Advertisement
Not a member of Pastebin yet?
Sign Up,
it unlocks many cool features!
- apiVersion: "serving.kubeflow.org/v1alpha2"
- kind: "InferenceNode"
- metadata:
- name: "my-model"
- spec:
- default:
- predictor:
- tensorflow:
- storageUri: "gs://mybucket/mymodel-2"
- # A/B testing add-on
- ABTest:
- # which metric (the log request can log multiple) are we looking for an improvement in?
- metricName: "my_metric"
- # what percent of users are randomly routed to this model?
- # default: 50
- trafficPercent: 10
- # over the course of a single A/B test, all users _must_ consistently be routed to the same model
- # however, we must _not_ have correlations between user assignment to group A vs group B
- # example: if user 1 and user 2 are both in group A in one A/B test,
- # it shouldn't be guaranteed that they are in the same group in the next A/B test
- # thus: assuming we have 1) user ID and 2) some unique A/B test id,
- # we may determine what model to route a user to by hashing user ID and A/B test ID
- # `hash(user_id, abtest_id) / MAX_VALUE < trafficPercent / 100` (0 -> model A, 1 -> model B)
- # default: number of seconds after epoch (just has to be unique)
- seed: 93020191153
- # STATISTICAL PARAMETERS
- # For metric estimation (before the A/B test starts; to calculate sample size required for the A/B test)
- # maximumPercentError - largest possible percent error at the specified confidence value, default: 1
- # estimationConfidence - percent chance that our true estimation error <= maximumPercentError, default: 95
- maximumPercentError: 1
- estimationConfidence: 95
- # For A/B testing
- # *note*: for convenience, define the output of our A/B test, C = (B's metric >= A's metric + minimumDetectableEffect)
- # minimumDetectableEffect - smallest detectable absolute improvement
- # in B's metric over A's at the specified confidence & power values
- # confidence - percent chance that we predict C is false given C is indeed false, default: 95
- # power - percent chance that we predict C is true given C is indeed true, default: 80
- # *note*: higher confidence/power and lower minimumDetectableEffect -> longer experiments (more samples required)
- minimumDetectableEffect: 0.01
- confidence: 95
- power: 80
- # CONSIDERING:
- # timeout for sessions which do not receive a log request?
- # specifying our second predictor
- predictor:
- tensorflow:
- storageUri: "gs://mybucket/mymodel-3"
Advertisement
Add Comment
Please, Sign In to add comment
Advertisement