Htykuut

I am currently trying to use my own pre-trained word embeddings in my tensorflow model_fn, but after a single epoch, the model building fails due to the following error;

ValueError: Fetch argument <tf.Variable 'embedding/embeddings:0' shape=(86565, 300) dtype=float32_ref> cannot be interpreted a

s a Tensor. (Tensor Tensor("embedding/embeddings:0", shape=(86565, 300), dtype=float32_ref) is not an element of this graph.)

How do I modify my code in order to maintain the embeddings in the graph over several epochs? Should I be initializing them in a different way? Or outside of the model function?

def word_embeddings_matrix():



    ..... load my embeddings .....



    return embedding_matrix





embedding_matrix = word_embeddings_matrix()





def model_fn(features, labels, mode, params):



    vocab_table = lookup.index_table_from_file(vocabulary_file='dataset/vocab.csv', num_oov_buckets=1, default_value=-1)

    text = features[commons.FEATURE_COL]

    words = tf.string_split(text)

    dense_words = tf.sparse_tensor_to_dense(words, default_value=commons.PAD_WORD)

    word_ids = vocab_table.lookup(dense_words)



    padding = tf.constant([[0, 0], [0, commons.MAX_DOCUMENT_LENGTH]])

    # Pad all the word_ids entries to the maximum document length

    word_ids_padded = tf.pad(word_ids, padding)

    word_id_vector = tf.slice(word_ids_padded, [0, 0], [-1, commons.MAX_DOCUMENT_LENGTH])



    embedding_matrix = word_embeddings_matrix()



    if mode == tf.estimator.ModeKeys.TRAIN:

        tf.keras.backend.set_learning_phase(True)

    else:

        tf.keras.backend.set_learning_phase(False)



    # embedded_sequences = tf.keras.layers.Embedding(params.N_WORDS,

    #                       10, input_length=commons.MAX_DOCUMENT_LENGTH)(word_id_vector)



    embedded_sequences = tf.keras.layers.Embedding(params.N_WORDS,

                                                   300,

                                                   weights=[embedding_matrix],

                                                   input_length=commons.MAX_DOCUMENT_LENGTH,

                                                   trainable=True)(word_id_vector)





    conv = tf.keras.layers.Conv1D(filters=64, kernel_size=2, activation='relu')(embedded_sequences)

    pool = tf.keras.layers.GlobalAveragePooling1D()(conv)

    drop = tf.keras.layers.Dropout(0.5)(pool)



    logits = tf.keras.layers.Dense(commons.TARGET_SIZE, activation=None)(drop)



    predictions = tf.nn.softmax(logits)

    prediction_indices = tf.argmax(predictions, axis=1)