[FEA] Integrating Model Zoo and MMAR

[vikashg]

Is your feature request related to a problem? Please describe.
Can we consider integrating a set of pre-trained models to monai-deploy-app-sdk?

Describe the solution you'd like

We should have something like MMAR integrated into monai-deploy. In this context the user will directly download a docker image (or model) from nvidia cloud for direct usage. The model that will be downloaded will contain all the relevant information (i.e., input/output node names, data type, input/output size) required for writing the config.pbtxt for a triton inference engine.

Describe alternatives you've considered
As of now we are using a google drive where the model is hosted.

Additional context
We can have this as a feature where models from a model zoo can be directly integrated to monai-deploy app sdk. There are so many ubiquitous models in deep learning and radiology, for example breast cancer detector, brain mask segmentation, lung nodule detection, fracture detection, tumor segmentation (BRATS). Though different groups have different models, these are some of the problems people have been working on for a while now and we do have models with reasonably good accuracy. We should integrate them with monai-deploy. A new lab/research team can just download these models and also the infrastructure (through the deploy sdk) to run inference. Otherwise people may get the model but still takes a lot of work to make them work.

• host/load pretrained weights for 3D resnet model-zoo#48
• Develop MVP of model bundle MONAI#3482
  • [WIP] 486 Add example model package tutorials#487

[dbericat]

@ericspod @wyli I know you have been thinking along these lines for an MMAR-ish-like common. Thoughts?

[ericspod]

We are definitely working on that. We've had a look at MMAR and solutions from Huggingface and others, we're in the process of developing a prototype repo for a shared model to discuss and consider its contents and what information we need in such things. MMAR is a good starting point for what the structure likely will be.

[MMelQin]

MONAI Model Zoo is now public hosting MONAI Bundle (model and metadata).

App SDK has a specific issue for creating an built-in operator to consume MONAI Bundles and automates inference in an App.

The PR for the aforementioned issue is being reviewed and once approved will be release in App SDK v0.4