Keras + Hyperopt: A very simple wrapper for convenient hyperparameter optimization

Overview

This project is now archived. It's been fun working on it, but it's time for me to move on. Thank you for all the support and feedback over the last couple of years. If someone is interested in taking ownership, let's discuss. ✌️

Hyperas Build Status PyPI version

A very simple convenience wrapper around hyperopt for fast prototyping with keras models. Hyperas lets you use the power of hyperopt without having to learn the syntax of it. Instead, just define your keras model as you are used to, but use a simple template notation to define hyper-parameter ranges to tune.

Installation

pip install hyperas

Quick start

Assume you have data generated as such

def data():
    x_train = np.zeros(100)
    x_test = np.zeros(100)
    y_train = np.zeros(100)
    y_test = np.zeros(100)
    return x_train, y_train, x_test, y_test

and an existing keras model like the following

def create_model(x_train, y_train, x_test, y_test):
    model = Sequential()
    model.add(Dense(512, input_shape=(784,)))
    model.add(Activation('relu'))
    model.add(Dropout(0.2))
    model.add(Dense(512))
    model.add(Activation('relu'))
    model.add(Dropout(0.2))
    model.add(Dense(10))
    model.add(Activation('softmax'))

    # ... model fitting

    return model

To do hyper-parameter optimization on this model, just wrap the parameters you want to optimize into double curly brackets and choose a distribution over which to run the algorithm.

In the above example, let's say we want to optimize for the best dropout probability in both dropout layers. Choosing a uniform distribution over the interval [0,1], this translates into the following definition. Note that before returning the model, to optimize, we also have to define which evaluation metric of the model is important to us. For example, in the following, we optimize for accuracy.

Note: In the following code we use 'loss': -accuracy, i.e. the negative of accuracy. That's because under the hood hyperopt will always minimize whatever metric you provide. If instead you want to actually want to minimize a metric, say MSE or another loss function, you keep a positive sign (e.g. 'loss': mse).

from hyperas.distributions import uniform

def create_model(x_train, y_train, x_test, y_test):
    model = Sequential()
    model.add(Dense(512, input_shape=(784,)))
    model.add(Activation('relu'))
    model.add(Dropout({{uniform(0, 1)}}))
    model.add(Dense(512))
    model.add(Activation('relu'))
    model.add(Dropout({{uniform(0, 1)}}))
    model.add(Dense(10))
    model.add(Activation('softmax'))

    # ... model fitting

    score = model.evaluate(x_test, y_test, verbose=0)
    accuracy = score[1]
    return {'loss': -accuracy, 'status': STATUS_OK, 'model': model}

The last step is to actually run the optimization, which is done as follows:

best_run = optim.minimize(model=create_model,
                          data=data,
                          algo=tpe.suggest,
                          max_evals=10,
                          trials=Trials())

In this example we use at most 10 evaluation runs and the TPE algorithm from hyperopt for optimization.

Check the "complete example" below for more details.

Complete example

Note: It is important to wrap your data and model into functions as shown below, and then pass them as parameters to the minimizer. data() returns the data the create_model() needs. An extended version of the above example in one script reads as follows. This example shows many potential use cases of hyperas, including:

  • Varying dropout probabilities, sampling from a uniform distribution
  • Different layer output sizes
  • Different optimization algorithms to use
  • Varying choices of activation functions
  • Conditionally adding layers depending on a choice
  • Swapping whole sets of layers
from __future__ import print_function
import numpy as np

from hyperopt import Trials, STATUS_OK, tpe
from keras.datasets import mnist
from keras.layers.core import Dense, Dropout, Activation
from keras.models import Sequential
from keras.utils import np_utils

from hyperas import optim
from hyperas.distributions import choice, uniform


def data():
    """
    Data providing function:

