kubeflow
/
examples
mirror of https://github.com/kubeflow/examples.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
examples/xgboost_synthetic/build-train-deploy.ipynb

879 lines
31 KiB
Plaintext
Raw Blame History

{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Train and deploy on Kubeflow from Notebooks\n",
"\n",
"This notebook shows you how to use Kubeflow to build, train, and deploy models on Kubernetes.\n",
"This notebook walks you through the following\n",
" \n",
"* Building an XGBoost model inside a notebook\n",
"* Training the model inside the notebook\n",
"* Performing inference using the model inside the notebook\n",
"* Using Kubeflow Fairing to launch training jobs on Kubernetes\n",
"* Using Kubeflow Fairing to build and deploy a model using [Seldon Core](https://www.seldon.io/)\n",
"* Using [Kubeflow metadata](https://github.com/kubeflow/metadata) to record metadata about your models\n",
"* Using [Kubeflow Pipelines](https://www.kubeflow.org/docs/pipelines/) to build a pipeline to train your model\n",
"\n",
"## Prerequisites \n",
"\n",
"* This notebook assumes you are running inside 0.6 Kubeflow deployed on GKE following the [GKE instructions](https://www.kubeflow.org/docs/gke/deploy/)\n",
"* If you are running somewhere other than GKE you will need to modify the notebook to use a different docker registry or else configure Kubeflow to work with GCR."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Verify we have a GCP account\n",
"\n",
"* The cell below checks that this notebook was spawned with credentials to access GCP\n",
"* To add credentials when you created the notebook you should have selected add gcp credential as shown below\n",
" ![add credential](images/addgcpsecret.png)\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"import os\n",
"if not os.getenv(\"GOOGLE_APPLICATION_CREDENTIALS\"):\n",
" raise ValueError(\"Notebook is missing google application credentials\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Install Required Libraries\n",
"\n",
"Import the libraries required to train this model."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"!pip3 install retrying"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"* Install a specific version of kubeflow-fairing that this example is tested against"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"!pip3 install git+git://github.com/kubeflow/fairing.git@b3db9a548b51eea93250c662defe6470283943b3"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"* Perform some notebook setup"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"scrolled": false
},
"outputs": [],
"source": [
"import util\n",
"from pathlib import Path\n",
"import os\n",
"\n",
"util.notebook_setup()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"* Import the python libraries we will use\n",
"* We add a comment \"fairing:include-cell\" to tell the kubefow fairing preprocessor to keep this cell when converting to python code later"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# fairing:include-cell\n",
"import fire\n",
"import joblib\n",
"import logging\n",
"import kfmd\n",
"import nbconvert\n",
"import os\n",
"import pathlib\n",
"import sys\n",
"from pathlib import Path\n",
"import pandas as pd\n",
"import pprint\n",
"from sklearn.metrics import mean_absolute_error\n",
"from sklearn.model_selection import train_test_split\n",
"from sklearn.impute import SimpleImputer\n",
"from xgboost import XGBRegressor\n",
"from importlib import reload\n",
"from sklearn.datasets import make_regression\n",
"from kfmd import metadata\n",
"from kfmd import openapi_client\n",
"from kfmd.openapi_client import Configuration, ApiClient, MetadataServiceApi\n",
"from datetime import datetime\n",
"import retrying\n",
"import urllib3"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# Imports not to be included in the built docker image\n",
"import kfp\n",
"import kfp.components as comp\n",
"import kfp.gcp as gcp\n",
"import kfp.dsl as dsl\n",
"import kfp.compiler as compiler\n",
"from kubernetes import client as k8s_client\n",
"from kubeflow import fairing \n",
"from kubeflow.fairing.builders import append\n",
"from kubeflow.fairing.deployers import job\n",
"from kubeflow.fairing.preprocessors.converted_notebook import ConvertNotebookPreprocessorWithFire\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Code to train and predict \n",
"\n",
"* In the cells below we define some functions to generate data and train a model\n",
"* These functions could just as easily be defined in a separate python module"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# fairing:include-cell\n",
"def read_synthetic_input(test_size=0.25):\n",
" \"\"\"generate synthetic data and split it into train and test.\"\"\"\n",
" # generate regression dataset\n",
" X, y = make_regression(n_samples=200, n_features=5, noise=0.1)\n",
" train_X, test_X, train_y, test_y = train_test_split(X,\n",
" y,\n",
" test_size=test_size,\n",
" shuffle=False)\n",
"\n",
" imputer = SimpleImputer()\n",
" train_X = imputer.fit_transform(train_X)\n",
" test_X = imputer.transform(test_X)\n",
"\n",
" return (train_X, train_y), (test_X, test_y)\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# fairing:include-cell\n",
"def train_model(train_X,\n",
" train_y,\n",
" test_X,\n",
" test_y,\n",
" n_estimators,\n",
" learning_rate):\n",
" \"\"\"Train the model using XGBRegressor.\"\"\"\n",
" model = XGBRegressor(n_estimators=n_estimators, learning_rate=learning_rate)\n",
"\n",
" model.fit(train_X,\n",
" train_y,\n",
" early_stopping_rounds=40,\n",
" eval_set=[(test_X, test_y)])\n",
"\n",
" print(\"Best RMSE on eval: %.2f with %d rounds\",\n",
" model.best_score,\n",
" model.best_iteration+1)\n",
" return model\n",
"\n",
"def eval_model(model, test_X, test_y):\n",
" \"\"\"Evaluate the model performance.\"\"\"\n",
" predictions = model.predict(test_X)\n",
" mae=mean_absolute_error(predictions, test_y)\n",
" logging.info(\"mean_absolute_error=%.2f\", mae)\n",
" return mae\n",
"\n",
"def save_model(model, model_file):\n",
" \"\"\"Save XGBoost model for serving.\"\"\"\n",
" joblib.dump(model, model_file)\n",
" logging.info(\"Model export success: %s\", model_file)\n",
"\n",
"@retrying.retry(stop_max_delay=180000)\n",
"def wait_for_istio(address=\"metadata-service.kubeflow.svc.cluster.local:8080\"):\n",
" \"\"\"Wait until we can connect to the metadata service.\n",
" \n",
" When we launch a K8s pod we may not be able to connect to the metadata service immediately\n",
" because the ISTIO side car hasn't started.\n",
" \n",
" This function allows us to wait for a time specified up to stop_max_delay to see if the service\n",
" is ready. \n",
" \"\"\"\n",
" config = Configuration()\n",
" config.host = address\n",
" api_client = ApiClient(config)\n",
" client = MetadataServiceApi(api_client)\n",
"\n",
" client.list_artifacts2()\n",
" \n",
"def create_workspace():\n",
" return metadata.Workspace(\n",
" # Connect to metadata-service in namesapce kubeflow in k8s cluster.\n",
" backend_url_prefix=\"metadata-service.kubeflow.svc.cluster.local:8080\",\n",
" name=\"xgboost-synthetic\",\n",
" description=\"workspace for xgboost-synthetic artifacts and executions\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Wrap Training and Prediction in a class\n",
"\n",
"* In the cell below we wrap training and prediction in a class\n",
"* A class provides the structure we will need to eventually use kubeflow fairing to launch separate training jobs and/or deploy the model on Kubernetes"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# fairing:include-cell\n",
"class ModelServe(object): \n",
" def __init__(self, model_file=None):\n",
" self.n_estimators = 50\n",
" self.learning_rate = 0.1\n",
" if not model_file:\n",
" if \"MODEL_FILE\" in os.environ:\n",
" print(\"model_file not supplied; checking environment variable\")\n",
" model_file = os.getenv(\"MODEL_FILE\")\n",
" else:\n",
" print(\"model_file not supplied; using the default\")\n",
" model_file = \"mockup-model.dat\"\n",
" \n",
" self.model_file = model_file\n",
" print(\"model_file={0}\".format(self.model_file))\n",
" \n",
" self.model = None\n",
" self._workspace = None\n",
" self.exec = self.create_execution()\n",
"\n",
" def train(self):\n",
" (train_X, train_y), (test_X, test_y) = read_synthetic_input()\n",
" \n",
" # Here we use Kubeflow's metadata library to record information\n",
" # about the training run to Kubeflow's metadata store.\n",
" self.exec.log_input(metadata.DataSet(\n",
