How to load sparse data with TensorFlow?

Question

There is a small snippet about loading sparse data but I have no idea how to use it.

SparseTensors don\'t play well with queues. If you use SparseTensors

Answer 1

For libsvm format you can write and parse like below, if you want sparse tensor result(as opposed to dense tensor result using padding strategy)

    #---write
    _float_feature = lambda v: tf.train.Feature(float_list=tf.train.FloatList(value=v))
    _int_feature = lambda v: tf.train.Feature(int64_list=tf.train.Int64List(value=v))

    indexes = []
    values = []

    for item in l[start:]:
      index,value = item.split(':')
      indexes.append(int(index))
      values.append(float(value))

    example = tf.train.Example(features=tf.train.Features(feature={
      'label': _int_feature([label]),
      'num_features': _int_feature([num_features]),
      'index': _int_feature(indexes),
      'value': _float_feature(values)
      }))

    writer.write(example.SerializeToString())

    #---read
    def decode(batch_serialized_examples):
        features = tf.parse_example(
            batch_serialized_examples,
            features={
                'label' : tf.FixedLenFeature([], tf.int64),
                'index' : tf.VarLenFeature(tf.int64),
                'value' : tf.VarLenFeature(tf.float32),
            })

        label = features['label']
        index = features['index']
        value = features['value']

        return label, index, value

So by this way you will get label as dense tensor, index and value as two sparse tensors, you can see one self contained example of writing libsvm format to TFRecord and read it for mlp classification from

https://github.com/chenghuige/tensorflow-example/tree/master/examples/tf-record/sparse https://github.com/chenghuige/tensorflow-example/tree/master/examples/text-classification