research/syntaxnet/examples/dragnn/trainer_tutorial.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true,
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "import os\n",
    "import os.path\n",
    "import random\n",
    "import time\n",
    "import tensorflow as tf\n",
    "\n",
    "from IPython.display import HTML\n",
    "from tensorflow.python.platform import gfile\n",
    "from tensorflow.python.platform import tf_logging as logging\n",
    "\n",
    "from google.protobuf import text_format\n",
    "\n",
    "from syntaxnet.ops import gen_parser_ops\n",
    "from syntaxnet import load_parser_ops  # This loads the actual op definitions\n",
    "from syntaxnet import task_spec_pb2\n",
    "from syntaxnet import sentence_pb2\n",
    "\n",
    "from dragnn.protos import spec_pb2\n",
    "from dragnn.python.sentence_io import ConllSentenceReader\n",
    "\n",
    "from dragnn.python import evaluation\n",
    "from dragnn.python import graph_builder\n",
    "from dragnn.python import lexicon\n",
    "from dragnn.python import load_dragnn_cc_impl\n",
    "from dragnn.python import render_parse_tree_graphviz\n",
    "from dragnn.python import render_spec_with_graphviz\n",
    "from dragnn.python import spec_builder\n",
    "from dragnn.python import trainer_lib\n",
    "from dragnn.python import visualization\n",
    "\n",
    "DATA_DIR = '/opt/tensorflow/syntaxnet/examples/dragnn/data/es'\n",
    "TENSORBOARD_DIR = '/notebooks/tensorboard'\n",
    "CHECKPOINT_FILENAME = '{}/spanish.checkpoint'.format(DATA_DIR)\n",
    "TRAINING_CORPUS_PATH = '{}/es-universal-train.conll'.format(DATA_DIR)\n",
    "DEV_CORPUS_PATH = '{}/es-universal-dev.conll'.format(DATA_DIR)\n",
    "\n",
    "# Some of the IO functions fail miserably if data is missing.\n",
    "assert os.path.isfile(TRAINING_CORPUS_PATH), 'Could not find training corpus'\n",
    "\n",
    "logging.set_verbosity(logging.WARN)\n",
    "\n",
    "# Constructs lexical resources for SyntaxNet in the given resource path, from\n",
    "# the training data.\n",
    "lexicon.build_lexicon(DATA_DIR, TRAINING_CORPUS_PATH)\n",
    "\n",
    "# Construct the 'lookahead' ComponentSpec. This is a simple right-to-left RNN\n",
    "# sequence model, which encodes the context to the right of each token. It has\n",
    "# no loss except for the downstream components.\n",
    "lookahead = spec_builder.ComponentSpecBuilder('lookahead')\n",
    "lookahead.set_network_unit(\n",
    "    name='FeedForwardNetwork', hidden_layer_sizes='256')\n",
    "lookahead.set_transition_system(name='shift-only', left_to_right='true')\n",
    "lookahead.add_fixed_feature(name='words', fml='input.word', embedding_dim=64)\n",
    "lookahead.add_rnn_link(embedding_dim=-1)\n",
    "lookahead.fill_from_resources(DATA_DIR)\n",
    "\n",
    "# Construct the ComponentSpec for tagging. This is a simple left-to-right RNN\n",
    "# sequence tagger.\n",
    "tagger = spec_builder.ComponentSpecBuilder('tagger')\n",
    "tagger.set_network_unit(name='FeedForwardNetwork', hidden_layer_sizes='256')\n",
    "tagger.set_transition_system(name='tagger')\n",
    "tagger.add_rnn_link(embedding_dim=-1)\n",
    "tagger.add_token_link(source=lookahead, fml='input.focus', embedding_dim=32)\n",
    "tagger.fill_from_resources(DATA_DIR)\n",
    "\n",
    "# Construct the ComponentSpec for parsing.\n",
    "parser = spec_builder.ComponentSpecBuilder('parser')\n",
    "parser.set_network_unit(name='FeedForwardNetwork', hidden_layer_sizes='256')\n",
    "parser.set_transition_system(name='arc-standard')\n",
    "parser.add_token_link(source=lookahead, fml='input.focus', embedding_dim=32)\n",
    "parser.add_token_link(\n",
    "    source=tagger,\n",
    "    fml='input.focus stack.focus stack(1).focus',\n",
    "    embedding_dim=32)\n",
    "\n",
    "# Recurrent connection for the arc-standard parser. For both tokens on the\n",
    "# stack, we connect to the last time step to either SHIFT or REDUCE that\n",