    This function is separated from create_model() so that hyperopt
    won't reload data for each evaluation run.
    """
    (x_train, y_train), (x_test, y_test) = mnist.load_data()
    x_train = x_train.reshape(60000, 784)
    x_test = x_test.reshape(10000, 784)
    x_train = x_train.astype('float32')
    x_test = x_test.astype('float32')
    x_train /= 255
    x_test /= 255
    nb_classes = 10
    y_train = np_utils.to_categorical(y_train, nb_classes)
    y_test = np_utils.to_categorical(y_test, nb_classes)
    return x_train, y_train, x_test, y_test


def create_model(x_train, y_train, x_test, y_test):
    """
    Model providing function:

    Create Keras model with double curly brackets dropped-in as needed.
    Return value has to be a valid python dictionary with two customary keys:
        - loss: Specify a numeric evaluation metric to be minimized
        - status: Just use STATUS_OK and see hyperopt documentation if not feasible
    The last one is optional, though recommended, namely:
        - model: specify the model just created so that we can later use it again.
    """
    model = Sequential()
    model.add(Dense(512, input_shape=(784,)))
    model.add(Activation('relu'))
    model.add(Dropout({{uniform(0, 1)}}))
    model.add(Dense({{choice([256, 512, 1024])}}))
    model.add(Activation({{choice(['relu', 'sigmoid'])}}))
    model.add(Dropout({{uniform(0, 1)}}))

    # If we choose 'four', add an additional fourth layer
    if {{choice(['three', 'four'])}} == 'four':
        model.add(Dense(100))

        # We can also choose between complete sets of layers

        model.add({{choice([Dropout(0.5), Activation('linear')])}})
        model.add(Activation('relu'))

    model.add(Dense(10))
    model.add(Activation('softmax'))

    model.compile(loss='categorical_crossentropy', metrics=['accuracy'],
                  optimizer={{choice(['rmsprop', 'adam', 'sgd'])}})

    result = model.fit(x_train, y_train,
              batch_size={{choice([64, 128])}},
              epochs=2,
              verbose=2,
              validation_split=0.1)
    #get the highest validation accuracy of the training epochs
    validation_acc = np.amax(result.history['val_acc']) 
    print('Best validation acc of epoch:', validation_acc)
    return {'loss': -validation_acc, 'status': STATUS_OK, 'model': model}


if __name__ == '__main__':
    best_run, best_model = optim.minimize(model=create_model,
                                          data=data,
                                          algo=tpe.suggest,
                                          max_evals=5,
                                          trials=Trials())
    X_train, Y_train, X_test, Y_test = data()
    print("Evalutation of best performing model:")
    print(best_model.evaluate(X_test, Y_test))
    print("Best performing model chosen hyper-parameters:")
    print(best_run)

FAQ

Here is a list of a few popular errors

TypeError: require string label

You're probably trying to execute the model creation code, with the templates, directly in python. That fails simply because python cannot run the templating in the braces, e.g. {{uniform..}}. The def create_model(...) function is in fact not a valid python function anymore.

You need to wrap your code in a def create_model(...): ... function, and then call it from optim.minimize(model=create_model,... like in the example.

The reason for this is that hyperas works by doing template replacement of everything in the {{...}} into a separate temporary file, and then running the model with the replaced braces (think jinja templating).

This is the basis of how hyperas simplifies usage of hyperopt by being a "very simple wrapper".

TypeError: 'generator' object is not subscriptable

This is currently a known issue.

Just pip install networkx==1.11

NameError: global name 'X_train' is not defined

Maybe you forgot to return the x_train argument in the def create_model(x_train...) call from the def data(): ... function.

You are not restricted to the same list of arguments as in the example. Any arguments you return from data() will be passed to create_model()

notebook adjustment

If you find error like "No such file or directory" or OSError, Err22, you may need add notebook_name='simple_notebook'(assume your current notebook name is simple_notebook) in optim.minimize function like this:

best_run, best_model = optim.minimize(model=model,
                                      data=data,
                                      algo=tpe.suggest,
                                      max_evals=5,
                                      trials=Trials(),
                                      notebook_name='simple_notebook')

How does hyperas work?

All we do is parse the data and model templates and translate them into proper hyperopt by reconstructing the space object that's then passed to fmin. Most of the relevant code is found in optim.py and utils.py.

How to read the output of a hyperas model?

