spotify / tfexample-derive   0.2.5

Apache License 2.0 GitHub

Provides compile-time derivation of conversions between Scala case classes and Tensorflow Example protocol buffers

Scala versions: 2.13 2.12 2.11


Functionality of this library has been superceded by Magnolify. Use magnolify-tensorflow for Example converter derivation.


Build Status GitHub license Maven Central

Magnolia-based conversions between case classes and tensorflow Example protobufs.

libraryDependencies += "com.spotify" %% "tfexample-derive" % "0.2.5"


ExampleConverter[T] is a typeclass that converts between case class T and Tensorflow Example Protobuf types:

import com.spotify.tfexample.derive._

case class Data(floats: Array[Float], longs: Array[Long], strings: List[String], label: String)

val converter = ExampleConverter[Data]
val data = Data(Array(1.5f, 2.5f), Array(1L, 2L), List("a", "b"), "x")
val example = converter.toExample(data)
val data2: Option[Data] = converter.fromExample(example)

The derivation makes use of magnolia, which provides a generic macro for materializing typeclasses for case classes.

Supported Types

Tensorflow Example is an inherently flat structure - essentially a map of (String -> Feature), where Feature is one of:

  • Int64List
  • FloatList
  • BytesList

A converter can be automatically derived for types which naturally correspond to these feature types - Int, Long, Float, Double, ByteString, String etc, and by extension collections of these types including Array, Seq, List. Option is also supported, simply by not encoding None values in the resulting Example. See the below section on custom types for an example of how to add encodings for new types.


As mentioned above, Example is a flat structure, but a converter can be derived in certain cases even for nested case classes. Flattening is achieved by using the field name of the nested case class as a namespace for the features belonging to that class. For example:

case class Record(xs: List[Int], inner: Inner)
case class Inner(ys: List[Float], labels: Option[List[String]])

val record = Record(List(1, 2, 3), Inner(List(1.0f, 2.0f), Some(List("hello"))))
features {
  feature {
    key: "xs"
    value {
      int64_list {
        value: 1
        value: 2
        value: 3
  feature {
    key: "inner.ys"
    value {
      float_list {
        value: 1.0
        value: 2.0
  feature {
    key: "inner.labels"
    value {
      bytes_list {
        value: "hello"

However, the following will result in a compilation error:

case class Record(xs: List[Int], inners: List[Inner])
case class Inner(y: Float, label: Option[String])
Error: could not find implicit value for parameter converter: com.spotify.tfexample.derive.ExampleConverter[Record]

To drill down and find the particular field that breaks the derivation, we can get more information from Magnolia by directly calling the macro. Replace the call to ExampleConverter with a call to FeatureBuilder.gen:

import com.spotify.tfexample.derive.FeatureBuilder

case class Record(xs: List[Int], inners: List[Inner])
case class Inner(y: Float, label: Option[String])
Error: cannot derive FeatureBuilder for type List[Inner]

Magnolia tells us we cannot derive a FeatureBuilder for List[Inner], and this makes sense - for List[T], T must correspond to one of the three feature types, and in the case of a nested case class, there's no obvious mapping. A mapping can either be provided (see below for an example) or the case class should be restructured.

Custom Types

In addition to the types supported out of the box, custom types are also supported by providing an implicit TensorflowMapping, which defines an appropriate encoding of the type to one of the three possible feature types. In this example, we encode a URI as a BytesList feature (via String), using some helper functions from TensorflowMapping.scala. We create a TensorflowMapping by supplying functions for converting to and from Feature from type T.

import com.spotify.tfexample.derive.TensorflowMapping._

implicit val uriType: TensorflowMapping[URI] =
  TensorflowMapping[URI](toStrings(_).map(URI.create), xs => fromStrings(

case class Record(uri: URI, uris: List[URI])
val converter = ExampleConverter[Record]
val record = Record(URI.create(""), List(URI.create("")))
val example = converter.toExample(record)

Code of Conduct

This project adheres to the Open Code of Conduct. By participating, you are expected to honor this code.