    "# token. This allows the parser to build up compositional representations of\n",
    "# phrases.\n",
    "parser.add_link(\n",
    "    source=parser,                          # recurrent connection\n",
    "    name='rnn-stack',                       # unique identifier\n",
    "    fml='stack.focus stack(1).focus',       # look for both stack tokens\n",
    "    source_translator='shift-reduce-step',  # maps token indices -> step\n",
    "    embedding_dim=32)                       # project down to 32 dims\n",
    "\n",
    "parser.fill_from_resources(DATA_DIR)\n",
    "\n",
    "master_spec = spec_pb2.MasterSpec()\n",
    "master_spec.component.extend([lookahead.spec, tagger.spec, parser.spec])\n",
    "HTML(render_spec_with_graphviz.master_spec_graph(master_spec))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# Build the TensorFlow graph based on the DRAGNN network spec.\n",
    "graph = tf.Graph()\n",
    "with graph.as_default():\n",
    "    hyperparam_config = spec_pb2.GridPoint(\n",
    "        learning_method='adam',\n",
    "        learning_rate=0.0005, \n",
    "        adam_beta1=0.9, adam_beta2=0.9, adam_eps=0.001,\n",
    "        dropout_rate=0.8, gradient_clip_norm=1,\n",
    "        use_moving_average=True,\n",
    "        seed=1)\n",
    "    builder = graph_builder.MasterBuilder(master_spec, hyperparam_config)\n",
    "    target = spec_pb2.TrainTarget(\n",
    "        name='all',\n",
    "        unroll_using_oracle=[False, True, True], # train tagger & parser on gold unrolling, skip lookahead\n",
    "        component_weights=[0, 0.5, 0.5]) # tagger and parser losses have equal weights\n",
    "    trainer = builder.add_training_from_config(target)\n",
    "    annotator = builder.add_annotation(enable_tracing=True)\n",
    "    builder.add_saver()\n",
    "\n",
    "# Train on Spanish data for N_STEPS steps and evaluate.\n",
    "N_STEPS = 20\n",
    "BATCH_SIZE = 64\n",
    "with tf.Session(graph=graph) as sess:\n",
    "    sess.run(tf.global_variables_initializer())\n",
    "    training_corpus = ConllSentenceReader(\n",
    "        TRAINING_CORPUS_PATH, projectivize=True).corpus()\n",
    "    dev_corpus = ConllSentenceReader(DEV_CORPUS_PATH).corpus()[:200]\n",
    "    for step in xrange(N_STEPS):\n",
    "        trainer_lib.run_training_step(sess, trainer, training_corpus, batch_size=BATCH_SIZE)\n",
    "        tf.logging.warning('Step %d/%d', step + 1, N_STEPS)\n",
    "    parsed_dev_corpus = trainer_lib.annotate_dataset(sess, annotator, dev_corpus)\n",
    "    pos, uas, las = evaluation.calculate_parse_metrics(dev_corpus, parsed_dev_corpus)\n",
    "    tf.logging.warning('POS %.2f UAS %.2f LAS %.2f', pos, uas, las)\n",
    "    builder.saver.save(sess, CHECKPOINT_FILENAME)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# Visualize the output of our mini-trained model on a test sentence.\n",
    "\n",
    "text = '¿ Viste ese gran coche rojo ?'\n",
    "tokens = [sentence_pb2.Token(word=word, start=-1, end=-1) for word in text.split()]\n",
    "sentence = sentence_pb2.Sentence()\n",
    "sentence.token.extend(tokens)\n",
    "\n",
    "with tf.Session(graph=graph) as sess:\n",
    "    # Restore the model we just trained.\n",
    "    builder.saver.restore(sess, CHECKPOINT_FILENAME)\n",
    "    annotations, traces = sess.run([annotator['annotations'], annotator['traces']],\n",
    "                      feed_dict={annotator['input_batch']: [sentence.SerializeToString()]})\n",
    "\n",
    "HTML(visualization.trace_html(traces[0]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "parsed_sentence = sentence_pb2.Sentence.FromString(annotations[0])\n",
    "HTML(render_parse_tree_graphviz.parse_tree_graph(parsed_sentence))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 2",
   "language": "python",
   "name": "python2"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}