image-classification-service/scripts/keras_MnWE.py

import multiprocessing

import numpy as np
from tensorflow import keras
from tensorflow.keras import layers


def process(predict_fn_wrapper):
    # We observed memory doesn't get properly deallocated unless we do this:
    manager = multiprocessing.Manager()
    return_dict = manager.dict()
    p = multiprocessing.Process(
        target=predict_fn_wrapper,
        args=(return_dict,),
    )
    p.start()
    p.join()
    try:
        return dict(return_dict)["result"]
    except KeyError:
        pass


def make_model():
    inputs = keras.Input(shape=(784,))
    dense = layers.Dense(64, activation="relu")
    x = dense(inputs)
    outputs = layers.Dense(10)(x)
    model = keras.Model(inputs=inputs, outputs=outputs, name="mnist_model")
    model.compile(
        loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
        optimizer=keras.optimizers.RMSprop(),
        metrics=["accuracy"],
    )
    return model


def make_predict_fn():
    # Keras bug: doesn't work in outer scope
    model = make_model()

    def predict(*args):
        # model = make_model()
        return model.predict(np.random.random(size=(1, 784)))

    return predict


def make_predict_fn_wrapper(predict_fn):
    def predict_fn_wrapper(return_dict):
        return_dict["result"] = predict_fn()

    return predict_fn_wrapper


if __name__ == "__main__":
    predict_fn = make_predict_fn()
    print(process(make_predict_fn_wrapper(predict_fn)))