kevinrobinson

# Entire graph
a = tf.constant([1, 2, 3], name="a")
b = tf.placeholder(name="b")
c = tf.mul(a, b, name="c")
# d = ...
# e = ...
f = tf.add(c, 4, name="f")
with tf.Session() as session:
  print session.run(f, feed_dict={ b: 10 })
  # >>> [14 24 34]

kevinrobinson

TEST_F(SubgraphTest, Targets1) {
  ExpectOK(
      "node { name: 'W1' op: 'TestParams' }"
      "node { name: 'W2' op: 'TestParams' }"
      "node { name: 'input' op: 'TestInput' }"
      "node { name: 't1' op: 'TestMul' input: [ 'W1', 'input:1' ] }"
      "node { name: 't2' op: 'TestMul' input: [ 'W2', 't1' ] }"
      "node { name: 't3_a' op: 'TestRelu' input: 't2' }"
      "node { name: 't3_b' op: 'TestRelu' input: 't2' }");
  EXPECT_EQ("OK", Subgraph("", "", "t1"));

kevinrobinson

$ python embedding/word2vec_optimized.py --train_data=data/text8 --save_path=fiddling-one --eval_data=data/questions-words.txt --epochs_to_train=1
.../local_device.cc:40] Local device intra op parallelism threads: 8
.../direct_session.cc:60] Direct session inter op parallelism threads: 8

kevinrobinson

// Test that a graph with no constraints will successfully assign nodes to the
// "best available" device (i.e. prefer GPU over CPU).
TEST_F(SimplePlacerTest, TestNoConstraints) {
  Graph g(OpRegistry::Global());
  {  // Scope for temporary variables used to construct g.
    GraphDefBuilder b(GraphDefBuilder::kFailImmediately);
    Node* input = ops::SourceOp("TestInput", b.opts().WithName("in"));
    ops::UnaryOp("TestRelu", ops::NodeOut(input, 0), b.opts().WithName("n1"));
    ops::UnaryOp("TestRelu", ops::NodeOut(input, 1), b.opts().WithName("n2"));
    EXPECT_OK(BuildGraph(b, &g));

kevinrobinson

.../simple_placer.cc:304] global_step: /job:localhost/replica:0/task:0/cpu:0
.../simple_placer.cc:304] Skipgram: /job:localhost/replica:0/task:0/cpu:0
...

kevinrobinson

tuple = word2vec.skipgram(filename=opts.train_data,
                          batch_size=opts.batch_size,
                          window_size=opts.window_size,
                          min_count=opts.min_count,
                          subsample=opts.subsample)
(vocab_word, vocab_freq, words_per_epoch, current_epoch, total_words_processed, examples, labels) = tuple

kevinrobinson

def skipgram(filename, batch_size, window_size=None, min_count=None,
             subsample=None, name=None):
  r"""Parses a text file and creates a batch of examples.
  ...
  """
  return _op_def_lib.apply_op("Skipgram", filename=filename,
                              batch_size=batch_size, window_size=window_size,
                              min_count=min_count, subsample=subsample,
                              name=name)

kevinrobinson

def _InitOpDefLibrary():
  op_list = op_def_pb2.OpList()
  text_format.Merge(op_list_ascii, op_list)
  op_def_registry.register_op_list(op_list)
  op_def_lib = op_def_library.OpDefLibrary()
  op_def_lib.add_op_list(op_list)
  return op_def_lib


_InitOpDefLibrary.op_list_ascii = """op {

kevinrobinson

# build
bazel build -c dbg tensorflow/models/embedding:all
# copy built gen_word2vec.py file from bazel-out/../genfiles into the project

# in word2vec.py, comment out the division line to fix python 2/3 error:
# error:  TypeError: unsupported operand type(s) for /: 'Tensor' and 'int'
# fix: comment out  from __future__ import division

# in word2vec.py, change the import to pull in gen_word2vec from the local folder:
# before: from tensorflow.models.embedding import gen_word2vec as word2vec

kevinrobinson

class SkipgramOp : public OpKernel {
 public:
  explicit SkipgramOp(OpKernelConstruction* ctx) : OpKernel(ctx) {
    // validate input and save it as state
    OP_REQUIRES_OK(ctx, ctx->GetAttr("filename", &filename));
    // ...
  }

  void Compute(OpKernelContext* ctx) override {
    Tensor examples(DT_INT32, TensorShape({batch_size_}));