Skip to content

Implement transfer learning functionality in TensorFlowTransform. #951

Closed
@zeahmed

Description

@zeahmed

There are couple of hurdles in implementing a transfer learning scenario (where tf model is actually modified) using TF#.

  • [SOLVED] There is no way to serialized TF model back to file using TF C-API (also TF#).
  • It seems impossible to load checkpoint models using C-API (also TF#) because the graph saved with checkpoint models is saved in meta_graph format for which straightforward API is currently missing to load/save.
  • When using frozen model, there are no training ops like optimizations or loss operations. Also, frozen model don't have variables because those are converted into constant while the model is frozen. So, frozen models can't be retrained.
  • When using un-frozen model (saved with simple_save method in python), training operations must be included in the graph so that training can happen from TensorFlowTransform.

Metadata

Metadata

Assignees

Labels

enhancementNew feature or request

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions