GitHub - chzbrgr71/got-image-classification: Demo: End to end Machine Learning on Kubernetes

Game of Thrones Image Classification Demo

This is demo code for my talk at Open Source Summit NA 2019.

"Machine Learning Made Easy on Kubernetes. DevOps for Data Scientists," August 21, 2019

Game of Thrones Characters

Robert Baratheon (robert-baratheon)
Tyrion Lannister (tyrion-lannister)
Jon Snow (jon-snow)
Daenerys Targaryen (daenerys-targaryen)
Hodor (hodor)
Samwell Tarley (samwell-tarley)
Cersei Lannister (cersei-lannister)
Theon Greyjoy (theon-greyjoy)
Night King (night-king)
Arya Stark (arya-stark)
Benjen Stark (benjen-stark)
Jamie Lannister (jamie-lannister)
Margaery Tyrell (margaery-tyrell)
Sansa Stark (sansa-stark)

Local testing/training

Testing in local Docker container interactively

docker run -it --rm --name got \
  --publish 6006:6006 \
  --publish 5000:5000 \
  --volume /Users/brianredmond/gopath/src/github.com/chzbrgr71/got-image-classification:/got-image-classification \
  --workdir /got-image-classification \
tensorflow/tensorflow:1.13.1

python ./training/retrain.py \
  --bottleneck_dir=/got-image-classification/tf-output/bottlenecks \
  --model_dir=/tmp/tensorflow/inception \
  --summaries_dir=/got-image-classification/tf-output \
  --output_graph=/got-image-classification/tf-output \
  --output_labels=/got-image-classification/tf-output \
  --image_dir=/got-image-classification/training/images \
  --saved_model_dir=/got-image-classification/tf-output \
  --how_many_training_steps 2000

python ./preprocess/processimages.py \
  --bottleneck_dir=/got-image-classification/tf-output/bottlenecks \
  --image_dir=/got-image-classification/preprocess/images 

# or conda
source activate tf

python ./training/retrain.py \
  --bottleneck_dir=./tf-output/bottlenecks \
  --model_dir=./tf-output/inception \
  --summaries_dir=./tf-output \
  --output_graph=./tf-output \
  --output_labels=./tf-output \
  --image_dir=./training/images \
  --saved_model_dir=./tf-output \
  --how_many_training_steps 2000  

tensorboard --logdir=/got-image-classification/tf-output/training_summaries

Create preprocess container image

# set image tag depending on target cpu/gpu
export IMAGE_TAG=2.00
export ACRNAME=gotcr

# build/push (ACR or Docker)
az acr build -t chzbrgr71/got-image-preprocess:$IMAGE_TAG -r $ACRNAME ./preprocess

docker build -t chzbrgr71/got-image-preprocess:$IMAGE_TAG -f ./preprocess/Dockerfile ./preprocess
docker push chzbrgr71/got-image-preprocess:$IMAGE_TAG

Create training container image

# set image tag depending on target cpu/gpu
export IMAGE_TAG=2.0
export IMAGE_TAG=2.00-gpu
export ACRNAME=gotcr

# build/push (ACR or Docker)
az acr build -t chzbrgr71/got-image-training:$IMAGE_TAG -r $ACRNAME ./training

docker build -t chzbrgr71/got-image-training:$IMAGE_TAG -f ./training/Dockerfile ./training
docker push chzbrgr71/got-image-training:$IMAGE_TAG

Create scoring container image

# set image tag depending on target cpu/gpu
export IMAGE_TAG=2.01
export ACRNAME=gotcr

# build/push (ACR or Docker)
az acr build -t chzbrgr71/got-model-scoring:$IMAGE_TAG -r $ACRNAME ./serving

docker build -t chzbrgr71/got-model-scoring:$IMAGE_TAG -f ./serving/Dockerfile ./serving
docker push chzbrgr71/got-model-scoring:$IMAGE_TAG

docker run -d --name score --volume /Users/brianredmond/gopath/src/github.com/chzbrgr71/got-image-classification:/got-image-classification chzbrgr71/got-model-scoring:$IMAGE_TAG '/got-image-classification/tf-output/latest_model'