Hyperas translates your script into hyperopt compliant code, see here for some guidance on how to interpret the result.

How to pass arguments to data?

Suppose you want your data function take an argument, specify it like this using positional arguments only (not keyword arguments):

import pickle
def data(fname):
    with open(fname,'rb') as fh:
        return pickle.load(fh)

Note that your arguments must be implemented such that repr can show them in their entirety (such as strings and numbers). If you want more complex objects, use the passed arguments to build them inside the data function.

And when you run your trials, pass a tuple of arguments to be substituted in as data_args:

best_run, best_model = optim.minimize(
    model=model,
    data=data,
    algo=tpe.suggest,
    max_evals=64,
    trials=Trials(),
    data_args=('my_file.pkl',)
)

What if I need more flexibility loading data and adapting my model?

Hyperas is a convenience wrapper around Hyperopt that has some limitations. If it's not convenient to use in your situation, simply don't use it -- and choose Hyperopt instead. All you can do with Hyperas you can also do with Hyperopt, it's just a different way of defining your model. If you want to squeeze some flexibility out of Hyperas anyway, take a look here.

Running hyperas in parallel?

You can use hyperas to run multiple models in parallel with the use of mongodb (which you'll need to install and setup users for). Here's a short example using MNIST:

  1. Copy and modify examples/mnist_distributed.py (bump up max_evals if you like):

  2. Run python mnist_distributed.py. It will create a temp_model.py file. Copy this file to any machines that will be evaluating models. It will then begin waiting for evaluation results

  3. On your other machines (make sure they have a python installed with all your dependencies, ideally with the same versions) run:

    export PYTHONPATH=/path/to/temp_model.py
    hyperopt-mongo-worker --exp-key='mnist_test' --mongo='mongo://username:[email protected]:27017/jobs'
  4. Once max_evals have been completed, you should get an output with your best model. You can also look through your mongodb and examine the results, to get the best model out and run it, do:

    from pymongo import MongoClient
    from keras.models import load_model
    import tempfile
    c = MongoClient('mongodb://username:[email protected]:27017/jobs')
    best_model = c['jobs']['jobs'].find_one({'exp_key': 'mnist_test'}, sort=[('result.loss', -1)])
    temp_name = tempfile.gettempdir()+'/'+next(tempfile._get_candidate_names()) + '.h5'
    with open(temp_name, 'wb') as outfile:
        outfile.write(best_model['result']['model_serial'])
    model = load_model(temp_name)
Owner
Max Pumperla
Data Science Professor, Data Scientist & Engineer. DL4J core developer, Hyperopt maintainer, Keras contributor. Author of "Deep Learning and the Game of Go"
Max Pumperla
A light weight data augmentation tool for training CNNs and Viola Jones detectors

hey-daug A light weight data augmentation tool for training CNNs and Viola Jones detectors (Haar Cascades). This tool inflates your data by up to six

Jaiyam Sharma 2 Nov 23, 2019
This repository contains the source code and data for reproducing results of Deep Continuous Clustering paper

Deep Continuous Clustering Introduction This is a Pytorch implementation of the DCC algorithms presented in the following paper (paper): Sohil Atul Sh

Sohil Shah 197 Nov 29, 2022
OpenL3: Open-source deep audio and image embeddings

OpenL3 OpenL3 is an open-source Python library for computing deep audio and image embeddings. Please refer to the documentation for detailed instructi

Music and Audio Research Laboratory - NYU 326 Jan 02, 2023
RM Operation can equivalently convert ResNet to VGG, which is better for pruning; and can help RepVGG perform better when the depth is large.

RM Operation can equivalently convert ResNet to VGG, which is better for pruning; and can help RepVGG perform better when the depth is large.

