Uploading the model#

Using mlboardclient it is possible to upload the catalog as a model.

Once the client have been initialized, it may be used as the following:

mlboard.model_upload("model-name", "1.0.0-version", "/model-catalog-dir")

Full model_upload method spec:

model_upload(self, model_name, version, path,
             workspace=None, auto_create=True, spec=None)

Arguments:

• auto_create: if True, then model will be created if there is no such model yet. Defaults to True.
• spec: Optional model spec dict, it contains model specification used for starting the model as a serving. If None and this method is called inside ML project, then mlboardclient tries to get spec automatically from the project if it exists. See about model spec format below.
• workspace: Optinal parameter. Must be provided if the method is called outside of ML project.

Model spec format#

Model spec should be dict or corresponding JSON-string. Example is below:

{
  "displayName": "Serving",
  "name": "tensorflow-serving",
  "ports": [
    {
      "protocol": "TCP",
      "targetPort": 9000,
      "name": "grpc",
      "port": 9000
    }
  ],
  "command": "kuberlab-serving --port=9000 --model-path=$SRC_DIR",
  "sources": [
    {
      "mountPath": "/src",
      "name": "src"
      "gitRepo": {
        "repository": "https://github.com/kuberlab-catalog/tensorflow"
      }
    }
  ],
  "images": {
    "gpu": "kuberlab/serving:latest-gpu",
    "cpu": "kuberlab/serving:latest"
  },
  "spec": {
    "outFilter": "string",
    "rawInput": true,
    "model": "string",
    "params": [
      {
        "name": "input",
        "type": "bytes"
      }
    ]
  },
  "resources": {
    "accelerators": {
      "gpu": 0
    },
    "requests": {
      "cpu": "100m",
      "memory": "125Mi"
    },
    "limits": {
      "cpu": "1",
      "memory": "4Gi"
    }
  }
}

Note: spec section is needed mostly for the web UI, to pass the data to the serving via web UI.