Sending an object dtype array as the request JSON for a Model API

[nsriram13]

I am building an sklearn pipeline that involves some pre-processing (specifically one-hot-encoding categorical variables and scaling numeric variables) which then gets forwarded to a classifier. I have provided the pipeline below (inspired by this blog):

import joblib
import pandas as pd

from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.compose import ColumnTransformer, make_column_transformer
from sklearn.model_selection import train_test_split
from sklearn.pipeline import make_pipeline
from sklearn.linear_model import LogisticRegression
from sklearn.impute import SimpleImputer

titanic = pd.read_csv("https://raw.githubusercontent.com/amueller/scipy-2017-sklearn/master/notebooks/datasets/titanic3.csv")
target = titanic.survived.values
features = titanic[['pclass', 'sex', 'age', 'fare', 'embarked']].copy()
X_train, X_test, y_train, y_test = train_test_split(features, target, random_state=0)

numerical_features = features.dtypes == 'float'
categorical_features = ~numerical_features

preprocess = make_column_transformer(
    (make_pipeline(SimpleImputer(strategy='median'), StandardScaler()), numerical_features),
    (make_pipeline(SimpleImputer(strategy='constant', fill_value='missing'), OneHotEncoder(handle_unknown='ignore')), categorical_features))

model = make_pipeline(
    preprocess,
    LogisticRegression(solver='lbfgs'))

model.fit(X_train, y_train)
print("logistic regression score: %f" % model.score(X_test, y_test))
joblib.dump(model, 'model.joblib')

I am saving the trained model and then using that to create a Docker image to be deployed using s2i based on the model template app (link). I am able to successfully build the image and have it run locally.

Now I am trying to test by sending in a curl request to a predict endpoint. But I am not sure what the input format should be like. An example input to the model is an array like so:

array([[1, 'female', 29.0, 211.3375, 'S']], dtype=object)

All examples seemed to point to the use case were all input array was all numeric. I did not find an example where we can send this mixed object type.

Can you please let me know if this is possible in seldon-core or should I always assume that the input has to be numeric. If so, is there a recommended way of deploying this pipeline.

[ukclivecox]

You can send mix typed data with the ndarray payload:

seldon-core/proto/prediction.proto

Line 29 in a4013f6

google.protobuf.ListValue ndarray = 3;

A JSON payload for this would be for example

{"data":{"ndarray":["a",1,"b"]}}

[nsriram13]

Thanks for the speedy response. Based on above, I tried to issue the following curl command

curl -g http://localhost:5000/predict --data-urlencode 'json={"data": {"ndarray": [[1, "female", 29, 211.33, "S"]]}}'

and I get the following error:

ValueError: SimpleImputer does not support data with dtype <U21. Please provide either a numeric array (with a floating point or integer dtype) or categorical data represented either as an array with integer dtype or an array of string values with an object dtype.

A more detailed log can be seen here. Any ideas on how to troubleshoot?

[ukclivecox]

When converting to a numpy array automatically its type is <U21 (unicode 21 chars).
So you maybe need to convert it yourself into dtype=object I assume before passing to the sklearn pipeline.

[nsriram13]

When converting to a numpy array automatically its type is <U21 (unicode 21 chars).
So you maybe need to convert it yourself into dtype=object I assume before passing to the sklearn pipeline.

Thanks for the tip @cliveseldon . I have this working now. Basically had to modify the predict function to cast X as an object dtype as shown below:

class TitanicServe(object):

    def __init__(self, model_file='model.joblib'):
        self.model = joblib.load(model_file)

    def predict(self,X,features_names):
        X = X.astype(object)
        return self.model.predict_proba(X)

PS: thanks for giving the community this awesome toolkit for model deployment! 👍

[ukclivecox]

Glad its working. Keep us updated on how you're using Seldon Core.
I'll close this now. Feel free to open future issues.