" description=\"xgboost synthetic data\",\n",
" name=\"synthetic-data\",\n",
" owner=\"someone@kubeflow.org\",\n",
" uri=\"file://path/to/dataset\",\n",
" version=\"v1.0.0\"))\n",
" \n",
" model = train_model(train_X,\n",
" train_y,\n",
" test_X,\n",
" test_y,\n",
" self.n_estimators,\n",
" self.learning_rate)\n",
"\n",
" mae = eval_model(model, test_X, test_y)\n",
" \n",
" # Here we log metrics about the model to Kubeflow's metadata store.\n",
" self.exec.log_output(metadata.Metrics(\n",
" name=\"xgboost-synthetic-traing-eval\",\n",
" owner=\"someone@kubeflow.org\",\n",
" description=\"training evaluation for xgboost synthetic\",\n",
" uri=\"gcs://path/to/metrics\",\n",
" metrics_type=metadata.Metrics.VALIDATION,\n",
" values={\"mean_absolute_error\": mae}))\n",
" \n",
" save_model(model, self.model_file)\n",
" self.exec.log_output(metadata.Model(\n",
" name=\"housing-price-model\",\n",
" description=\"housing price prediction model using synthetic data\",\n",
" owner=\"someone@kubeflow.org\",\n",
" uri=self.model_file,\n",
" model_type=\"linear_regression\",\n",
" training_framework={\n",
" \"name\": \"xgboost\",\n",
" \"version\": \"0.9.0\"\n",
" },\n",
" hyperparameters={\n",
" \"learning_rate\": self.learning_rate,\n",
" \"n_estimators\": self.n_estimators\n",
" },\n",
" version=datetime.utcnow().isoformat(\"T\")))\n",
" \n",
" def predict(self, X, feature_names):\n",
" \"\"\"Predict using the model for given ndarray.\n",
" \n",
" The predict signature should match the syntax expected by Seldon Core\n",
" https://github.com/SeldonIO/seldon-core so that we can use\n",
" Seldon h to wrap it a model server and deploy it on Kubernetes\n",
" \"\"\"\n",
" if not self.model:\n",
" self.model = joblib.load(self.model_file)\n",
" # Do any preprocessing\n",
" prediction = self.model.predict(data=X)\n",
" # Do any postprocessing\n",
" return [[prediction.item(0), prediction.item(1)]]\n",
"\n",
" @property\n",
" def workspace(self):\n",
" if not self._workspace:\n",
" wait_for_istio()\n",
" self._workspace = create_workspace()\n",
" return self._workspace\n",
" \n",
" def create_execution(self): \n",
" r = metadata.Run(\n",
" workspace=self.workspace,\n",
" name=\"xgboost-synthetic-faring-run\" + datetime.utcnow().isoformat(\"T\"),\n",
" description=\"a notebook run\")\n",
"\n",
" return metadata.Execution(\n",
" name = \"execution\" + datetime.utcnow().isoformat(\"T\"),\n",
" workspace=self.workspace,\n",
" run=r,\n",
" description=\"execution for training xgboost-synthetic\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Train your Model Locally\n",
"\n",
"* Train your model locally inside your notebook\n",
"* To train locally we just instatiante the ModelServe class and then call train"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"model = ModelServe(model_file=\"mockup-model.dat\")\n",
"model.train()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Predict locally\n",
"\n",
"* Run prediction inside the notebook using the newly created model\n",
"* To run prediction we just invoke redict"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"(train_X, train_y), (test_X, test_y) =read_synthetic_input()\n",
"\n",
"ModelServe().predict(test_X, None)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Use Kubeflow Fairing to Launch a K8s Job to train your model\n",
"\n",
"* Now that we have trained a model locally we can use Kubeflow fairing to\n",
" 1. Launch a Kubernetes job to train the model\n",
" 1. Deploy the model on Kubernetes\n",
"* Launching a separate Kubernetes job to train the model has the following advantages\n",
"\n",
" * You can leverage Kubernetes to run multiple training jobs in parallel \n",
" * You can run long running jobs without blocking your kernel"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Configure The Docker Registry For Kubeflow Fairing\n",
"\n",
"* In order to build docker images from your notebook we need a docker registry where the images will be stored\n",
"* Below you set some variables specifying a [GCR container registry](https://cloud.google.com/container-registry/docs/)\n",
"* Kubeflow Fairing provides a utility function to guess the name of your GCP project"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# Setting up google container repositories (GCR) for storing output containers\n",