184 Jan 04, 2023
This is the 3D Implementation of 《Inconsistency-aware Uncertainty Estimation for Semi-supervised Medical Image Segmentation》

CoraNet This is the 3D Implementation of 《Inconsistency-aware Uncertainty Estimation for Semi-supervised Medical Image Segmentation》 Environment pytor

25 Nov 08, 2022
Video Swin Transformer - PyTorch

Video-Swin-Transformer-Pytorch This repo is a simple usage of the official implementation "Video Swin Transformer". Introduction Video Swin Transforme

Haofan Wang 116 Dec 20, 2022
PyTorch Implementation of CvT: Introducing Convolutions to Vision Transformers

CvT: Introducing Convolutions to Vision Transformers Pytorch implementation of CvT: Introducing Convolutions to Vision Transformers Usage: img = torch

Rishikesh (ऋषिकेश) 193 Jan 03, 2023
[NeurIPS 2021] PyTorch Code for Accelerating Robotic Reinforcement Learning with Parameterized Action Primitives

Robot Action Primitives (RAPS) This repository is the official implementation of Accelerating Robotic Reinforcement Learning via Parameterized Action

Murtaza Dalal 55 Dec 27, 2022
Deep Markov Factor Analysis (NeurIPS2021)

Deep Markov Factor Analysis (DMFA) Codes and experiments for deep Markov factor analysis (DMFA) model accepted for publication at NeurIPS2021: A. Farn

Sarah Ostadabbas 2 Dec 16, 2022
Fully Convolutional DenseNet (A.K.A 100 layer tiramisu) for semantic segmentation of images implemented in TensorFlow.

FC-DenseNet-Tensorflow This is a re-implementation of the 100 layer tiramisu, technically a fully convolutional DenseNet, in TensorFlow (Tiramisu). Th

Hasnain Raza 121 Oct 12, 2022
Proximal Backpropagation - a neural network training algorithm that takes implicit instead of explicit gradient steps

Proximal Backpropagation Proximal Backpropagation (ProxProp) is a neural network training algorithm that takes implicit instead of explicit gradient s

Thomas Frerix 40 Dec 17, 2022
Nerf pl - NeRF (Neural Radiance Fields) and NeRF in the Wild using pytorch-lightning

nerf_pl Update: an improved NSFF implementation to handle dynamic scene is open! Update: NeRF-W (NeRF in the Wild) implementation is added to nerfw br

AI葵 1.8k Dec 30, 2022
Efficient 3D Backbone Network for Temporal Modeling

VoV3D is an efficient and effective 3D backbone network for temporal modeling implemented on top of PySlowFast. Diverse Temporal Aggregation and

102 Dec 06, 2022
MRI reconstruction (e.g., QSM) using deep learning methods

deepMRI: Deep learning methods for MRI Authors: Yang Gao, Hongfu Sun This repo is devloped based on Pytorch (1.8 or later) and matlab (R2019a or later

Hongfu Sun 17 Dec 18, 2022
Experiments on continual learning from a stream of pretrained models.

Ex-model CL Ex-model continual learning is a setting where a stream of experts (i.e. model's parameters) is available and a CL model learns from them

Antonio Carta 6 Dec 04, 2022
Code for "Universal inference meets random projections: a scalable test for log-concavity"

How to use this repository This repository contains code to replicate the results of "Universal inference meets random projections: a scalable test fo

Robin Dunn 0 Nov 21, 2021
Code for KHGT model, AAAI2021

KHGT Code for KHGT accepted by AAAI2021 Please unzip the data files in Datasets/ first. To run KHGT on Yelp data, use python labcode_yelp.py For Movi

32 Nov 29, 2022
Efficient Sharpness-aware Minimization for Improved Training of Neural Networks

Efficient Sharpness-aware Minimization for Improved Training of Neural Networks Code for “Efficient Sharpness-aware Minimization for Improved Training

Angusdu 32 Oct 18, 2022
Code from PropMix, accepted at BMVC'21

PropMix: Hard Sample Filtering and Proportional MixUp for Learning with Noisy Labels This repository is the official implementation of Hard Sample Fil

6 Dec 21, 2022
Disease Informed Neural Networks (DINNs) — neural networks capable of learning how diseases spread, forecasting their progression, and finding their unique parameters (e.g. death rate).

DINN We introduce Disease Informed Neural Networks (DINNs) — neural networks capable of learning how diseases spread, forecasting their progression, a

19 Dec 10, 2022