transform_output_raw not working . It gets refered to transform-input endpoint when analysing logs

[divyadilip91]

Hi,
I found an issue with transform_output_raw which did not seem to be working. I created a graph with a model and an output transformer and found that the ouptut from model (using predict_raw) was not passed to the transformer and in the logs I could see that "/transform-input" endpoint is being called instead of /transform-output
The seldon version I am using is 1.1.0.
Has this been fixed in the higher versions?
Expecting your reply

[axsaucedo]

Hi @divyadilip91, please fill out the template for bug creations in seldon, so we know exactly what image you used to wrap your model vs the image of the service orchestrator, as well as the yaml so we can reproduce. Thanks

[divyadilip91]

Hi @axsaucedo
Is there a service type OUTPUT_TRANSFORMER?I used TRANSFORMER as the service in my seldon graph for an outputtransformer component and not OUTPUT_TRANSFORMER. I didnt find a service of this type in the docs. Could the issue be because of this?

[divyadilip91]

@axsaucedo @RafalSkolasinski Hi . I am facing issue in creating the seldon graph too.. I am using seldon v1.1.0.
Could you help me in constructing the seldon graph for this diagram. Is it possible to pass the output of combiner to a model type service alone ?

#1927

[divyadilip91]

This is the graph I have created. But from the seldon docs and examples its unclear how to pass the combiner output to another node which can be a model or a transformer

[ukclivecox]

This can be achieved by making your model3 an output-transformer so it runs after the combiner.

[divyadilip91]

This can be achieved by making your model3 an output-transformer so it runs after the combiner.

Hi @cliveseldon
So a combiner output can only be passed to an output transformer ?
I have my model3 already created as service type model and hence i want to reuse the container rather than creating model3 as an output transformer. If I have to rebuild the same model code as type output transformer, it looks as if I am not reusing my images. Is there any workaround for this?

[RafalSkolasinski]

Please try setting type: OUTPUT_TRANSFORMER - such node can be a parent of combiner and will be able to process its output

[divyadilip91]

@RafalSkolasinski Thanks for your reply. So only an output transformer can be a parent of combiner?Could I connect the output from a combiner to be passed as input to another model/input transformer?

[RafalSkolasinski]

As far as I am aware only the OUTPUT_TRANSFORMER node (and obviously the combiner) can process output of their children.

In other words, parents of COMBINER of different type (MODEL, TRANSFORMER, etc.) will received the request first before the COMBINER itself.

The difference of OUTPUT_TRANSFORMER is such that the request first goes to its children and then to the OUTPUT_TRANSFORMER.

For the diagram above you could try:

- name: output-transformer
  type: OUTPUT_TRANSFORMER
  children:
  - name: model-3
    type: OUTPUT_TRANSFORMER
    children:
    - name: combiner
      type: COMBINER
      children:
      - name: model-1
        type: MODEL
      - name: model-2
        type: MODEL

[divyadilip91]

@divyadilip91 Ok got it. Thankyou for your response.