"# You can use any docker container registry istead of GCR\n",
"GCP_PROJECT = fairing.cloud.gcp.guess_project_name()\n",
"DOCKER_REGISTRY = 'gcr.io/{}/fairing-job'.format(GCP_PROJECT)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Use Kubeflow fairing to build the docker image\n",
"\n",
"* First you will use kubeflow fairing's kaniko builder to build a docker image that includes all your dependencies\n",
" * You use kaniko because you want to be able to run `pip` to install dependencies\n",
" * Kaniko gives you the flexibility to build images from Dockerfiles\n",
"* kaniko, however, can be slow\n",
"* so you will build a base image using Kaniko and then every time your code changes you will just build an image\n",
" starting from your base image and adding your code to it\n",
"* you use the kubeflow fairing build to enable these fast rebuilds"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from kubeflow.fairing.builders import cluster\n",
"preprocessor = ConvertNotebookPreprocessorWithFire(class_name='ModelServe', notebook_file='build-train-deploy.ipynb')\n",
"\n",
"if not preprocessor.input_files:\n",
" preprocessor.input_files = set()\n",
"input_files=[\"xgboost_util.py\", \"mockup-model.dat\", \"requirements.txt\"]\n",
"preprocessor.input_files = set([os.path.normpath(f) for f in input_files])\n",
"preprocessor.preprocess()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Build the base image\n",
"\n",
"* You use cluster_builder to build the base image\n",
"* You only need to perform this again if we change our Docker image or the dependencies we need to install\n",
"* ClusterBuilder takes as input the DockerImage to use as a base image\n",
"* You should use the same Jupyter image that you are using for your notebook server so that your environment will be\n",
" the same when you launch Kubernetes jobs"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"scrolled": true
},
"outputs": [],
"source": [
"# Use a stock jupyter image as our base image\n",
"base_image = \"gcr.io/kubeflow-images-public/tensorflow-1.13.1-notebook-cpu:v0.5.0\"\n",
"\n",
"cluster_builder = cluster.cluster.ClusterBuilder(registry=DOCKER_REGISTRY,\n",
" base_image=base_image,\n",
" preprocessor=preprocessor,\n",
" pod_spec_mutators=[fairing.cloud.gcp.add_gcp_credentials_if_exists],\n",
" context_source=cluster.gcs_context.GCSContextSource())\n",
"cluster_builder.build()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Build the actual image\n",
"\n",
"Here you use the append builder to add your code to the base image\n",
"\n",
"* Calling preprocessor.preprocess() converts your notebook file to a python file\n",
"\n",
" * You are using the [ConvertNotebookPreprocessorWithFire](https://github.com/kubeflow/fairing/blob/master/fairing/preprocessors/converted_notebook.py#L85) \n",
" * This preprocessor converts ipynb files to py files by doing the following\n",
" 1. Removing all cells which don't have a comment `# fairing:include-cell`\n",
" 1. Using [python-fire](https://github.com/google/python-fire) to add entry points for the class specified in the constructor \n",
" \n",
" * Call preprocess() will create the file build-train-deploy.py\n",
" \n",
"* You use the AppendBuilder to rapidly build a new docker image by quickly adding some files to an existing docker image\n",
" * The AppendBuilder is super fast so its very convenient for rebuilding your images as you iterate on your code\n",
" * The AppendBuilder will add the converted notebook, build-train-deploy.py, along with any files specified in `preprocessor.input_files` to `/app` in the newly created image"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"scrolled": true
},
"outputs": [],
"source": [
"preprocessor.preprocess()\n",
"\n",
"builder = append.append.AppendBuilder(registry=DOCKER_REGISTRY,\n",
" base_image=cluster_builder.image_tag, preprocessor=preprocessor)\n",
"builder.build()\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Launch the K8s Job\n",
"\n",
"* You can use kubeflow fairing to easily launch a [Kubernetes job](https://kubernetes.io/docs/concepts/workloads/controllers/jobs-run-to-completion/) to invoke code\n",
"* You use fairings Kubernetes job library to build a Kubernetes job\n",
