word2vec: self.embedding for restoration in the future

Author: JasonShin

032-word2vec/model.py

@@ -6,6 +6,7 @@ class Word2vec:
  embedding_size = 2 # word embedding size
  num_sampled = 15 # Number of negative examples to sample.
  vocab_size = 0 # Vocab size created from the data API
+ embedding = None
  gpu_dynamic_memory_growth = False
  loss = None # TF Loss
  optimizer = None # Optimizer
@@ -33,8 +34,8 @@ def compile(self):
  # Look up embeddings for inputs.
  # The conversion of 10,000 columned matrix into a 200 columned matrix is called word embedding.
  with tf.device("/cpu:0"):
- embeddings = tf.Variable(tf.random_uniform([self.vocab_size, self.embedding_size], -1.0, 1.0))
- embed = tf.nn.embedding_lookup(embeddings, self.X) # lookup table
+ self.embeddings = tf.Variable(tf.random_uniform([self.vocab_size, self.embedding_size], -1.0, 1.0))
+ embed = tf.nn.embedding_lookup(self.embeddings, self.X) # lookup table
 
  # Construct the variables for the NCE loss
  nce_weights = tf.Variable(tf.random_uniform([self.vocab_size, self.embedding_size], -1.0, 1.0))