Skip to content

I. List of supported engines

Jump to bottom
carlosuc3m edited this page Oct 19, 2023 · 5 revisions

Currently supported frameworks

Currently JDLL supports 3 Deep Learning frameworks:

  • Tensorflow 1 and 2. The tags used to refer to this framework are tensorflow or tensorflow_saved_model_bundled (second one is the one used by Bioimage.io).

  • Pytorch 1 and 2. The tags used to refer to this framework are pytorch or torchscript (second one is the one used by Bioimage.io).

  • Onnx. The tag used to refer to this framework is onnx.

DL framework Tag in JDLL Versions for Windows-x86_64 Versions for Linux-x86_64 Versions for MacOS-x86_64 Versions for Linux-arm64 Versions for MacOS-arm64
Tensorfow tensorflow or tensorflow_saved_model_bundle 1.12.0, 1.13.1, 1.14.0, 1.15.0, 2.3.1, 2.4.1, 2.7.0, 2.7.1, 2.7.4, 2.10.1 1.12.0, 1.13.1, 1.14.0, 1.15.0, 2.3.1, 2.4.1, 2.7.0, 2.7.1, 2.7.4, 2.10.1 1.12.0, 1.13.1, 1.14.0, 1.15.0, 2.3.1, 2.4.1, 2.7.0, 2.7.1, 2.7.4, 2.10.1 2.7.0 2.7.0
Pytorch pytorch or torchscript 1.7.1, 1.8.1, 1.9.0, 1.9.1, 1.10.0, 1.11.0, 1.12.1, 1.13.0, 1.13.1, 2.0.0 1.7.1, 1.8.1, 1.9.0, 1.9.1, 1.10.0, 1.11.0, 1.12.1, 1.13.0, 1.13.1, 2.0.0 1.7.1, 1.8.1, 1.9.0, 1.9.1, 1.10.0, 1.11.0, 1.12.1, 1.13.0, 1.13.1, 2.0.0 1.11.0, 1.12.1, 1.13.0, 1.13.1, 2.0.0 1.11.0, 1.12.1, 1.13.0, 1.13.1, 2.0.0
Onnx* onnx 8(1.3.1), 9(1.4.0), 10(1.5.2), 11(1.6.0), 12(1.7.0), 13(1.8.1), 14(1.9.0), 15(1.10.0), 16(1.11.0), 17(1.12.1), 18(1.13.1) 8(1.3.1), 9(1.4.0), 10(1.5.2), 11(1.6.0), 12(1.7.0), 13(1.8.1), 14(1.9.0), 15(1.10.0), 16(1.11.0), 17(1.12.1), 18(1.13.1) 8(1.3.1), 9(1.4.0), 10(1.5.2), 11(1.6.0), 12(1.7.0), 13(1.8.1), 14(1.9.0), 15(1.10.0), 16(1.11.0), 17(1.12.1), 18(1.13.1) 17(1.12.1), 18(1.13.1) 17(1.12.1), 18(1.13.1)

*For the versions column in Onnx, the number not in parenthesis refers to the opset version (used in the Bioimage.io rdf.yaml) and the number in parenthesis represents the actual version.

MacOS systems do not support GPU acceleration, although arm64 chips in newer Macs accelerate greatly the computation. On the other hand, for Windows and Linux-x86_64, for every DL framework and version there is an engine that supports GPU.

Non-explicitly supported Deep Learning version frameworks

Looking at the table above, one can observe that for some frameworks as Tensorfow, not every Python version has a Java API associated. JDLL tries to dive support for every model existing. It is assumed that the Java API of a certain version will be quite compatible with the close Python versions that do not have a coresponding Java API just for themselves.

This is why for example, when looking for Tensorflow 2.7.2 the among the available engines, the result will return the Tensorflow 2.7.4 version. JDLL assumes that Tensorflow 2.7.4 is equivalent to Tensorflow 2.7.2, so they can be used interchangeably. The list of interchangeable versions can be found here.

Here there are a couple of examples explaining this JDLL behaviour.

The first tries to find if JDLL has any engine that supports Tensorflow 2.7.2 in both CPU and GPU. The example uses methods explained in the Wiki section: VIV. Engine Management II (AvailableEngines).

