Reasonable Scala API for serving TensorFlow models
More details about the library could be found here.
The core API algebra is agnostic to TensorFlow itself and could be used for binding to any TensorFlow-like library:
/**
* Abstract model serving algebra.
* Defines a set of operations on a pre-built model that is stored
* somewhere on an external storage.
*
* @tparam F Type of the effect used by this algebra.
*/
trait ModelServing[F[_]] {
// Model type.
type TModel
// Tensor type.
type TTensor
/**
* Loads a model from an external source.
*
* @param source A source to load a model from.
* @param tag A tag that could be used to uniquely identify
* a concrete computational graph inside a saved model.
* @return A model.
*/
def load(source: ModelSource, tag: String): F[TModel]
/**
* Queries a model metadata.
*
* @param model A model to query metadata for.
* @return A model metadata wrapped into an effect.
*/
def metadata(model: TModel): F[ModelMetadata]
/**
* Creates a tensor from its representation in the form of a different data structure
* e.g. Array, List etc.
*
* @param data An instance of a data structure for building the tensor.
* @param shape Shape of the resulting tensor.
* @param E Tensor encoder to use.
* @tparam TRepr Type of the data structure to build tensor from.
* @return A tensor wrapped into an effect.
*/
def tensor[TRepr](data: TRepr, shape: List[Long])(implicit E: TensorEncoder[TTensor, TRepr]): F[TTensor]
/**
* Evaluates a tensor value by feeding a set of input tensors as defined by a signature
* into the model.
* @param model A model to use.
* @param output Output tensor metadata.
* @param feed Map of a tensor metadata to the tensor to use as an input.
* @param D Tensor decoder to use.
* @param C Closeable type class for tensor type being used.
* @tparam TRepr Type of the data structure to map output tensor to.
* @return A data structure representing an output tensor calculated by feeding a set
* of inputs into the model.
*/
def eval[TRepr](model: TModel, output: TensorMetadata, feed: Map[TensorMetadata, TTensor])
(implicit D: TensorDecoder[TTensor, TRepr], C: Closeable[TTensor]): F[TRepr]
/**
* Closes the model and releases all related resources.
*
* @param model A model to close.
* @return Unit.
*/
def close(model: TModel): F[Unit]
}
The binding for TensorFlow is implemented in TFModelServing.scala.
- Create a TensorFlow graph and save it using Saved Model API:
import tensorflow as tf
import numpy as np
export_dir = '/tmp/saved_model_1'
builder = tf.saved_model.builder.SavedModelBuilder(export_dir=export_dir)
with tf.Graph().as_default(), tf.Session().as_default() as sess:
x = tf.placeholder(shape=(2, 3), dtype=tf.float32, name='x')
y = tf.Variable(np.identity(3), dtype=tf.float32)
z = tf.matmul(x, y, name='z')
tf.global_variables_initializer().run()
zval = z.eval(feed_dict={x: np.random.randn(2, 3)})
print(zval)
x_proto_info = tf.saved_model.utils.build_tensor_info(x)
z_proto_info = tf.saved_model.utils.build_tensor_info(z)
prediction_signature = (
tf.saved_model.signature_def_utils.build_signature_def(
inputs={'x': x_proto_info},
outputs={'z': z_proto_info},
method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME))
builder.add_meta_graph_and_variables(sess, [tf.saved_model.tag_constants.SERVING],
signature_def_map={
tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: prediction_signature
})
builder.save()
The script is available here.
In this example the computational graph simply does matrix multiplication of an input [2x3] matrix and identity [3x3] matrix. The result matrix has to be equal to the input matrix.
The model with all variables will be saved into /tmp/saved_model_1
dir.
