# Ensemble Neural Network Model Weights in Keras (Polyak Averaging)

Last Updated on August 28, 2020

The training process of neural networks is a challenging optimization process that can often fail to converge.

This can mean that the model at the end of training may not be a stable or best-performing set of weights to use as a final model.

One approach to address this problem is to use an average of the weights from multiple models seen toward the end of the training run. This is called Polyak-Ruppert averaging and can be further improved by using a linearly or exponentially decreasing weighted average of the model weights. In addition to resulting in a more stable model, the performance of the averaged model weights can also result in better performance.

In this tutorial, you will discover how to combine the weights from multiple different models into a single model for making predictions.

After completing this tutorial, you will know:

• The stochastic and challenging nature of training neural networks can mean that the optimization process does not converge.
• Creating a model with the average of the weights from models observed towards the end of a training run can result in a more stable and sometimes better-performing solution.
• How to develop final models created with the equal, linearly, and exponentially weighted average of model parameters from multiple saved models.

Kick-start your project with my new book Better Deep Learning, including step-by-step tutorials and the Python source code files for all examples.

Let’s get started.

• Updated Oct/2019: Updated for Keras 2.3 and TensorFlow 2.0.
• Update Jan/2020: Updated for changes in scikit-learn v0.22 API.

How to Create an Equally, Linearly, and Exponentially Weighted Average of Neural Network Model Weights in Keras
Photo by netselesoobrazno, some rights reserved.

## Tutorial Overview

This tutorial is divided into seven parts; they are:

1. Average Model Weight Ensemble
2. Multi-Class Classification Problem
3. Multilayer Perceptron Model
4. Save Multiple Models to File
5. New Model With Average Model Weights
6. Predicting With an Average Model Weight Ensemble
7. Linearly and Exponentially Decreasing Weighted Average

## Average Model Weight Ensemble

Learning the weights for a deep neural network model requires solving a high-dimensional non-convex optimization problem.

A challenge with solving this optimization problem is that there are many “good” solutions and it is possible for the learning algorithm to bounce around and fail to settle in on one. In the area of stochastic optimization, this is referred to as problems with the convergence of the optimization algorithm on a solution, where a solution is defined by a set of specific weight values.

A symptom you may see if you have a problem with the convergence of your model is train and/or test loss value that shows higher than expected variance, e.g. it thrashes or bounces up and down over training epochs.

One approach to address this problem is to combine the weights collected towards the end of the training process. Generally, this might be referred to as temporal averaging and is known as Polyak Averaging or Polyak-Ruppert averaging, named for the original developers of the method.

Polyak averaging consists of averaging together several points in the trajectory through parameter space visited by an optimization algorithm.

— Page 322, Deep Learning, 2016.

Averaging multiple noisy sets of weights during the learning process may paradoxically sound less desirable than tuning the optimization process itself, but may prove a desirable solution, especially for very large neural networks that may take days, weeks, or even months to train.

The essential advancement was reached on the basis of the paradoxical idea: a slow algorithm having less than optimal convergence rate must be averaged.

Averaging the weights of multiple models from a single training run has the effect of calming down the noisy optimization process that may be noisy because of the choice of learning hyperparameters (e.g. learning rate) or the shape of the mapping function that is being learned. The result is a final model or set of weights that may offer a more stable, and perhaps more accurate result.

The basic idea is that the optimization algorithm may leap back and forth across a valley several times without ever visiting a point near the bottom of the valley. The average of all of the locations on either side should be close to the bottom of the valley though.

— Page 322, Deep Learning, 2016.

The simplest implementation of Polyak-Ruppert averaging involves calculating the average of the weights of the models over the last few training epochs.

This can be improved by calculating a weighted average, where more weight is applied to more recent models, which is linearly decreased through prior epochs. An alternative and more widely used approach is to use an exponential decay in the weighted average.

Polyak-Ruppert averaging has been shown to improve the convergence of standard SGD […] . Alternatively, an exponential moving average over the parameters can be used, giving higher weight to more recent parameter value.

