luuil / merging-models-for-tensorflow-serving Goto Github PK

What I understood after playing around the scripts is that it just merges two frozen model (.pb files) into single model but the merged model is still storing the information about each member models.

Earlier, I was under the impression that we can unify and aggregate the models using this script and in the Tensorflow-serving, we can use the merged model as if it is a single model. But it is not, when we deploy it in Tensorflow-serving then we are still allowed to post the query to only one of the member model.
Note: signature_name = serving_default is hard-coded in the Tensorflow-serving which doesn't work with merged model.

My question is that

1. How do I get single model with a unique signature name, probably serving_default?
2. What I want to achieve is that I am having multiple frozen models (pre-trained model, each on different dataset). I want to merge them so that it behaves as if there is a single model which is trained on all the member models's dataset. How do I achieve it?

https://stackoverflow.com/questions/59381246/how-to-sequentially-combine-2-tensorflow-models
The problem mentioned in the above is what I have been trying to achieve.

From the repo. name, I thought it is what I needed but it doesn't. Please let me know if I am missing anything?

I have a merged model running in Tensorflow-Serving and when I do the following then I get an error.

python client.py --server_url "http://localhost:8501/v1/models/faster_rcnn_resnet:predict" --image_path ~/Downloads/image1.jpg --output_json ~/Downloads/output_image1.json --save_output_image "True" --label_map "$(pwd)/data/labels.pbtxt"

Error:
'Serving signature name: "serving_default" not found in signature def'

How to get rid of it?

Thanks

Hi,
Do you think it is possible to combine two models in such a way that the output of the first model, goes as input to the second model?