VertexAI AutoML regression model for tabular data, environment outside of the container.

[jeden8l]

Hello. 
Recently I downloaded the regression model for tabular data to my machine. 

I know how to run the container image and the model inside of it, but I want to avoid it due to latency (http communication and Windows-Linux overhead). 

I replicated the libraries stated in 'environment.json' file that comes with model files. 
Tensorflow, struct2tensor, tensorflow-addons. 
Then with a python script I triggered the prediction to the 001/saved_model.pb. 

No matter what I do, I end up with 'KeyError:'DecodeProtoSparseV4''. 

Can anyone help me to fix it? It's obviously some discrepancy inside the model processes. 

Maybe creating a custom container for cloud training process can help? If I create a custom cloud container with environment of my choice, then it's easier to replicate it on my local machine? 

[jeden8l]

Anyone on this please? It seems like DecodeProtoSparseV4 is custom process that I can't replicate by installing *all* packages used in container uri, not only TensorFlow related. 
Any thoughts please on how I can solve it? It's crucial for me.