Sentiment Prediction Demo

A worked step-by-step example using the Amazon Fine Foods Review dataset is provided here. The dataset provides around 500,000 reviews and ratings from Amazon for fine food products. We will use it to create a sentiment predictor that takes the review text and predicts whether it is a positive or negative review. A model will be created and run as a microservice within Seldon.


Create Model

The model creation step can be accomplished by running the Kubernetes deployment in kubernetes/conf/examples/finefoods/train-finefoods.json which should have been created when you followed the install and configuration steps.

Create the kubernetes job:

cd kubernetes/conf/examples/finefoods
kubectl create -f train-finefoods.json

The job may take a 10-20 minutes to fully run. You can check its status with kubectl get jobs -l job-name=train-finefoods.

This job will:

Serve Predictions

To serve predictions we will load the saved pipeline into a microservice. This can be accomplished by using the script start-microservice in seldon-server/kubernetes/bin.

start-microservice --type prediction --client test -p finefoods-xgboost /seldon-data/seldon-models/finefoods/1/ 1.0

This will load the pipeline saved in /seldon-data/seldon-models/finefoods/1/ and create a single replica microservice called finefoods-xgboost. It will activate this for the “test” client as a prediction algorithm.

You can then test the predictions, e.g.:

seldon-cli api --client-name test --endpoint /js/predict --json '{"data":{"review":"The candy is just red.   No flavor. Just  plan and chewy.  I would never buy them again."}}'

Which should produce a result like below indication a “negative” sentiment:

  "size": 2,
  "requested": 0,
  "list": [
      "prediction": 0.594055,
      "predictedClass": "0",
      "confidence": 0.594055
      "prediction": 0.405945,
      "predictedClass": "1",
      "confidence": 0.405945

While the following:

seldon-cli api --client-name test --endpoint /js/predict --json '{"data":{"review":"My daughter loves twizzlers and this shipment of six pounds really hit the spot"}}'

Should produce a “positive” sentiment:

  "size": 2,
  "requested": 0,
  "list": [
      "prediction": 0.268916,
      "predictedClass": "0",
      "confidence": 0.268916
      "prediction": 0.731084,
      "predictedClass": "1",
      "confidence": 0.731084

Detailed Steps

The model creation steps are provided by the Docker image seldonio/finefoods_xgboost whose code can be found in seldon-server/docker/examples/finefoods. There is a script which is shown below:


set -o nounset
set -o errexit

if [ "$#" -ne 2 ]; then
    echo "need <model_folder> <sample in range 0.0 to 1.0>"
    exit -1

STARTUP_DIR="$( cd "$( dirname "$0" )" && pwd )"

function get_data {
    cd ${STARTUP_DIR}
    echo "nameserver" >> /etc/resolv.conf
    gunzip finefoods.txt.gz
    iconv -f iso-8859-1 -t utf-8 finefoods.txt -o finefoods_utf8.txt
    mkdir -p data
    echo "sentiment,review" > data/train.csv
    cat finefoods_utf8.txt | awk 'BEGIN{OFS=","}/^review\/score/{if ($2<3){s=0} else if ($2>3){s=1}else{s=-1}}/^review\/text/{if (s>-1){gsub(/,/," ");print s,$0}}' | sed -e 's/review\/text://g' -e 's/<br \/>//g' -e 's/\./ /g' >> data/train.csv


function train {

    mkdir ${STARTUP_DIR}/folds
    python ${STARTUP_DIR}/ --data ${STARTUP_DIR}/data --model ${MODEL_FOLDER} --sample ${SAMPLE}
    perf -SEN -SPC -NPV -PPV -ACC -confusion -files folds/1_correct.txt <(cut -d' ' -f2 folds/1_predictions_proba.txt)



The script:

The model creation in uses the pyseldon package to build a sklearn pipeline as show below:

        tTfidf = ptfidf.Tfidf_transform(input_feature="review",output_feature="tfidf",target_feature="sentiment",min_df=10,max_df=0.7,select_features=False,topn_features=50000,stop_words="english",ngram_range=[1,2])

        tFilter2 = bt.Include_features_transform(included=["tfidf","sentiment"])

        svmTransform = bt.Svmlight_transform(output_feature="svmfeatures",excluded=["sentiment"],zero_based=False)

        classifier_xg = xg.XGBoostClassifier(target="sentiment",svmlight_feature="svmfeatures",silent=1,max_depth=5,n_estimators=200,objective='binary:logistic',scale_pos_weight=0.2)

        cv = cf.Seldon_KFold(classifier_xg,metric='auc',save_folds_folder="./folds")
        transformers = [("tTfidf",tTfidf),("tFilter2",tFilter2),("svmTransform",svmTransform),("cv",cv)]

        p = Pipeline(transformers)

The script saves the pipleline after modelling so that it can be reloaded and run in prediction mode to score new data.