ansrivas / keras-rest-server

A very simple project to create a rest backend for serving a neural network model based on keras

Keras-rest-server: A simple rest implementation for loading and serving keras models

---

About:

This repository contains a very simple server implemented in flask which loads a a simple neural network model trained using Keras from its saved-weights and model.

In this example a very simple case of XOR is considered.

Getting started:

---

1. Install Anaconda:

https://docs.continuum.io/anaconda/install

2. Clone this repository

git clone https://github.com/ansrivas/keras-rest-server.git
cd keras-rest-server

3. Create a new environment ( Change python=2 or python=3) and activate it:

conda create --name keras-server -y python=2
source activate keras-server

4. Install all the dependencies:

conda env update -n keras-server  --file requirements.txt

5. To remove the environment run:

conda remove -n keras-server --all -y

Usage

---

Run to generate pickle files:

python createpickles.py

Run the server (defaults to http://localhost:7171)

python server.py

Send a post request to this server to test your model

python client.py

keras-rest-server's Issues

document the possible requests

This could become a very interesting project if combined with https://neurals-ro.github.io/keras-model-editor/
(https://github.com/neurals-ro/keras-model-editor )
Is the server able to save a model given in JSON?

This should grow into a full-remote for Keras:

• update the model
• ability to start batch learning
• monitor the learning, get email when done
• ability to download the weights

what part of this is already implemented?

treat this as a classification problem, not regression

since youre predicting 1/0 (ie this is a classification example) for xor, you should use binary_crossentropy for your loss function, not mse.

keras-rest-server/createpickles.py

Line 68 in 9aa819b

model.compile(loss='mse', optimizer=rms)

additionally, you should use softmax as your activation function, not linear. this will return a predicted class (eg 0, 1) and not a real continuous value.

keras-rest-server/createpickles.py

Line 62 in 9aa819b

model.add(Activation('linear'))

Loading weights

Hi, I have not had much time to look trough your project but I really want to know some things:

1. Does the model need to be reloaded into memory every time i call the endpoint? because with vgg16() that would mean loading 500mb into memory on every request.
2. Can you run multiple models on the same application/server? Lets say I have 2 different weight files can I call them both?

Thanks!