search

onnx / onnx-tensorflow

Tensorflow Backend for ONNX
Other
1.28k stars 296 forks source link

How to get trainable variables from the saved model? #994

Open adjhawar opened 2 years ago

adjhawar commented 2 years ago

I am able to convert onnx model to tensorflow. However, when I try to access trainable variables/variables from the model, it displays that the attribute is not present. Also, when I try to access the variables/trainable variables from model.signatures['serving_default'], that list is also empty.

How can I get the trainable variables/variables from the converted tensorflow model?

chinhuang007 commented 2 years ago

While you could run in-place training as seen in this example, the current exported tensorflow saved model is for inference only, and does not include training variables.

adjhawar commented 2 years ago

Is it possible to modify the converter code and get the required support? If yes, then what changes will be required on the onnx-tensorflow converter side to support training on the converted model?

IaroslavElistratov commented 1 year ago

chinhuang007 could you please clarify what does "in-place training" mean? Could you also please clarify why does the example you've shared use tf1 and not tf2?