Hands-On Deep Learning with Go

Interchangeable models

The flexibility of Pachyderm pipelines allows you to easily swap out one model for another with a simple update or push of the JSON pipeline we used previously.

What's the point in specifying a pipeline in JSON? It's to make it repeatable! Pipelines reprocess data (in our case, to make new predictions about the classes of labels) each time their data is updated.

Here, we update the image flag in cifa_cnn.json to refer to a version of our containerized CNN that, for whatever reason, does not contain dropout:

"image": "pachyderm/cifar_cnn_train:nodropout"

We can then update the pipeline on the cluster, like so:

pachctl update pipeline -f cifar_cnn.json --reprocesses