Pipeline examples
This notebook illustrates a series of Pipelines showing of different ways of combining flows of data and conditional logic. We assume you have Seldon Core 2 running locally.
Models Used
gs://seldon-models/triton/simplean example Triton tensorflow model that takes 2 inputs INPUT0 and INPUT1 and adds them to produce OUTPUT0 and also subtracts INPUT1 from INPUT0 to produce OUTPUT1. See here for the original source code and license.Other models can be found at https://github.com/SeldonIO/triton-python-examples
Model Chaining
Chain the output of one model into the next. Also shows chaning the tensor names via tensorMap to
conform to the expected input tensor names of the second model.
cat ./models/tfsimple1.yaml
cat ./models/tfsimple2.yamlapiVersion: mlops.seldon.io/v1alpha1
kind: Model
metadata:
name: tfsimple1
spec:
storageUri: "gs://seldon-models/triton/simple"
requirements:
- tensorflow
memory: 100Ki
---
apiVersion: mlops.seldon.io/v1alpha1
kind: Model
metadata:
name: tfsimple2
spec:
storageUri: "gs://seldon-models/triton/simple"
requirements:
- tensorflow
memory: 100Ki
The pipeline below chains the output of tfsimple1 into tfsimple2. As these models have compatible
shape and data type this can be done. However, the output tensor names from tfsimple1 need to be
renamed to match the input tensor names for tfsimple2. We do this with the tensorMap feature.
The output of the Pipeline is the output from tfsimple2.
We use the Seldon CLI pipeline inspect feature to look at the data for all steps of the pipeline for the last data item passed through the pipeline (the default). This can be useful for debugging.
Next, we look get the output as json and use the jq tool to get just one value.
Model Chaining from inputs
Chain the output of one model into the next. Shows using the input and outputs and combining.
Model Join
Join two flows of data from two models as input to a third model. This shows how individual flows of data can be combined.
In the pipeline below for the input to tfsimple3 we join 1 output tensor each from the two previous models tfsimple1 and tfsimple2. We need to use the tensorMap feature to rename each output tensor to one of the expected input tensors for the tfsimple3 model.
The outputs are the sequence "2,4,6..." which conforms to the logic of this model (addition and subtraction) when fed the output of the first two models.
Conditional
Shows conditional data flows - one of two models is run based on output tensors from first.
Here we assume the conditional model can output two tensors OUTPUT0 and OUTPUT1 but only outputs the former if the CHOICE input tensor is set to 0 otherwise it outputs tensor OUTPUT1. By this means only one of the two downstream models will receive data and run. The output steps does an any join from both models and whichever data appears first will be sent as output to pipeline. As in this case only 1 of the two models add10 and mul10 runs we will receive their output.
The mul10 model will run as the CHOICE tensor is set to 0.
The add10 model will run as the CHOICE tensor is not set to zero.
Pipeline Input Tensors
Access to indivudal tensors in pipeline inputs
This pipeline shows how we can access pipeline inputs INPUT0 and INPUT1 from different steps.
Trigger Joins
Shows how joins can be used for triggers as well.
Here we required tensors names ok1 or ok2 to exist on pipeline inputs to run the mul10 model but require tensor ok3 to exist on pipeline inputs to run the add10 model. The logic on mul10 is handled by a trigger join of any meaning either of these input data can exist to satisfy the trigger join.
Last updated
Was this helpful?