PipelineAI Home

PipelineAI Products

PipelineAI Deployments

PipelineAI Feature Comparison

Consistent, Immutable, Reproducible Model Runtimes

Each model is built into a separate Docker image with the appropriate Python, C++, and Java/Scala Runtime Libraries for training or prediction.

Use the same Docker Image from Local Laptop to Production to avoid dependency surprises.

Sample Machine Learning and AI Models

Click HERE to view model samples for the following:

Scikit-Learn
TensorFlow
Keras
Spark ML (formerly called Spark MLlib)
XgBoost
Custom Java
Custom Python
Ensembles

Coming Soon: Amazon MXNet, Microsoft CNTK, and Gluon

Supported Model Runtimes (CPU and GPU)

Python
Java
Scala
C++
Caffe2
Theano
TensorFlow Serving (TensorFlow Only)
Nvidia TensorRT (TensorFlow, Caffe2)

Coming Soon: Amazon MXNet, Microsoft CNTK, and Gluon

Pre-Requisites

Docker

Install Docker

Python2 or Python3 (Conda is Optional)

Install Miniconda with Python 2 or 3 (Preferred) Support

Install PipelineCLI

Note: This command line interface requires Python 2 or 3 and Docker as detailed above.

pip install cli-pipeline==1.4.17 --ignore-installed --no-cache -U

Verify Successful PipelineCLI Installation

pipeline version

### EXPECTED OUTPUT ###
cli_version: 1.4.x    <-- MAKE SURE THIS MATCHES THE VERSION YOU INSTALLED ABOVE
api_version: v1

default build type: docker
default build context path: . => ...

default train base image: docker.io/pipelineai/train:cpu-1.4.0     
default predict base image: docker.io/pipelineai/predict:cpu-1.4.0 

capabilities_enabled: ['train-server-*', 'predict-server-*', 'predict-test-http', 'version']
capabilities_available: ['train-cluster-*', 'predict-cluster-*', 'predict-test-stream', 'optimize-*']

Email [email protected] to enable the advanced capabilities.

Review CLI Functionality

Community Edition

Standalone Edition

Enterprise Edition

pipeline

### EXPECTED OUTPUT ###
Usage:       pipeline                             <-- This List of CLI Commands

(Standalone) pipeline optimize-predict-server     <-- Perform Model and Runtime Optimizations

(Community)  pipeline predict-test-http           <-- Predict Http-based Model Server or Cluster
             pipeline predict-test-stream         <-- Predict Kafka-based Model Server or Cluster
             
(Enterprise) pipeline predict-cluster-autoscale   <-- Configure AutoScaling for Model Cluster
             pipeline predict-cluster-connect     <-- Create Secure Tunnel to Model Cluster 
             pipeline predict-cluster-describe    <-- Describe Model Cluster
             pipeline predict-cluster-logs        <-- View Model Cluster Logs 
             pipeline predict-cluster-scale       <-- Scale Model Cluster
             pipeline predict-cluster-shell       <-- Shell into Model Cluster
             pipeline predict-cluster-start       <-- Start Model Cluster from Docker Registry
             pipeline predict-cluster-status      <-- Status of Model Cluster
             pipeline predict-cluster-stop        <-- Stop Model Cluster
             
(Community)  pipeline predict-server-build        <-- Build Model Server
             pipeline predict-server-logs         <-- View Model Server Logs
             pipeline predict-server-pull         <-- Pull Model Server from Docker Registry
             pipeline predict-server-push         <-- Push Model Server to Docker Registry
             pipeline predict-server-shell        <-- Shell into Model Server (Debugging)
             pipeline predict-server-start        <-- Start Model Server
             pipeline predict-server-stop         <-- Stop Model Server

(Enterprise) pipeline train-cluster-connect       <-- Create Secure Tunnel to Training Cluster
             pipeline train-cluster-describe      <-- Describe Training Cluster
             pipeline train-cluster-logs          <-- View Training Cluster Logs
             pipeline train-cluster-scale         <-- Scale Training Cluster
             pipeline train-cluster-shell         <-- Shell into Training Cluster
             pipeline train-cluster-start         <-- Start Training Cluster from Docker Registry
             pipeline train-cluster-status        <-- Status of Training Cluster
             pipeline train-cluster-stop          <-- Stop Training Cluster

