djl icon indicating copy to clipboard operation
djl copied to clipboard

To train in TF engine

Open SidneyLann opened this issue 5 years ago • 5 comments

Description

TF engine can only be used for inference, when can be used for trainning?

Will this change the current api? How?

Yes. Wrap TF java api 2.x.

Who will benefit from this enhancement?

Many people.

References

  • list reference and related literature
  • list known implementations https://github.com/tensorflow/java https://gitter.im/tensorflow/sig-jvm

SidneyLann avatar Nov 17 '20 05:11 SidneyLann

Currently DJL does not offer training feature on TensorFlow. But we will have this done when it is necessary. You are more than welcome to contribute to the TF training feature and we will be happy to help you on the way :)

lanking520 avatar Nov 17 '20 18:11 lanking520

We had an RFC in TF JNI build in discussion now. @SidneyLann Have you tried to train with DJL with MXNet engine? It is mostly working.

lanking520 avatar Feb 17 '21 23:02 lanking520

I found DL4J has more functinalities than MXNet java, lik conv1d, conv2d, conv3d, conv4d and gnn.

SidneyLann avatar Feb 18 '21 12:02 SidneyLann

@SidneyLann we also have conv1d, conv2d, conv3d. Modern frameworks like TF, PT doesn't have conv4d. Can you elaborate your use case? Thanks

stu1130 avatar Feb 18 '21 17:02 stu1130

Human action in a video(a sequence of images), like conv3d+sequence images, it maybe not a realy conv4d.

SidneyLann avatar Feb 18 '21 21:02 SidneyLann