tf.io.parse_example  |  TensorFlow v2.16.1 (original) (raw)

Parses Example protos into a dict of tensors.

tf.io.parse_example(
    serialized, features, example_names=None, name=None
)

Used in the notebooks

Parses a number of serialized Exampleprotos given in serialized. We refer to serialized as a batch withbatch_size many entries of individual Example protos.

example_names may contain descriptive names for the corresponding serialized protos. These may be useful for debugging purposes, but they have no effect on the output. If not None, example_names must be the same length asserialized.

This op parses serialized examples into a dictionary mapping keys to Tensor SparseTensor, and RaggedTensor objects. features is a Mapping from keys to VarLenFeature, SparseFeature, RaggedFeature, and FixedLenFeatureobjects. Each VarLenFeature and SparseFeature is mapped to aSparseTensor; each FixedLenFeature is mapped to a Tensor; and eachRaggedFeature is mapped to a RaggedTensor.

Each VarLenFeature maps to a SparseTensor of the specified type representing a ragged matrix. Its indices are [batch, index] where batchidentifies the example in serialized, and index is the value's index in the list of values associated with that feature and example.

Each SparseFeature maps to a SparseTensor of the specified type representing a Tensor of dense_shape [batch_size] + SparseFeature.size. Its values come from the feature in the examples with key value_key. A values[i] comes from a position k in the feature of an example at batch entry batch. This positional information is recorded in indices[i] as[batch, index_0, index_1, ...] where index_j is the k-th value of the feature in the example at with key SparseFeature.index_key[j]. In other words, we split the indices (except the first index indicating the batch entry) of a SparseTensor by dimension into different features of theExample. Due to its complexity a VarLenFeature should be preferred over aSparseFeature whenever possible.

Each FixedLenFeature df maps to a Tensor of the specified type (ortf.float32 if not specified) and shape (serialized.size(),) + df.shape.

FixedLenFeature entries with a default_value are optional. With no default value, we will fail if that Feature is missing from any example inserialized.

Each FixedLenSequenceFeature df maps to a Tensor of the specified type (or tf.float32 if not specified) and shape(serialized.size(), None) + df.shape. All examples in serialized will be padded with default_value along the second dimension.

Each RaggedFeature maps to a RaggedTensor of the specified type. It is formed by stacking the RaggedTensor for each example, where theRaggedTensor for each individual example is constructed using the tensors specified by RaggedTensor.values_key and RaggedTensor.partition. See the tf.io.RaggedFeature documentation for details and examples.

Examples:

For example, if one expects a tf.float32 VarLenFeature ft and three serialized Examples are provided:

serialized = [
  features
    { feature { key: "ft" value { float_list { value: [1.0, 2.0] } } } },
  features
    { feature []},
  features
    { feature { key: "ft" value { float_list { value: [3.0] } } }
]

then the output will look like:

{"ft": SparseTensor(indices=[[0, 0], [0, 1], [2, 0]],
                    values=[1.0, 2.0, 3.0],
                    dense_shape=(3, 2)) }

If instead a FixedLenSequenceFeature with default_value = -1.0 andshape=[] is used then the output will look like:

{"ft": [[1.0, 2.0], [3.0, -1.0]]}

Given two Example input protos in serialized:

[
  features {
    feature { key: "kw" value { bytes_list { value: [ "knit", "big" ] } } }
    feature { key: "gps" value { float_list { value: [] } } }
  },
  features {
    feature { key: "kw" value { bytes_list { value: [ "emmy" ] } } }
    feature { key: "dank" value { int64_list { value: [ 42 ] } } }
    feature { key: "gps" value { } }
  }
]

And arguments

example_names: ["input0", "input1"],
features: {
    "kw": VarLenFeature(tf.string),
    "dank": VarLenFeature(tf.int64),
    "gps": VarLenFeature(tf.float32),
}

Then the output is a dictionary:

{
  "kw": SparseTensor(
      indices=[[0, 0], [0, 1], [1, 0]],
      values=["knit", "big", "emmy"]
      dense_shape=[2, 2]),
  "dank": SparseTensor(
      indices=[[1, 0]],
      values=[42],
      dense_shape=[2, 1]),
  "gps": SparseTensor(
      indices=[],
      values=[],
      dense_shape=[2, 0]),
}

For dense results in two serialized Examples:

[
  features {
    feature { key: "age" value { int64_list { value: [ 0 ] } } }
    feature { key: "gender" value { bytes_list { value: [ "f" ] } } }
   },
   features {
    feature { key: "age" value { int64_list { value: [] } } }
    feature { key: "gender" value { bytes_list { value: [ "f" ] } } }
  }
]

We can use arguments:

example_names: ["input0", "input1"],
features: {
    "age": FixedLenFeature([], dtype=tf.int64, default_value=-1),
    "gender": FixedLenFeature([], dtype=tf.string),
}

And the expected output is:

{
  "age": [[0], [-1]],
  "gender": [["f"], ["f"]],
}

An alternative to VarLenFeature to obtain a SparseTensor isSparseFeature. For example, given two Example input protos inserialized:

[
  features {
    feature { key: "val" value { float_list { value: [ 0.5, -1.0 ] } } }
    feature { key: "ix" value { int64_list { value: [ 3, 20 ] } } }
  },
  features {
    feature { key: "val" value { float_list { value: [ 0.0 ] } } }
    feature { key: "ix" value { int64_list { value: [ 42 ] } } }
  }
]

And arguments

example_names: ["input0", "input1"],
features: {
    "sparse": SparseFeature(
        index_key="ix", value_key="val", dtype=tf.float32, size=100),
}

Then the output is a dictionary:

{
  "sparse": SparseTensor(
      indices=[[0, 3], [0, 20], [1, 42]],
      values=[0.5, -1.0, 0.0]
      dense_shape=[2, 100]),
}

See the tf.io.RaggedFeature documentation for examples showing howRaggedFeature can be used to obtain RaggedTensors.