def setup_training(self, pretrained_path):
"""
Augment word embedding matrix in `pretrained_path` with
randomly intialized ones for those words that are in self.words
but not in the original embedding matrix and initialize self.emb_word
with it. Write out model definitions to `tagger_defs.txt`
Inputs:
pretrained_path (str): path to pretrained embedding file
"""
emb_w = read_pretrained_embeddings(pretrained_path)
new_emb_w = 0.02 * np.random.random_sample(
(len(self.words), emb_w.shape[1])).astype('f') - 0.01
for i in range(len(emb_w)):
new_emb_w[i] = emb_w[i]
self.emb_word.W.data = new_emb_w
with open(os.path.join(self.model_path, "tagger_defs.txt"), "w") as out:
json.dump({"model": self.__class__.__name__,
"word_dim": self.word_dim,
"suffix_dim": self.suffix_dim,
"caps_dim": self.caps_dim}, out)
def load_pretrained_embeddings(self, path):
self.emb_word.W.data = read_pretrained_embeddings(path)