(Standalone) pipeline train-server-build          <-- Build Training Server
             pipeline train-server-logs           <-- View Training Server Logs
             pipeline train-server-pull           <-- Pull Training Server from Docker Registry
             pipeline train-server-push           <-- Push Training Server to Docker Registry
             pipeline train-server-shell          <-- Shell into Training Server (Debugging)
             pipeline train-server-start          <-- Start Training Server
             pipeline train-server-stop           <-- Stop Training Server
             
(Community)  pipeline version                     <-- View This CLI Version

Step 1: Retrieve Sample Models

Clone the PipelineAI Predict Repo

git clone https://github.com/PipelineAI/models

Change into `models` Directory

cd ./models

Switch to Latest Branch (master)

Note: Master may be unstable. See Releases Tab for stable releases.

git checkout master

Step 2: Train a Model

Inspect Model Directory

ls -l ./tensorflow/census

### EXPECTED OUTPUT ###
...
pipeline_conda_environment.yml     <-- Required.  Sets up the conda environment
pipeline_condarc                   <-- Required.  Configure Conda proxy servers (.condarc)
pipeline_train.py                  <-- Required.  `main()` is required. Args passed through `--train-args`
...

Build Training Server

Note: --model-path must be relative. On Windows, be sure to use the forward slash \ for --model-path.

pipeline train-server-build --model-type=tensorflow --model-name=census --model-tag=v1 --model-path=./tensorflow/census

Note: If you see CondaHTTPError: HTTP 000 CONNECTION FAILED for url or [Errno 111] Connection refused' or ConnectionError(MaxRetryError("HTTPSConnectionPool, you need to update ./tensorflow/census/pipeline_condarc to include proxy servers per THIS document.

Start Training UI

Note the following:

--train-args is a single argument passed into the pipeline_train.py. Therefore, you must escape spaces (\ ) between arguments.
--input-path and --output-path are relative to the current working directory (outside the Docker container) and will be mapped as directories inside the Docker container from /root.
--train-files and --eval-files are relative to --input-path inside the Docker container.
Models, logs, and event are written to --output-path (or a subdirectory within). These will be available outside of the Docker container.
To prevent overwriting the output of a previous run, you should either 1) change the --output-path between calls or 2) create a new unique subfolder with --output-path in your pipeline_train.py (ie. timestamp). See examples below.

(We are working on making these more intuitive.)

Note: On Windows, be sure to use the forward slash \ for --input-path and --output-path (not the args inside of --train-args).

pipeline train-server-start --model-type=tensorflow --model-name=census --model-tag=v1 --input-path=./tensorflow/census/data --output-path=./tensorflow/census/versions --train-args="--train-files=train/adult.data.csv\ --eval-files=eval/adult.test.csv\ --num-epochs=2\ --learning-rate=0.025"

Note: If you see port is already allocated or already in use by container, you already have a container running. List and remove any conflicting containers. For example, docker ps and/or docker rm -f train-tfserving-tensorflow-census-v1.

View Training Logs

pipeline train-server-logs --model-type=tensorflow --model-name=census --model-tag=v1

Press Ctrl-C to exit out of the logs.

View Trained Model Output (Locally)

Make sure you pressed Ctrl-C to exit out of the logs.

ls -l ./tensorflow/census/versions/

### EXPECTED OUTPUT ###
...
drwxr-xr-x  11 cfregly  staff  352 Nov 22 11:20 1511367633 
drwxr-xr-x  11 cfregly  staff  352 Nov 22 11:21 1511367665
drwxr-xr-x  11 cfregly  staff  352 Nov 22 11:22 1511367765 <= Sub-directories of training output
...

Multiple training runs will produce multiple subdirectories - each with a different timestamp.

View Training UI (including TensorBoard for TensorFlow Models)

http://localhost:6007

Stop Training UI

pipeline train-server-stop --model-type=tensorflow --model-name=census --model-tag=v1

Step 3: Predict with Model

Inspect Model Directory

Note: This is relative to where you cloned the models repo above.

ls -l ./tensorflow/mnist

### EXPECTED OUTPUT ###
...
pipeline_conda_environment.yml     <-- Required.  Sets up the conda environment
pipeline_condarc                   <-- Required.  Configure Conda proxy servers (.condarc)
pipeline_predict.py                <-- Required.  `predict(request: bytes) -> bytes` is required
versions/                          <-- Optional.  TensorFlow Serving requires this directory
...

