Tensorflow: Input pipeline with sparse data for the SVM estimator

Question

Introduction:

I am trying to train the tensorflow svm estimator tensorflow.contrib.learn.python.learn.estimators.svm with sparse data. Sample usage with sparse data at the github repo at tensorflow/contrib/learn/python/learn/estimators/svm_test.py#L167 (I am not allowed to post more links, so here the relative path).

The svm estimator expects as parameter example_id_column and feature_columns, where the feature columns should be derived of class FeatureColumn such as tf.contrib.layers.feature_column.sparse_column_with_hash_bucket. See Github repo at tensorflow/contrib/learn/python/learn/estimators/svm.py#L85 and the documentation at tensorflow.org at python/contrib.layers#Feature_columns.

Question:

1. How do I have to set up my input pipeline to format sparse data in such a way that I can use one of the tf.contrib.layers feature_columns as input for the svm estimator.
2. How would a dense input function with many features look like?

Background

The data that I use is the a1a dataset from the LIBSVM website. The data set has 123 features (that would correspond to 123 feature_columns if the data would be dense). I wrote an user op to read the data like tf.decode_csv() but for the LIBSVM format. The op returns the labels as dense tensor and the features as sparse tensor. My input pipeline:

NUM_FEATURES = 123
batch_size = 200

# my op to parse the libsvm data
decode_libsvm_module = tf.load_op_library('./libsvm.so')

def input_pipeline(filename_queue, batch_size):
    with tf.name_scope('input'):
        reader = tf.TextLineReader(name="TextLineReader_")
        _, libsvm_row = reader.read(filename_queue, name="libsvm_row_")
        min_after_dequeue = 1000
        capacity = min_after_dequeue + 3 * batch_size
        batch = tf.train.shuffle_batch([libsvm_row], batch_size=batch_size,
                                       capacity=capacity,
                                       min_after_dequeue=min_after_dequeue,
                                       name="text_line_batch_")
        labels, sp_indices, sp_values, sp_shape = \
            decode_libsvm_module.decode_libsvm(records=batch,
                                               num_features=123,
                                               OUT_TYPE=tf.int64, 
                                               name="Libsvm_decoded_")
        # Return the features as sparse tensor and the labels as dense
        return tf.SparseTensor(sp_indices, sp_values, sp_shape), labels

Here is an example batch with batch_size = 5.

def input_fn(dataset_name):
    maybe_download()

    filename_queue_train = tf.train.string_input_producer([dataset_name], 
                                                        name="queue_t_")
    features, labels = input_pipeline(filename_queue_train, batch_size)

    return {
        'example_id': tf.as_string(tf.range(1,123,1,dtype=tf.int64)),
        'features': features
    }, labels

This is what I tried so far:

with tf.Session().as_default() as sess:
    sess.run(tf.global_variables_initializer())

    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(sess=sess, coord=coord)

    feature_column = tf.contrib.layers.sparse_column_with_hash_bucket(
        'features', hash_bucket_size=1000, dtype=tf.int64)

    svm_classifier = svm.SVM(feature_columns=[feature_column],
                             example_id_column='example_id',
                             l1_regularization=0.0,
                             l2_regularization=1.0)
    svm_classifier.fit(input_fn=lambda: input_fn(TRAIN),
                       steps=30)

    accuracy = svm_classifier.evaluate(
        input_fn= lambda: input_fn(features, labels), 
        steps=1)['accuracy']                       
    print(accuracy)
    coord.request_stop()

    coord.join(threads)
    sess.close()

Answer 1

Here's an example, with made up data, that works for me in TensorFlow 1.1.0-rc2. I think my comment was misleading; you're best off converting ~100 binary features to real valued features (tf.sparse_tensor_to_dense) and using a real_valued_column, since sparse_column_with_integerized_feature is hiding most of the useful information from the SVM Estimator.

import tensorflow as tf

batch_size = 10
num_features = 123
num_examples = 100

def input_fn():
  example_ids = tf.random_uniform(
      [batch_size], maxval=num_examples, dtype=tf.int64)
  # Construct a SparseTensor with features
  dense_features = (example_ids[:, None]
                    + tf.range(num_features, dtype=tf.int64)[None, :]) % 2
  non_zeros = tf.where(tf.not_equal(dense_features, 0))
  sparse_features = tf.SparseTensor(
      indices=non_zeros,
      values=tf.gather_nd(dense_features, non_zeros),
      dense_shape=[batch_size, num_features])
  features = {
      'some_sparse_features': tf.sparse_tensor_to_dense(sparse_features),
      'example_id': tf.as_string(example_ids)}
  labels = tf.equal(dense_features[:, 0], 1)
  return features, labels
svm = tf.contrib.learn.SVM(
    example_id_column='example_id',
    feature_columns=[
      tf.contrib.layers.real_valued_column(
          'some_sparse_features')],
    l2_regularization=0.1, l1_regularization=0.5)
svm.fit(input_fn=input_fn, steps=1000)
positive_example = lambda: {
    'some_sparse_features': tf.sparse_tensor_to_dense(
        tf.SparseTensor([[0, 0]], [1], [1, num_features])),
    'example_id': ['a']}
print(svm.evaluate(input_fn=input_fn, steps=20))
print(next(svm.predict(input_fn=positive_example)))
negative_example = lambda: {
    'some_sparse_features': tf.sparse_tensor_to_dense(
        tf.SparseTensor([[0, 0]], [0], [1, num_features])),
    'example_id': ['b']}
print(next(svm.predict(input_fn=negative_example)))

Prints:

{'accuracy': 1.0, 'global_step': 1000, 'loss': 1.0645389e-06}
{'logits': array([ 0.01612902], dtype=float32), 'classes': 1}
{'logits': array([ 0.], dtype=float32), 'classes': 0}