Using an average or weighted average of model weights in the final model is a common technique in practice for ensuring the very best results are achieved from the training run. The approach is one of many “tricks” used in the Google Inception V2 and V3 deep convolutional neural network models for photo classification, a milestone in the field of deep learning.

Model evaluations are performed using a running average of the parameters computed over time.

### Want Better Results with Deep Learning?

Take my free 7-day email crash course now (with sample code).

Click to sign-up and also get a free PDF Ebook version of the course.

## Multi-Class Classification Problem

We will use a small multi-class classification problem as the basis to demonstrate the model weight ensemble.

The scikit-learn class provides the make_blobs() function that can be used to create a multi-class classification problem with the prescribed number of samples, input variables, classes, and variance of samples within a class.

The problem has two input variables (to represent the x and y coordinates of the points) and a standard deviation of 2.0 for points within each group. We will use the same random state (seed for the pseudorandom number generator) to ensure that we always get the same data points.

The results are the input and output elements of a dataset that we can model.

In order to get a feeling for the complexity of the problem, we can plot each point on a two-dimensional scatter plot and color each point by class value.

The complete example is listed below.

Running the example creates a scatter plot of the entire dataset. We can see that the standard deviation of 2.0 means that the classes are not linearly separable (separable by a line) causing many ambiguous points.

This is desirable as it means that the problem is non-trivial and will allow a neural network model to find many different “good enough” candidate solutions resulting in a high variance.

Scatter Plot of Blobs Dataset With Three Classes and Points Colored by Class Value

## Multilayer Perceptron Model

Before we define a model, we need to contrive a problem that is appropriate for the ensemble.

In our problem, the training dataset is relatively small. Specifically, there is a 10:1 ratio of examples in the training dataset to the holdout dataset. This mimics a situation where we may have a vast number of unlabeled examples and a small number of labeled examples with which to train a model.

We will create 1,100 data points from the blobs problem. The model will be trained on the first 100 points and the remaining 1,000 will be held back in a test dataset, unavailable to the model.

The problem is a multi-class classification problem, and we will model it using a softmax activation function on the output layer. This means that the model will predict a vector with three elements with the probability that the sample belongs to each of the three classes. Therefore, we must one hot encode the class values before we split the rows into the train and test datasets. We can do this using the Keras to_categorical() function.

Next, we can define and compile the model.

The model will expect samples with two input variables. The model then has a single hidden layer with 25 nodes and a rectified linear activation function, then an output layer with three nodes to predict the probability of each of the three classes and a softmax activation function.

Because the problem is multi-class, we will use the categorical cross entropy loss function to optimize the model and stochastic gradient descent with a small learning rate and momentum.

The model is fit for 500 training epochs and we will evaluate the model each epoch on the test set, using the test set as a validation set.

At the end of the run, we will evaluate the performance of the model on the train and test sets.

Then finally, we will plot learning curves of the model accuracy over each training epoch on both the training and validation datasets.

Tying all of this together, the complete example is listed below.

Running the example prints the performance of the final model on the train and test datasets.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

In this case, we can see that the model achieved about 86% accuracy on the training dataset, which we know is optimistic, and about 81% on the test dataset, which we would expect to be more realistic.

A line plot is also created showing the learning curves for the model accuracy on the train and test sets over each training epoch.

We can see that training accuracy is more optimistic over most of the run, as we also noted with the final scores. Importantly, we do see a reasonable amount of variance in the accuracy during training on both the train and test datasets, potentially providing a good basis for using model weight averaging.

Line Plot Learning Curves of Model Accuracy on Train and Test Dataset over Each Training Epoch

## Save Multiple Models to File

One approach to the model weight ensemble is to keep a running average of model weights in memory.

There are three downsides to this approach:

• It requires that you know beforehand the way in which the model weights will be combined; perhaps you want to experiment with different approaches.
• It requires that you know the number of epochs to use for training; maybe you want to use early stopping.
• It requires that you keep at least one copy of the entire network in memory; this could be very expensive for large models and fragile if the training process crashes or is killed.