Inspect Trained Models

ls -l ./tensorflow/mnist/versions/

### EXPECTED OUTPUT ###
...

drwxr-xr-x  11 cfregly  staff  352 Nov 20 12:07 1510612525  
drwxr-xr-x  11 cfregly  staff  352 Nov 20 12:18 1510612528   <-- Serves the highest (latest) version

Build the Model into a Runnable Docker Image

This command bundles the TensorFlow runtime with the model.

Note: --model-path must be relative. On Windows, be sure to use the forward slash \ for --model-path.

pipeline predict-server-build --model-type=tensorflow --model-name=mnist --model-tag=v1 --model-path=./tensorflow/mnist

Start the Model Server

pipeline predict-server-start --model-type=tensorflow --model-name=mnist --model-tag=v1 --memory-limit=2G

Inspect `pipeline_predict.py`

Note: Only the predict() method is required. Everything else is optional.

cat ./pipeline_predict.py

### EXPECTED OUTPUT ###
import os
import logging
from pipeline_model import TensorFlowServingModel             <-- Optional.  Wraps TensorFlow Serving
from pipeline_monitor import prometheus_monitor as monitor    <-- Optional.  Monitor runtime metrics
from pipeline_logger import log                               <-- Optional.  Log to console, file, kafka

...

__all__ = ['predict']                                         <-- Optional.  Being a good Python citizen.

...

def _initialize_upon_import() -> TensorFlowServingModel:      <-- Optional.  Called once at server startup
    return TensorFlowServingModel(host='localhost',           <-- Optional.  Wraps TensorFlow Serving
                                  port=9000,
                                  model_name=os.environ['PIPELINE_MODEL_NAME'],
                                  inputs_name='inputs',       <-- Optional.  TensorFlow SignatureDef inputs
                                  outputs_name='outputs',     <-- Optional.  TensorFlow SignatureDef outputs
                                  timeout=100)                <-- Optional.  TensorFlow Serving timeout

_model = _initialize_upon_import()                            <-- Optional.  Called once upon server startup

_labels = {'model_runtime': os.environ['PIPELINE_MODEL_RUNTIME'],  <-- Optional.  Tag metrics
           'model_type': os.environ['PIPELINE_MODEL_TYPE'],   
           'model_name': os.environ['PIPELINE_MODEL_NAME'],
           'model_tag': os.environ['PIPELINE_MODEL_TAG']}

_logger = logging.getLogger('predict-logger')                 <-- Optional.  Standard Python logging

@log(labels=_labels, logger=_logger)                          <-- Optional.  Sample and compare predictions
def predict(request: bytes) -> bytes:                         <-- Required.  Called on every prediction

    with monitor(labels=_labels, name="transform_request"):   <-- Optional.  Expose fine-grained metrics
        transformed_request = _transform_request(request)     <-- Optional.  Transform input (json) into TensorFlow (tensor)

    with monitor(labels=_labels, name="predict"):
        predictions = _model.predict(transformed_request)       <-- Optional.  Calls _model.predict()

    with monitor(labels=_labels, name="transform_response"):
        transformed_response = _transform_response(predictions) <-- Optional.  Transform TensorFlow (tensor) into output (json)

    return transformed_response                                 <-- Required.  Returns the predicted value(s)
...

Monitor Runtime Logs

Wait for the model runtime to settle...

pipeline predict-server-logs --model-type=tensorflow --model-name=mnist --model-tag=v1

### EXPECTED OUTPUT ###
...
2017-10-10 03:56:00.695  INFO 121 --- [     run-main-0] i.p.predict.jvm.PredictionServiceMain$   : Started PredictionServiceMain. in 7.566 seconds (JVM running for 20.739)
[debug] 	Thread run-main-0 exited.
[debug] Waiting for thread container-0 to terminate.
...
INFO[0050] Completed initial partial maintenance sweep through 4 in-memory fingerprints in 40.002264633s.  source="storage.go:1398"
...

You need to Ctrl-C out of the log viewing before proceeding.

Perform Prediction

Before proceeding, make sure you hit Ctrl-C after viewing the logs in the previous step.

You may see 502 Bad Gateway or '{"results":["fallback"]}' if you predict too quickly. Let the server settle a bit - and try again.