Test local

docker run -d --name process --volume /Users/brianredmond/gopath/src/github.com/chzbrgr71/got-image-classification:/got-image-classification chzbrgr71/got-image-preprocess:$IMAGE_TAG "--bottleneck_dir=/got-image-classification/tf-output/bottlenecks" "--image_dir=/got-image-classification/preprocess/images"

docker run -d --name train --volume /Users/brianredmond/gopath/src/github.com/chzbrgr71/got-image-classification:/got-image-classification chzbrgr71/got-image-training:$IMAGE_TAG "--bottleneck_dir=/tmp/tensorflow/bottlenecks" "--model_dir=/tmp/tensorflow/inception" "--summaries_dir=/got-image-classification/tf-output/training_summaries/baseline" "--output_graph=/got-image-classification/tf-output/retrained_graph.pb" "--output_labels=/got-image-classification/tf-output/retrained_labels.txt" "--image_dir=/images" "--saved_model_dir=/got-image-classification/tf-output/saved_models/1"

Tensorboard local

export IMAGE_TAG=2.00
export ACRNAME=gotcr

# build/push (ACR or Docker)
az acr build -t chzbrgr71/tensorboard:$IMAGE_TAG -r $ACRNAME -f ./tensorboard/Dockerfile ./tensorboard

docker build -t chzbrgr71/tensorboard:$IMAGE_TAG -f ./tensorboard/Dockerfile ./tensorboard
docker push chzbrgr71/tensorboard:$IMAGE_TAG

# run
docker run -d --name tb -p 6006:6006 --volume /Users/brianredmond/gopath/src/github.com/chzbrgr71/got-image-classification/tf-output:/tf-output chzbrgr71/tensorboard:$IMAGE_TAG "--logdir" "/tf-output/training_summaries"

Kubernetes Setup

Create Azure Kubernetes Service

Use node pools to add GPU nodes. https://docs.microsoft.com/en-us/azure/aks/use-multiple-node-pools

Add Virtual Node add-on

export VK_RELEASE=virtual-kubelet-latest
export CHART_URL=https://github.com/virtual-kubelet/virtual-kubelet/raw/master/charts/$VK_RELEASE.tgz
helm install --name vk "$CHART_URL" --namespace kube-system -f ./k8s/vk-helm-values.yaml

Enable GPU's with daemonset. https://docs.microsoft.com/en-us/azure/aks/gpu-cluster

kubectl create namespace gpu-resources
kubectl apply -f ./k8s/nvidia-device-plugin-ds.yaml

Storage (Azure Files Static)

Azure Files Docs: https://docs.microsoft.com/en-us/azure/aks/azure-files-volume

export AKS_PERS_STORAGE_ACCOUNT_NAME=briarmlwestnew
export AKS_PERS_RESOURCE_GROUP=oss-summit-west-new
export AKS_PERS_LOCATION=westus
export AKS_PERS_SHARE_NAME=aksshare

# Create the storage account
az storage account create -n $AKS_PERS_STORAGE_ACCOUNT_NAME -g $AKS_PERS_RESOURCE_GROUP -l $AKS_PERS_LOCATION --sku Standard_LRS

# Export the connection string as an environment variable, this is used when creating the Azure file share
export AZURE_STORAGE_CONNECTION_STRING=`az storage account show-connection-string -n $AKS_PERS_STORAGE_ACCOUNT_NAME -g $AKS_PERS_RESOURCE_GROUP -o tsv`

# Create the file share
az storage share create -n $AKS_PERS_SHARE_NAME

# Get storage account key
STORAGE_KEY=$(az storage account keys list --resource-group $AKS_PERS_RESOURCE_GROUP --account-name $AKS_PERS_STORAGE_ACCOUNT_NAME --query "[0].value" -o tsv)

