kevinrobinson

py_binary(
  name = “word2vec”,
  srcs = [“word2vec.py”],
  deps = [“:gen_word2vec”, ...],
  ...
)

kevinrobinson

native.py_library(
    name = "gen_word2vec",
    src = ["gen_word2vec.py"]
    ...
)

kevinrobinson

native.genrule(
    name = "gen_word2vec_pygenrule",
    tools = ["gen_gen_word2vec_py_wrappers_cc"],
    out = ["gen_word2vec.py"],
    ...
)

kevinrobinson

native.cc_binary(
    name = "gen_gen_word2vec_py_wrappers_cc",
    deps = [":word2vec_ops", ...],
    ...
)

kevinrobinson

tf_gen_op_wrapper_py(
    name = "gen_word2vec",
    out = "gen_word2vec.py",
    deps = [":word2vec_ops"],
)

kevinrobinson

node {
  name: "Cast_1"
  op: "Cast"
  input: "Skipgram:4"
  attr {
    key: "DstT"
    value {
      type: DT_FLOAT
    }
  }

kevinrobinson

lr = opts.learning_rate * tf.maximum(0.0001, 1.0 — tf.cast(self._words, tf.float32) / words_to_train)

kevinrobinson

REGISTER_OP(“Skipgram”)
 .Output(“vocab_word: string”)
 .Output(“vocab_freq: int32”)
 .Output(“words_per_epoch: int64”)
 .Output(“current_epoch: int32”)
 .Output(“total_words_processed: int64”)
 .Output(“examples: int32”)
 .Output(“labels: int32”)
 .Attr(“filename: string”)
 .Attr(“batch_size: int”)

kevinrobinson

(words, counts, words_per_epoch, self._epoch, self._words, examples, labels) = word2vec.skipgram(filename=opts.train_data,
  batch_size=opts.batch_size,
  window_size=opts.window_size,
  min_count=opts.min_count,
  subsample=opts.subsample)

kevinrobinson

# example input:
sentence = 'The quick brown fox jumped over the lazy dog.'

# tokenize and normalize words, building the set of all vocabulary ever seen
words_set = {}
def tokenize(sentence):
  return map(str.lower, sentence[0:-1].split(' '))
for word in tokenize(sentence):
  words_set[word] = True