skflow load pre-train word embeddings into text_classification_cnn

Question

word_vectors = skflow.ops.categorical_variable(X, n_classes=n_words,
            embedding_size=EMBEDDING_SIZE, name='words')

word_vectors = tf.expand_dims(word_vectors, 3)

this is a skflow example on convolutional text classification

when I debug this patch of code, i can't explain how it works. How to use pre-trained word embeddings instead of it?