String framework = "tensorflow";
String version = "2.7.2";
Boolean cpu = True;
Boolean gpu = True;

List<DeepLearningVersion> engineList = 
            AvailableEngines.getEnginesForOsByParams(framework, version, cpu, gpu);

System.out.println("Number of engines found: " + engineList.size());
System.out.println("Framework of the engine: " + engineList.get(0).getFramework());
System.out.println("Python version of the DL framework of the engine: " + engineList.get(0).getPythonVersion());
System.out.println("Java API version of the DL framework of the engine: " + engineList.get(0).getPythonVersion());

Output:

Number of engines found: 1
Framework of the engine: tensorflow
Python version of the DL framework of the engine: 2.7.4
Java API version of the DL framework of the engine: 0.4.2

It can be seen that the only engine returned does not have the same DL framework Python version as the one asked for (2.7.2 != 2.7.4), however JDLL assumes they are interchangeable, so it returns it. Looking at the JSON file with interchangeabilities, one can observe that they both use the same Java API version: 0.4.2

The second example shows how an EngineInfo instance can be created for Tensorflow 2.7.2, which by the table above, does not have an exact Java API available in JDLL. All the methods and classes for the example are explained in detail in the Wiki section: XIV. Load and run models I (EngineInfo)

String framework = "tensorflow";
String version = "2.7.2";
Boolean cpu = True;
Boolean gpu = True;
EngineInfo engineInfo = EngineInfo.defineDLEngine(framework, version, cpu, gpu);
System.out.println(engineInfo.getFramework());
System.out.println(engineInfo.getVersion());

Output:

tensorflow
2.7.4

Again it can be observed that JDLL considers Tensorflow 2.7.2 and 2.7.4 equivalent.

DL framework API version Minimum Java version required Interchangeable versions
tensorflow 2.10.1 11 2.11.0, 2.10.1, 2.10.0, 2.9.4, 2.9.3, 2.9.2, 2.9.1, 2.9.0, 2.8.5, 2.8.4, 2.8.3, 2.8.2, 2.8.1, 2.8.0
tensorflow 2.7.4 8 2.7.4, 2.7.2
tensorflow 2.7.1 8 2.7.1
tensorflow 2.7.0 8 2.7.0, 2.6.0, 2.6.1, 2.6.2, 2.6.3, 2.5.2, 2.5.1, 2.5.0
tensorflow 2.4.1 8 2.4.1, 2.4.4, 2.4.3, 2.4.2, 2.4.0
tensorflow 2.3.1 8 2.3.1, 2.3.4, 2.3.3, 2.3.2, 2.3.0, 2.2.3, 2.2.2, 2.2.1, 2.2.0, 2.1.4, 2.1.3, 2.1.2, 2.1.1, 2.1.0, 2.0.4, 2.0.3, 2.0.2, 2.0.1, 2.0.0
tensorflow 1.15.0 8 1.15.0, 1.15.5, 1.15.4, 1.15.3, 1.15.2, 1.15.1
tensorflow 1.14.0 8 1.14.0
tensorflow 1.13.1 8 1.13.1, 1.13.2
tensorflow 1.12.0 8 1.12.0, 1.12.3, 1.12.2, 1.12.1, 1.12.0
pytorch 2.0.0 8 2.0.0
pytorch 1.13.1 8 1.13.1
pytorch 1.13.0 8 1.13.0
pytorch 1.12.1 8 1.12.1
pytorch 1.11.0 8 1.11.0
pytorch 1.10.0 8 1.10.0
pytorch 1.9.1 8 1.9.1
pytorch 1.9.0 8 1.9.0
pytorch 1.8.1 8 1.8.1, 1.8.0
pytorch 1.7.1 8 1.7.1, 1.7.0, 1.6.0, 1.5.1, 1.5.0
onnx 18(1.13.1) 8 18(1.13.1)
onnx 17(1.12.1) 8 17(1.12.1)
onnx 16(1.11.0) 8 16(1.11.0)
onnx 15(1.10.0) 8 15(1.10.0)
onnx 14(1.9.0) 8 14(1.9.0)
onnx 13(1.8.1) 8 13(1.8.1)
onnx 12(1.7.0) 8 12(1.7.0)
onnx 11(1.6.0) 8 11(1.6.0)
onnx 10(1.5.2) 8 10(1.5.2)
onnx 9(1.4.0) 8 9(1.4.0)
onnx 8(1.3.1) 8 8(1.3.1), 7(1.2), 6(1.1.2)
Clone this wiki locally