An alternative is to save model weights to file during training as a first step, and later combine the weights from the saved models in order to make a final model.

Perhaps the simplest way to implement this is to manually drive the training process, one epoch at a time, then save models at the end of the epoch if we have exceeded an upper limit on the number of epochs.

For example, with our test problem, we will train the model for 500 epochs and perhaps save models from epoch 490 onwards (e.g. between and including epochs 490 and 499).

Models can be saved to file using the save() function on the model and specifying a filename that includes the epoch number.

Note, saving and loading neural network models in Keras requires that you have the h5py library installed. You can install this library using pip as follows:

Tying all of this together, the complete example of fitting the model on the training dataset and saving all models from the last 10 epochs is listed below.

Running the example saves 10 models into the current working directory.

## New Model With Average Models Weights

We can create a new model from multiple existing models with the same architecture.

First, we need to load the models into memory. This is reasonable as the models are small. If you are working with very large models, it might be easier to load models one at a time and average the weights in memory.

The load_model() Keras function can be used to load a saved model from file. The function load_all_models() below will load models from the current working directory. It takes the start and end epochs as arguments so that you can experiment with different groups of models saved over contiguous epochs.

We can call the function to load all of the models.

Once loaded, we can create a new model with the weighted average of the model weights.

Each model has a get_weights() function that returns a list of arrays, one for each layer in the model. We can enumerate each layer in the model, retrieve the same layer from each model, and calculate the weighted average. This will give us a set of weights.

We can then use the clone_model() Keras function to create a clone of the architecture and call set_weights() function to use the average weights we have prepared. The model_weight_ensemble() function below implements this.

Tying these elements together, we can load the 10 models and calculate the equally weighted average (arithmetic average) of the model weights. The complete listing is provided below.

Running the example first loads the 10 models from file.

A model weight ensemble is created from these 10 models giving equal weight to each model and a summary of the model structure is reported.

## Predicting With an Average Model Weight Ensemble

Now that we know how to calculate a weighted average of model weights, we can evaluate predictions with the resulting model.

One issue is that we don’t know how many models are appropriate to combine in order to achieve good performance. We can address this by evaluating model weight averaging ensembles with the last n models and vary n to see how many models results in good performance.

The evaluate_n_members() function below will create a new model from a given number of loaded models. Each model is given an equal weight in contributing to the final model, then the model_weight_ensemble() function is called to create the final model that is then evaluated on the test dataset.

Importantly, the list of loaded models is reversed first to ensure that the last n models in the training run are used, which we would assume might have better performance on average.

We can then evaluate models created from different numbers of the last n models saved from the training run from the last 1-model to the last 10 models. In addition to evaluating the combined final model, we can also evaluate each saved standalone model on the test dataset to compare performance.

The collected scores can be plotted, with blue dots for the accuracy of the single saved models and the orange line for the test accuracy for the model that combines the weights the last n models.

Tying all of this together, the complete example is listed below.

Running the example first loads the 10 saved models.

The performance of each individually saved model is reported as well as an ensemble model with weights averaged from all models up to and including each model, working backward from the end of the training run.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

The results show that the best test accuracy was about 81.4% achieved by the last two models. We can see that the test accuracy of the model weight ensemble levels out the performance and performs just as well.

A line plot is also created showing the test accuracy of each single model (blue dots) and the performance of the model weight ensemble (orange line).

We can see that averaging the model weights does level out the performance of the final model and performs at least as well as the final model of the run.

Line Plot of Single Model Test Performance (blue dots) and Model Weight Ensemble Test Performance (orange line)

## Linearly and Exponentially Decreasing Weighted Average

We can update the example and evaluate a linearly decreasing weighting of the model weights in the ensemble.

The weights can be calculated as follows:

This can be used instead of the equal weights in the evaluate_n_members() function.

The complete example is listed below.

Running the example reports the performance of each single model again, and this time the test accuracy of each average model weight ensemble with a linearly decreasing contribution of models.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

We can see that, at least in this case, the ensemble achieves a small bump in performance above any standalone model to about 81.5% accuracy.

