionEnabled: true,
       autounselectify: true,
       // We'll do more custom layout later.
-      layout: {name: 'dragnn'},
+      layout: {name: 'dragnn', masterSpec: this.masterSpec},
       style: [
         {
           selector: 'node',
@@ -285,12 +288,15 @@ class InteractiveDragnnGraph {
  * situations, the script tag containing the graph definition will be generated
  * inline.
  *
- * @param {Object} masterTrace Master trace proto from DRAGNN.
+ * @param {!Object} masterTrace Master trace proto from DRAGNN.
  * @param {string} divId ID of the page element to populate with the graph.
+ * @param {?Object} masterSpec Master spec proto from DRAGNN; if provided, used
+ *     to improve the layout.
  */
-const visualizeToDiv = function(masterTrace, divId) {
-  const interactiveGraph =
-      new InteractiveDragnnGraph(masterTrace, document.getElementById(divId));
+
+const visualizeToDiv = function(masterTrace, divId, masterSpec) {
+  const interactiveGraph = new InteractiveDragnnGraph(
+      masterTrace, document.getElementById(divId), masterSpec);
   interactiveGraph.initDomElements();
 };
 

syntaxnet/examples/dragnn/BUILD

+py_binary(
+    name = "tutorial_1",
+    srcs = ["tutorial_1.py"],
+    data = [":data"],
+    deps = [":tutorial-deps"],
+)
+
+py_binary(
+    name = "tutorial_2",
+    srcs = ["tutorial_2.py"],
+    data = [":data"],
+    deps = [":tutorial-deps"],
+)
+
+py_library(
+    name = "tutorial-deps",
+    deps = [
+        "//dragnn/core:dragnn_bulk_ops",
+        "//dragnn/core:dragnn_ops",
+        "//dragnn/protos:spec_py_pb2",
+        "//dragnn/python:graph_builder",
+        "//dragnn/python:lexicon",
+        "//dragnn/python:load_dragnn_cc_impl_py",
+        "//dragnn/python:spec_builder",
+        "//dragnn/python:visualization",
+        "//syntaxnet:load_parser_ops_py",
+        "//syntaxnet:parser_ops",
+        "//syntaxnet:sentence_py_pb2",
+        "@org_tensorflow//tensorflow:tensorflow_py",
+        "@org_tensorflow//tensorflow/core:protos_all_py",
+    ],
+)
+
+filegroup(
+    name = "data",
+    data = glob(["tutorial_data/*"]),
+)
+
+sh_test(
+    name = "test_run_all_tutorials",
+    size = "medium",
+    srcs = ["test_run_all_tutorials.sh"],
+    args = [
+        "$(location :tutorial_1)",
+        "$(location :tutorial_2)",
+    ],
+    data = [
+        ":tutorial_1",
+        ":tutorial_2",
+    ],
+)

syntaxnet/examples/dragnn/data/en/category-map

+0

syntaxnet/examples/dragnn/data/en/char-map

+99
+e 88088
+t 63849
+a 60376
+o 55522
+n 50760
+i 49620
+s 43451
+r 43314
+h 34531
+l 30483
+d 27017
+u 20955
+c 19696
+m 17547
+y 15363
+f 14604
+g 14285
+p 13521
+w 13380
+9 11333
+. 10880
+b 10234
+v 7728
+, 6776
+k 6200
+I 5027
+- 3666
+T 3262
+A 2967
+S 2823
+' 2382
+C 1954
+E 1763
+M 1748
+P 1662
+B 1457
+x 1455
+N 1420
+W 1285
+H 1248
+" 1233
+D 1214
+O 1212
+R 1153
+! 1135
+/ 1124
+L 1094
+: 1067
+j 924
+? 915
+) 892
+F 887
+G 875
+q 873
+( 827
+U 709
+J 660
+Y 588
+z 566
+_ 539
+K 499
+V 372
+= 369
+* 303
+$ 254
+@ 177
+& 155
+> 151
+< 143
+Q 142
+; 100
+’ 94
+Z 92
+X 73
+# 69
++ 51
+% 39
+[ 34
+] 34
+“ 30
+” 30
+| 25
+~ 17
+` 15
+‘ 13
+– 9
+— 9
+^ 8
+… 7
+· 6
+{ 4
+} 3
+é 2
+£ 1
+­ 1
+³ 1
+Ã 1
+á 1
+ç 1