# Echo storage account name and key
echo Storage account name: $AKS_PERS_STORAGE_ACCOUNT_NAME
echo Storage account key: $STORAGE_KEY

kubectl create secret generic azure-file-secret --from-literal=azurestorageaccountname=$AKS_PERS_STORAGE_ACCOUNT_NAME --from-literal=azurestorageaccountkey=$STORAGE_KEY

kubectl create secret generic azure-file-secret --from-literal=azurestorageaccountname=$AKS_PERS_STORAGE_ACCOUNT_NAME --from-literal=azurestorageaccountkey=$STORAGE_KEY --namespace kubeflow

# Add persistent volume
kubectl apply -f ./k8s/persistent-volume.yaml

Kubernetes Jobs

Kubernetes job

kubectl apply -f ./k8s/job-preprocess.yaml

kubectl apply -f ./k8s/job-training.yaml

kubectl apply -f ./k8s/tensorboard.yaml

Kubeflow

Install Kubeflow (I am using v0.5.0) https://www.kubeflow.org/docs/started/getting-started-k8s

export KFAPP=kf-app-got-3
kfctl init ${KFAPP}
cd ${KFAPP}
kfctl generate all -V
kfctl apply all -V

Validate Kubeflow
```
kubectl -n kubeflow get all
```

Execute TFJob

# cpu
kubectl apply -f ./k8s/tfjob-training.yaml

# gpu
kubectl apply -f ./k8s/tfjob-training-gpu.yaml

# aci
kubectl apply -f ./k8s/tfjob-training-vk.yaml

# Azure Premium Files
kubectl apply -f ./k8s/tfjob-training-prem.yaml

Deploy Serving

kubectl apply -f ./k8s/serving.yaml

docker cp tf-output/latest_model/exported_model serving_base:/models/inception

docker commit --change "ENV MODEL_NAME inception" serving_base chzbrgr71/got-tfserving:1.0

Deploy Tensorboard
```
kubectl apply -f ./k8s/tensorboard.yaml
```
Kubeflow Pipelines
- Create a clean Python 3 environment
```
conda create --name mlpipeline python=3.7
source activate mlpipeline
pip install -r ./pipelines/requirements.txt --upgrade
```
- Compile pipeline
```
source activate mlpipeline
python3 ./pipelines/pipeline.py
```
- For now, there are a couple manual edits needed on the pipeline.yaml
  - environment variables (KUBE_POD_NAME in training)
  - volumes for Azure files

Jupyter Notebooks

Via port-forward to notebook instance
```
kubectl port-forward brian-testing-01-0 -n kubeflow 8888:8888
```
http://localhost:8888/notebook/kubeflow/brian-testing-01
Access via Traefik ingress. http://jupyter.brianredmond.io/notebook/kubeflow/brian-testing-01

Hyperparameter Optimization

Using Helm chart and AKS Virtual Nodes with GPU:

helm install --name hyperparam ./hyperparameter/chart

kubectl apply -f ./hyperparameter/tensorboard-hp.yaml

katib

kubectl apply -f ./katib/random-example.yaml

kubectl apply -f ./katib/got.yaml

Inference

Local python script

# testing
python ./serving/label-image.py ./serving/benjen.jpg

hodor (score = 0.35062)
benjen stark (score = 0.21019)
samwell tarley (score = 0.13798)
jon snow (score = 0.10155)
robert baratheon (score = 0.04643)
theon greyjoy (score = 0.04288)
daenerys targaryen (score = 0.03613)
tyrion lannister (score = 0.02663)
night king (score = 0.02428)
margaery tyrell (score = 0.00809)
cersei lannister (score = 0.00707)
arya stark (score = 0.00544)
sansa stark (score = 0.00271)

TF Serving (Local)

docker run -d --name serving \
  --publish 8500:8500 \
  --volume /Users/brianredmond/gopath/src/github.com/chzbrgr71/got-image-classification/tf-output/saved_models:/models/inception \
  --env MODEL_NAME=inception \
tensorflow/serving:1.13.0

python serving/inception_client.py --server localhost:8500 --image ./serving/hodor.jpg
python serving/inception_client.py --server localhost:8500 --image ./serving/tyrion.jpg
python serving/inception_client.py --server localhost:8500 --image ./serving/night-king.jpg