" * You use pod mutators to attach GCP credentials to the pod\n",
" * You can also use pod mutators to attch PVCs\n",
"* Since the [ConvertNotebookPreprocessorWithFire](https://github.com/kubeflow/fairing/blob/master/fairing/preprocessors/converted_notebook.py#L85) is using [python-fire](https://github.com/google/python-fire) you can easily invoke any method inside the ModelServe class just by configuring the command invoked by the Kubernetes job\n",
" * In the cell below you extend the command to include `train` as an argument because you want to invoke the train\n",
" function\n",
" \n",
"**Note** When you invoke train_deployer.deploy; kubeflow fairing will stream the logs from the Kubernetes job. The job will initially show some connection errors because the job will try to connect to the metadataserver. You can ignore these errors; the job will retry until its able to connect and then continue"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"pod_spec = builder.generate_pod_spec()\n",
"train_deployer = job.job.Job(cleanup=False,\n",
" pod_spec_mutators=[\n",
" fairing.cloud.gcp.add_gcp_credentials_if_exists])\n",
"\n",
"# Add command line arguments\n",
"pod_spec.containers[0].command.extend([\"train\"])\n",
"result = train_deployer.deploy(pod_spec)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"* You can use kubectl to inspect the job that fairing created"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"!kubectl get jobs -l fairing-id={train_deployer.job_id} -o yaml"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Deploy the trained model to Kubeflow for predictions"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"* Now that you have trained a model you can use kubeflow fairing to deploy it on Kubernetes\n",
"* When you call deployer.deploy fairing will create a [Kubernetes Deployment](https://kubernetes.io/docs/concepts/workloads/controllers/deployment/) to serve your model\n",
"* Kubeflow fairing uses the docker image you created earlier\n",
"* The docker image you created contains your code and [Seldon core](https://www.seldon.io/)\n",
"* Kubeflow fairing uses Seldon to wrap your prediction code, ModelServe.predict, in a REST and gRPC server"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from kubeflow.fairing.deployers import serving\n",
"pod_spec = builder.generate_pod_spec()\n",
"\n",
"module_name = os.path.splitext(preprocessor.executable.name)[0]\n",
"deployer = serving.serving.Serving(module_name + \".ModelServe\",\n",
" service_type=\"ClusterIP\",\n",
" labels={\"app\": \"mockup\"})\n",
" \n",
"url = deployer.deploy(pod_spec)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"* You can use kubectl to inspect the deployment that fairing created"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"!kubectl get deploy -o yaml {deployer.deployment.metadata.name}"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Send an inference request to the prediction server\n",
"\n",
"* Now that you have deployed the model into your Kubernetes cluster, you can send a REST request to \n",
" preform inference\n",
"* The code below reads some data, sends, a prediction request and then prints out the response"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"(train_X, train_y), (test_X, test_y) = read_synthetic_input()\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"full_url = url + \":5000/predict\"\n",
"result = util.predict_nparray(full_url, test_X)\n",
"pprint.pprint(result.content)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Clean up the prediction endpoint\n",
"\n",
"* You can use kubectl to delete the Kubernetes resources for your model\n",
"* If you want to delete the resources uncomment the following lines and run them"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# !kubectl delete service -l app=ames\n",
"# !kubectl delete deploy -l app=ames"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Track Models and Artifacts\n",
"\n",
"* Using Kubeflow's metadata server you can track models and artifacts\n",
"* The ModelServe code was instrumented to log executions and outputs\n",
"* You can access Kubeflow's metadata UI by selecting **Artifact Store** from the central dashboard\n",
