Converting tokens to word vectors effectively with TensorFlow Transform

Question

I would like to use TensorFlow Transform to convert tokens to word vectors during my training, validation and inference phase.

I followed this StackOverflow post and implemented the initial conversion from tokens to vectors. The conversion works as expected and I obtain vectors of EMB_DIM for each token.

import numpy as np
import tensorflow as tf

tf.reset_default_graph()
EMB_DIM = 10

def load_pretrained_glove():
    tokens = ["a", "cat", "plays", "piano"]
    return tokens, np.random.rand(len(tokens), EMB_DIM)

# sample string 
string_tensor = tf.constant(["plays", "piano", "unknown_token", "another_unknown_token"])


pretrained_vocab, pretrained_embs = load_pretrained_glove()

vocab_lookup = tf.contrib.lookup.index_table_from_tensor(
    mapping = tf.constant(pretrained_vocab),
    default_value = len(pretrained_vocab))
string_tensor = vocab_lookup.lookup(string_tensor)

# define the word embedding
pretrained_embs = tf.get_variable(
    name="embs_pretrained",
    initializer=tf.constant_initializer(np.asarray(pretrained_embs), dtype=tf.float32),
    shape=pretrained_embs.shape,
    trainable=False)

unk_embedding = tf.get_variable(
    name="unk_embedding",
    shape=[1, EMB_DIM],
    initializer=tf.random_uniform_initializer(-0.04, 0.04),
    trainable=False)

embeddings = tf.cast(tf.concat([pretrained_embs, unk_embedding], axis=0), tf.float32)
word_vectors = tf.nn.embedding_lookup(embeddings, string_tensor)

with tf.Session() as sess:
    tf.tables_initializer().run()
    tf.global_variables_initializer().run()
    print(sess.run(word_vectors))

When I refactor the code to run as a TFX Transform Graph, I am getting the error the ConversionError below.

import pprint
import tempfile
import numpy as np
import tensorflow as tf
import tensorflow_transform as tft
import tensorflow_transform.beam.impl as beam_impl
from tensorflow_transform.tf_metadata import dataset_metadata
from tensorflow_transform.tf_metadata import dataset_schema

tf.reset_default_graph()

EMB_DIM = 10

def load_pretrained_glove():
    tokens = ["a", "cat", "plays", "piano"]
    return tokens, np.random.rand(len(tokens), EMB_DIM)


def embed_tensor(string_tensor, trainable=False):
    """
    Convert List of strings into list of indices then into EMB_DIM vectors
    """

    pretrained_vocab, pretrained_embs = load_pretrained_glove()

    vocab_lookup = tf.contrib.lookup.index_table_from_tensor(
        mapping=tf.constant(pretrained_vocab),
        default_value=len(pretrained_vocab))
    string_tensor = vocab_lookup.lookup(string_tensor)

    pretrained_embs = tf.get_variable(
        name="embs_pretrained",
        initializer=tf.constant_initializer(np.asarray(pretrained_embs), dtype=tf.float32),
        shape=pretrained_embs.shape,
        trainable=trainable)
    unk_embedding = tf.get_variable(
        name="unk_embedding",
        shape=[1, EMB_DIM],
        initializer=tf.random_uniform_initializer(-0.04, 0.04),
        trainable=False)

    embeddings = tf.cast(tf.concat([pretrained_embs, unk_embedding], axis=0), tf.float32)
    return tf.nn.embedding_lookup(embeddings, string_tensor)

def preprocessing_fn(inputs):
    input_string = tf.string_split(inputs['sentence'], delimiter=" ") 
    return {'word_vectors': tft.apply_function(embed_tensor, input_string)}


raw_data = [{'sentence': 'This is a sample sentence'},]
raw_data_metadata = dataset_metadata.DatasetMetadata(dataset_schema.Schema({
  'sentence': dataset_schema.ColumnSchema(
      tf.string, [], dataset_schema.FixedColumnRepresentation())
}))

with beam_impl.Context(temp_dir=tempfile.mkdtemp()):
    transformed_dataset, transform_fn = (  # pylint: disable=unused-variable
        (raw_data, raw_data_metadata) | beam_impl.AnalyzeAndTransformDataset(
            preprocessing_fn))

    transformed_data, transformed_metadata = transformed_dataset  # pylint: disable=unused-variable
    pprint.pprint(transformed_data)

Error Message

TypeError: Failed to convert object of type <class 
'tensorflow.python.framework.sparse_tensor.SparseTensor'> to Tensor. 
Contents: SparseTensor(indices=Tensor("StringSplit:0", shape=(?, 2), 
dtype=int64), values=Tensor("hash_table_Lookup:0", shape=(?,), 
dtype=int64), dense_shape=Tensor("StringSplit:2", shape=(2,), 
dtype=int64)). Consider casting elements to a supported type.

Questions

1. Why would the TF Transform step require an additional conversion/casting?
2. Is this approach of converting tokens to word vectors feasible? The word vectors might be multiple gigabytes in memory. How is Apache Beam handling the vectors? If Beam in a distributed setup, would it require N x vector memory with N the number of workers?

Answer 1

One cannot put a SparseTensor into the dictionary, returned by the TFX Transform, in your case by the function "preprocessing_fn". The reason is that SparseTensor is not a Tensor, it is actually a small subgraph.

To fix your code, you can convert your SparseTensor into a Tensor. There is a number of ways to do so, I would recommend to use tf.serialize_sparse for regular SparseTensor and tf.serialize_many_sparse for batched one.

To consume such serialized Tensor in Trainer, you could call the function tf. deserialize_many_sparse.

Answer 2

The SparseTensor related error is because you are calling string_split which returns a SparseTensor. Your test code does not call string_split so that's why it only happens with your Transform code.

Regarding memory, you are correct, the embedding matrix must be loaded into each worker.