TF Serving (AKS)

kubectl apply -f ./k8s/serving.yaml

python serving/inception_client.py --server 104.45.210.253:8500 --image ./serving/night-king.jpg
python serving/inception_client.py --server gotserving.brianredmond.io:8500 --image ./serving/jon-snow.jpg
python serving/inception_client.py --server gotserving.brianredmond.io:8500 --image ./serving/benjen.jpg

# serving api metadata
curl http://gotserving.eastus.azurecontainer.io:8501/v1/models/inception/versions/1/metadata

# convert image to base64: https://onlinepngtools.com/convert-png-to-base64

curl -X POST http://gotserving.eastus.azurecontainer.io:8501/v1/models/inception:predict -d "@./serving/daenerys-targaryen.json"
curl -X POST http://gotserving.brianredmond.io:8501/v1/models/inception:predict -d "@./serving/daenerys-targaryen.json"
curl -X POST http://104.45.210.253:8501/v1/models/inception:predict -d "@./serving/daenerys-targaryen.json"

Web App

export IMAGE_TAG=2.01
export ACRNAME=gotcr

# build/push (ACR or Docker)
az acr build -t chzbrgr71/got-web-app:$IMAGE_TAG -r $ACRNAME ./webapp

docker build -t chzbrgr71/got-web-app:$IMAGE_TAG -f ./webapp/Dockerfile ./webapp
docker push chzbrgr71/got-web-app:$IMAGE_TAG

docker run -d --name web -p 3000:3000 -e ML_SERVING_ENDPOINT=http://gotserving.brianredmond.io:8501/v1/models/inception:predict chzbrgr71/got-web-app:$IMAGE_TAG

kubectl apply -f ./k8s/web.yaml

az webapp config appsettings set --name got-web -g game-of-thrones --settings ML_SERVING_ENDPOINT='http://gotserving.brianredmond.io:8501/v1/models/inception:predict'

Tensorflow Lite

Convert model

IMAGE_SIZE=299
tflite_convert \
  --graph_def_file=./tf-output/latest_model/got_retrained_graph.pb \
  --output_file=./tf-output/latest_model/optimized_graph.lite \
  --input_format=TENSORFLOW_GRAPHDEF \
  --output_format=TFLITE \
  --input_shape=1,${IMAGE_SIZE},${IMAGE_SIZE},3 \
  --input_array=Mul \
  --output_array=final_result \
  --inference_type=FLOAT \
  --input_data_type=FLOAT

In container

export IMAGE_TAG=2.00
export ACRNAME=gotcr

# build/push (ACR or Docker)
az acr build -t chzbrgr71/tflite-convert:$IMAGE_TAG -r $ACRNAME -f ./convert/Dockerfile ./convert

docker build -t chzbrgr71/tflite-convert:$IMAGE_TAG -f ./convert/Dockerfile ./convert
docker push chzbrgr71/tflite-convert:$IMAGE_TAG

# run
docker run -d --name convert --volume /Users/brianredmond/gopath/src/github.com/chzbrgr71/got-image-classification/tf-output:/tf-output chzbrgr71/tflite-convert:$IMAGE_TAG \
  --graph_def_file=./tf-output/latest_model/got_retrained_graph.pb \
  --output_file=./tf-output/latest_model/optimized_graph.lite \
  --input_format=TENSORFLOW_GRAPHDEF \
  --output_format=TFLITE \
  --input_shape=1,299,299,3 \
  --input_array=Mul \
  --output_array=final_result \
  --inference_type=FLOAT \
  --input_data_type=FLOAT

kubectl apply -f ./k8s/convert.yaml

ONNX

This doesn't work at all:

source activate mlpipeline

python -m tf2onnx.convert \
  --saved-model ./tf-output/latest_model/exported_model/1/ \
  --output ./tf-output/onnx/model.onnx \
  --verbose

python -m tf2onnx.convert \
  --input ./tf-output/latest_model/got_retrained_graph.pb \
  --inputs DecodeJpeg/contents:0 \
  --outputs final_result:0 \
  --output ./tf-output/onnx/model.onnx \
  --verbose