" * See [here](https://www.kubeflow.org/docs/other-guides/accessing-uis/) for instructions on connecting to Kubeflow's UIs\n",
"* You can also use the python SDK to read and write entries\n",
"* This [notebook](https://github.com/kubeflow/metadata/blob/master/sdk/python/demo.ipynb) illustrates a bunch of metadata functionality"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Create a workspace\n",
"\n",
"* Kubeflow metadata uses workspaces as a logical grouping for artifacts, executions, and datasets that belong together\n",
"* Earlier in the notebook we defined the function `create_workspace` to create a workspace for this example\n",
"* You can use that function to return a workspace object and then call list to see all the artifacts in that workspace"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"ws = create_workspace()\n",
"ws.list()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Create a pipeline to train your model\n",
"\n",
"* [Kubeflow pipelines](https://www.kubeflow.org/docs/pipelines/) makes it easy to define complex workflows to build and deploy models\n",
"* Below you will define and run a simple one step pipeline to train your model\n",
"* Kubeflow pipelines uses experiments to group different runs of a pipeline together\n",
"* So you start by defining a name for your experiement"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"#### Define the pipeline\n",
"\n",
"* To create a pipeline you create a function and decorate it with the `@dsl.pipeline` decorator\n",
" * You use the decorator to give the pipeline a name and description\n",
" \n",
"* Inside the function, each step in the function is defined by a ContainerOp that specifies\n",
" a container to invoke\n",
" \n",
"* You will use the container image that you built earlier using Kubeflow Fairing\n",
"* Since the Kubeflow Fairing preprocessor added a main function using [python-fire](https://github.com/google/python-fire), a step in your pipeline can invocation any function in the ModelServe class just by setting the command for the container op\n",
"* See the pipelines [SDK reference](https://kubeflow-pipelines.readthedocs.io/en/latest/) for more information"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"@dsl.pipeline(\n",
" name='Training pipeline',\n",
" description='A pipeline that trains an xgboost model for the Ames dataset.'\n",
")\n",
"def train_pipeline(\n",
" ): \n",
" command=[\"python\", preprocessor.executable.name, \"train\"]\n",
" train_op = dsl.ContainerOp(\n",
" name=\"train\", \n",
" image=builder.image_tag, \n",
" command=command,\n",
" ).apply(\n",
" gcp.use_gcp_secret('user-gcp-sa'),\n",
" )\n",
" train_op.container.working_dir = \"/app\""
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"#### Compile the pipeline"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"* Pipelines need to be compiled"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"pipeline_func = train_pipeline\n",
"pipeline_filename = pipeline_func.__name__ + '.pipeline.zip'\n",
"compiler.Compiler().compile(pipeline_func, pipeline_filename)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"#### Submit the pipeline for execution"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"* Pipelines groups runs using experiments\n",
"* So before you submit a pipeline you need to create an experiment or pick an existing experiment\n",
"* Once you have compiled a pipeline, you can use the pipelines SDK to submit that pipeline\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"EXPERIMENT_NAME = 'MockupModel'\n",
"\n",
"#Specify pipeline argument values\n",
"arguments = {}\n",
"\n",
"# Get or create an experiment and submit a pipeline run\n",
"client = kfp.Client()\n",
"experiment = client.create_experiment(EXPERIMENT_NAME)\n",
"\n",
"#Submit a pipeline run\n",
"run_name = pipeline_func.__name__ + ' run'\n",
"run_result = client.run_pipeline(experiment.id, run_name, pipeline_filename, arguments)\n",
"\n",
"#vvvvvvvvv This link leads to the run information page. (Note: There is a bug in JupyterLab that modifies the URL and makes the link stop working)"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.6.7"
}
},
"nbformat": 4,
"nbformat_minor": 2
}
Reference in New Issue View Git Blame Copy Permalink