def predict_online(data, config=None):
"""Predict from in-memory data on the fly.
"""
if config is None:
logger.debug(
"Config path was not explicitly passed. Falling back to default config."
)
config = load_yaml(DEFAULT_CONFIG_PATH)
config = config["predict"]
log_config_path = config["logging"]["config_path"]
initialize_logging(config_path=log_config_path)
model_dirname = config["model"]["dirname"]
model_version = config["model"]["version"]
MODEL_EXT = "keras" # joblib
# model_path = Path(model_dirname) / f"v{model_version}.{MODEL_EXT}"
model_path = os.path.join(model_dirname, f"v{model_version}.{MODEL_EXT}")
try:
# @todo: fix the hard coding
# checkpoint = load_sklearn_model(model_path)
checkpoint = load_keras_hub_model(model_path)
pred = checkpoint.predict(data)
# can't jsonify np array
pred = pred.tolist()
logger.info({"input": data, "pred": pred})
except Exception as e:
logger.error(f"{e}")
pred = []
return pred
def main(exp_config_path):
"""
Simple
"""
# read off a config file that controls experiment parameters.
config = load_yaml(exp_config_path)
config = config["train"]
# determine an output path where results of an experiment are stored.
cur_time = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
exp_output_dirname = Path(config["experiment"]["output_dirname"]) / cur_time
config["experiment"]["output_dirname"] = exp_output_dirname
if not exp_output_dirname.exists():
Path.mkdir(exp_output_dirname, exist_ok=True)
# initialize loggers
# @note: log files are stored to each exp_dirname.
# this is more useful for machine learning pipelines.
# as opposed to a standard web app where it has centralized log.
# for other modules other than train.py we log to the centralized.
log_config_path = config["logging"]["config_path"]
initialize_logging(config_path=log_config_path, log_dirname=exp_output_dirname)
# a demo of a training pipeline using sklearn Pipeline
# @todo: add more pipelines: e.g. tensorflow, pytorch
if config["data"]["dataset_name"] == "iris":
run_sklearn_pipeline(config)
else:
raise ValueError(f"Unsupported dataset was given {config['data']}.")
def predict(exp_config_path):
"""Make predictions from data.
data (np.array): n x d array.
Returns:
pred (list): n - dimensional predictions.
"""
raise NotImplementedError
# read off a config file that controls experiment parameters.
config = load_yaml(exp_config_path)
config = config["predict"]
log_config_path = config["logging"]["config_path"]
initialize_logging(config_path=log_config_path)
# set data to evaluate on
# @todo: not implemented yet.
val_data = config["dataset_path"]
pred = predict_online(val_data, config)
return pred
import logging
from flask import Blueprint, request, jsonify
from ml_deploy_demo.predict import predict_online
from ml_deploy_demo.util.utils import initialize_logging
logger = logging.getLogger(__name__)
initialize_logging(config_path='/app/logging.yaml')
ml_app = Blueprint("ml_app", __name__)
@ml_app.route("/predict", methods=["GET", "POST"])
def predict():
"""Performs an inference
"""
if request.method == "POST":
data = request.get_json()
logger.debug(f"Input to predict/: {data}")
pred = predict_online(data=data["data"])
return jsonify({"input": data, "pred": pred})
if request.method == "GET":
msg = f"Please compose your request in POST type with data."
logger.debug(f"Wrong request type {request}.")
return jsonify({"msg": msg})
import numpy as np
import pandas as pd
import tensorflow as tf
import tensorflow_hub as hub
import tensorflow_datasets as tfds
from ml_deploy_demo.util.utils import load_yaml, initialize_logging
DEFAULT_CONFIG_PATH = "/app/experiment_configs/default.yaml"
logger = logging.getLogger(__name__)
config = load_yaml(DEFAULT_CONFIG_PATH)
config = config["train"]
log_config_path = config["logging"]["config_path"]
initialize_logging(config_path=log_config_path)
# This example is taken from:
# https://www.tensorflow.org/tutorials/keras/text_classification_with_hub
def run_pipeline():
""" runs pipeline to train keras DNN model
for sentiment classification """
# Split the training set into 60% and 40%, so we'll end up with 15,000 examples
# for training, 10,000 examples for validation and 25,000 examples for testing.
train_validation_split = tfds.Split.TRAIN.subsplit([6, 4])
(train_data, validation_data), test_data = tfds.load(
name="imdb_reviews",
split=(train_validation_split, tfds.Split.TEST),