saved_model_cli show --dir /got-image-classification/tf-output/latest_model/exported_model/1/ --tag_set serve --signature_def serving_default

export IMAGE_TAG=2.00
export ACRNAME=gotcr

# build/push (ACR or Docker)
az acr build -t chzbrgr71/onnx-convert:$IMAGE_TAG -r $ACRNAME -f ./onnx/Dockerfile ./onnx

docker build -t chzbrgr71/onnx-convert:$IMAGE_TAG -f ./onnx/Dockerfile ./onnx
docker push chzbrgr71/onnx-convert:$IMAGE_TAG

docker run -d --name onnx --volume /Users/brianredmond/gopath/src/github.com/chzbrgr71/got-image-classification:/got-image-classification chzbrgr71/onnx-convert:1.1 "show" "--dir" "/got-image-classification/tf-output/latest_model/exported_model/1/" "--tag_set" "serve" "--signature_def" "serving_default"

Tensorflow.js

This doesn't work at all:

pip install tensorflowjs==0.8.5 --force-reinstall
pip install tensorflowjs==1.0.1 --force-reinstall

tensorflowjs_converter \
    --input_format=tf_saved_model \
    --output_format=tfjs_graph_model \
    --skip_op_check SKIP_OP_CHECK \
    ./tf-output/latest_model/got_retrained_graph.pb \
    ./tf-output/javascript

tensorflowjs_converter \
    --input_format=tf_saved_model \
    --output_format=tfjs_graph_model \
    --skip_op_check SKIP_OP_CHECK \
    ./tf-output/latest_model/exported_model/1 \
    ./tf-output/javascript

    --output_node_names='final_result' \

Reference Links

From: https://gameofthrones.fandom.com

JSON: https://raw.githubusercontent.com/jeffreylancaster/game-of-thrones/master/data/characters.json

Image downloader: https://github.com/teracow/googliser

https://codelabs.developers.google.com/codelabs/tensorflow-for-poets/#0

https://www.tensorflow.org/lite/guide/get_started

https://heartbeat.fritz.ai/intro-to-machine-learning-on-android-how-to-convert-a-custom-model-to-tensorflow-lite-e07d2d9d50e3

https://www.tensorflow.org/js/tutorials/conversion/import_keras#alternative_use_the_python_api_to_export_directly_to_tfjs_layers_format

https://becominghuman.ai/creating-restful-api-to-tensorflow-models-c5c57b692c10

https://codelabs.developers.google.com/codelabs/tensorflowjs-teachablemachine-codelab/index.html#0

https://www.tensorflow.org/hub/tutorials/image_retraining

https://medium.com/codait/bring-machine-learning-to-the-browser-with-tensorflow-js-part-iii-62d2b09b10a3

https://github.com/vabarbosa/tfjs-model-playground/tree/master/image-segmenter/demo

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Game of Thrones Image Classification Demo

Game of Thrones Characters

Local testing/training

Kubernetes Setup

Kubernetes Jobs

Kubeflow

Jupyter Notebooks

Hyperparameter Optimization

Inference

Tensorflow Lite

ONNX

Tensorflow.js

Reference Links

About

Releases

Packages

Contributors 2

Languages

Name		Name	Last commit message	Last commit date
Latest commit History 31 Commits
convert		convert
hyperparameter		hyperparameter
img		img
k8s		k8s
katib		katib
notebook		notebook
onnx		onnx
pipelines		pipelines
preprocess		preprocess
serving		serving
tensorboard		tensorboard
training		training
webapp		webapp
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
demos.md		demos.md
ml-slides.pdf		ml-slides.pdf

License

chzbrgr71/got-image-classification

Folders and files

Latest commit

History

Repository files navigation

Game of Thrones Image Classification Demo

Game of Thrones Characters

Local testing/training

Kubernetes Setup

Kubernetes Jobs

Kubeflow

Jupyter Notebooks

Hyperparameter Optimization

Inference

Tensorflow Lite

ONNX

Tensorflow.js

Reference Links

About

Topics

Resources

License

Stars

Watchers

Forks

Releases

Packages 0

Contributors 2

Languages

Packages