- Load saved model in Scala, read its metadata, feed input matrix into the loaded computational graph and show the output matrix:
import scala.util.Try
import tf._
import tf.implicits._
import dsl._
import utils._
import org.tfModelServing4s.utils.show._
import org.tfModelServing4s.utils.show.implicits._
object Example1 {
def main(args: Array[String]): Unit = {
val serving = new TFModelServing
val progr = for {
_ <- use(serving.load(FileModelSource("/tmp/saved_model_1"), tag = "serve")) { model =>
for {
meta <- serving.metadata(model)
_ = println(s"model metadata: $meta")
signature <- Try { meta.signatures("serving_default") }
_ = println(s"serving signature: $signature")
_ = println(s"serving signature inputs: ${signature.inputs}")
_ = println(s"serving signature outputs: ${signature.outputs}")
inputArray <- Try { Array.range(0, 6).map(_.toFloat) }
_ = println(s"input array = ${shows(inputArray)}")
_ <- use(serving.tensor(inputArray, shape = List(2, 3))) { inputTensor =>
for {
inputDef <- Try { signature.inputs("x") }
outputDef <- Try { signature.outputs("z") }
outputArray <- serving.eval[Array[Array[Float]]](model, outputDef, Map(inputDef -> inputTensor))
_ = println(s"output: ${shows(outputArray)}")
} yield ()
}
} yield ()
}
} yield ()
println(s"Program result = $progr")
}
}
Output:
model metadata: ModelMetadata(Map(serving_default -> SignatureMetadata(tensorflow/serving/predict,Map(x -> TensorMetadata(x:0,x,DTypeFloat,List(2, 3))),Map(z -> TensorMetadata(z:0,z,DTypeFloat,List(2, 3))))))
serving signature: SignatureMetadata(tensorflow/serving/predict,Map(x -> TensorMetadata(x:0,x,DTypeFloat,List(2, 3))),Map(z -> TensorMetadata(z:0,z,DTypeFloat,List(2, 3))))
serving signature inputs: Map(x -> TensorMetadata(x:0,x,DTypeFloat,List(2, 3)))
serving signature outputs: Map(z -> TensorMetadata(z:0,z,DTypeFloat,List(2, 3)))
input array = [0.0, 1.0, 2.0, 3.0, 4.0, 5.0]
releasing TF tensor
output: [[0.0, 1.0, 2.0], [3.0, 4.0, 5.0]]
releasing TF tensor
closing TF model
Program result = Success(())
The example sources are here.
This repo contains the description how to build dog breed classifier using pre-trained Inception model. The final model gets exported as a "frozen" graph instead of a SavedModel format, though. This script converts frozen model to SavedModel and could be used with this library.
- Follow steps to build dog breed classifier as described here
- Convert resulting "frozen" model to SavedModel format:
python -m src.freezing.frozen_to_saved_model
- Use the dog breed classifier saved model to classify an arbitrary dog image:
import scala.util.Try
import tf._
import tf.implicits._
import dsl._
import utils._
import org.tfModelServing4s.utils.show._
import org.tfModelServing4s.utils.show.implicits._
import java.nio.file.{Files, Paths}
object Example2 {
private def probsToClass(probs: Array[Float]): String = {
val classes = io.Source.fromInputStream(getClass.getResourceAsStream("/breeds.csv")).getLines().drop(1).toArray
val top5 = probs.zip(classes).sortBy { case (prob, idx) => prob }.reverse.take(5)
top5.mkString("\n")
}
def main(args: Array[String]): Unit = {
val imagePath = args(0)
val serving = new TFModelServing
val progr = for {
_ <- use(serving.load(FileModelSource("/tmp/dogs_1"), tag = "serve")) { model =>
for {
meta <- serving.metadata(model)
_ = println(s"model metadata: $meta")
signature <- Try { meta.signatures("serving_default") }
_ = println(s"serving signature: $signature")
_ = println(s"serving signature inputs: ${signature.inputs}")
_ = println(s"serving signature outputs: ${signature.outputs}")
inputArray <- Try { Array.range(0, 6).map(_.toFloat) }
_ = println(s"input array = ${shows(inputArray)}")
img <- Try {
Files.readAllBytes(Paths.get(imagePath))
}
_ <- use(serving.tensor(img, shape = List(1))) { inputTensor =>
for {
inputDef <- Try { signature.inputs("image_raw") }
outputDef <- Try { signature.outputs("probs") }
outputArray <- serving.eval[Array[Array[Float]]](model, outputDef, Map(inputDef -> inputTensor))
_ = println(s"output: ${shows(outputArray)}")
clazz <- Try { probsToClass(outputArray.flatten) }
_ = println(clazz)
} yield ()
}
} yield ()
}
} yield ()
println(s"Program result = $progr")
}
}
Here is the output for the sample dog image:
(0.9796268,3,airedale)
(0.00985535,83,otterhound)
(0.007160045,57,irish_terrier)
(0.0015072817,28,chesapeake_bay_retriever)
(8.2421233E-4,118,wire-haired_fox_terrier)
The model outputs probabilities array which gets mapped into top 5 breeds together with their probabilities.
The example source code is here.