support assets like MXNet model server or tensorflow savedmodel

[ydp]

I read the IR spec documents, did not find any field to represent assets like tensorflow savedModel format or MXNet model server, but that is necessary when we have vocabularies, etc.

tensorflow savedModel

Assets
    Subfolder called assets.
    Contains auxiliary files such as vocabularies, etc.
Extra assets
    Subfolder where higher-level libraries and users can add their own assets that co-exist with the model, but are not loaded by the graph.
    This subfolder is not managed by the SavedModel libraries.

MXNet model server:

Model Definition (json file) - contains the layers and overall structure of the neural network
Model Params and Weights (params file) - contains the parameters and the weights
Model Signature (json file) - defines the inputs and outputs that MMS is expecting to hand-off to the API
assets (text files) - auxiliary files that support model inference such as vocabularies, labels, etc. and vary depending on the model

[gramalingam]

Seems related to: https://github.com/onnx/onnx/wiki/%5Bdraft%5D-Proposal-for-ONNX-archive-file-format (at least partially).

[dzhulgakov]

Yes, for packaging parameters in a better form we're trying to use https://github.com/onnx/onnx/wiki/%5Bdraft%5D-Proposal-for-ONNX-archive-file-format

For model signature - there's some work in #879 to experiment with possible annotations.

For other assets - we can basically allow arbitrary files inside the archive. It'd be a bit harder to align the semantics of using assets across frameworks. What's the use case you have in mind to start?

[ydp]

Thanks for reply.

In my cases, I need to handle embedding in DNN models, to be more specifically, we have user ID and item ID in almost all recommender system or CTR algorithm, those IDs need to store in file, the format is one ID each line, the ID could be numbers or strings.

In the proposed format above, I guess we can refer each IDs list in a file, and compress all IDs list file in the zip file, don't know if this is the correct usage of this format. Would it be better if we have an example model for the format.

[quantum-fusion]

Where do these assets come from? The problem is that I only have an ONNX model from the model zoo. (https://github.com/onnx/models)

These two files are missing from the model zoo.

The missing files needed are:
signature.json - defining the input and output of the model
synset.txt - defining the set of classes the model was trained on, and returned by the model

---

Model Definition (json file) - contains the layers and overall structure of the neural network
Model Params and Weights (params file) - contains the parameters and the weights
Model Signature (json file) - defines the inputs and outputs that MMS is expecting to hand-off to the API
assets (text files) - auxiliary files that support model inference such as vocabularies, labels, etc. and vary depending on the model

missing file dependencies
#2944