The line plot shows the bump in performance and shows a more stable performance in terms of test accuracy over the different sized ensembles created, as compared to the use of an evenly weighted ensemble.

Line Plot of Single Model Test Performance (blue dots) and Model Weight Ensemble Test Performance (orange line) With a Linear Decay

We can also experiment with an exponential decay of the contribution of models. This requires that a decay rate (alpha) is specified. The example below creates weights for an exponential decay with a decrease rate of 2.

The complete example with an exponential decay for the contribution of models to the average weights in the ensemble model is listed below.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

Running the example shows a small improvement in performance much like the use of a linear decay in the weighted average of the saved models.

The line plot of the test accuracy scores shows the stronger stabilizing effect of using the exponential decay instead of the linear or equal weighting of models.

Line Plot of Single Model Test Performance (blue dots) and Model Weight Ensemble Test Performance (orange line) With an Exponential Decay

## Extensions

This section lists some ideas for extending the tutorial that you may wish to explore.

• Number of Models. Evaluate the effect of many more models contributing their weights to the final model.
• Decay Rate. Evaluate the effect on test performance of using different decay rates for an exponentially weighted average.

If you explore any of these extensions, I’d love to know.

This section provides more resources on the topic if you are looking to go deeper.

## Summary

In this tutorial, you discovered how to combine the weights from multiple different models into a single model for making predictions.

Specifically, you learned:

• The stochastic and challenging nature of training neural networks can mean that the optimization process does not converge.
• Creating a model with the average of the weights from models observed towards the end of a training run can result in a more stable and sometimes better-performing solution.
• How to develop final models created with the equal, linearly, and exponentially weighted average of model parameters from multiple saved models.

Do you have any questions?

## Develop Better Deep Learning Models Today!

#### Train Faster, Reduce Overftting, and Ensembles

...with just a few lines of python code

Discover how in my new Ebook:
Better Deep Learning

It provides self-study tutorials on topics like:
weight decay, batch normalization, dropout, model stacking and much more...

### 18 Responses to Ensemble Neural Network Model Weights in Keras (Polyak Averaging)

1. Shane January 11, 2019 at 8:38 am #

Thanks for the article Jason!
Just wanted to understand how is this averaging of weights more effective than having an adaptive learning rate in optimizer? In effect, the two should be equivalent.

• Jason Brownlee January 12, 2019 at 5:34 am #

Do you mean one model with adaptive learning rate vs an ensemble of models?

The ensemble will still reduce the variance and may have the benefit of lifting skill over a single solution. It really depends on the complexity of the problem.

2. Annie January 15, 2019 at 6:14 pm #

Hi, Jason,

What’s the difference or which one is more efficient? averaging the weights to one model or averaging the output/score of multiple models?

Additionally, if I train several independent models, should I still average weights? Or should I ensemble the models?

Thanks!

• Jason Brownlee January 16, 2019 at 5:44 am #

Perhaps try each on your problem and use the approach that works the best.

thank you sir,

i want to understand about ensemble and overfitting so can you please say something?

• Jason Brownlee January 19, 2019 at 5:36 am #

Ensembles can make overfitting more challenging, not less so.

4. Irina September 7, 2020 at 7:38 pm #

I was trying to use the function model_weight_ensemble, and I have got a following error:
“Weights for model sequential have not yet been created. Weights are created when the Model is first called on inputs or build() is called with an input_shape.” :-/

5. Irina September 7, 2020 at 7:40 pm #

I wanted to say thank you very much for this article. Even if I still get an error in my code, the idea is really cool!

• Jason Brownlee September 8, 2020 at 6:49 am #

Thanks!

6. Irina Kärkkänen September 9, 2020 at 9:09 pm #

Now it worked, I did not give the input dimensions to my model. My mistake :). Thank again!

7. Isaac M December 15, 2020 at 2:20 pm #

It is an interesting overview. However, in the end, the ensemble results are way worse than the single model results. I run the model to converge then save the resulting model and repeat this couple of times, then average the weight of converging models not the ones from individual epochs. Does that create any problems? I am also initializing the biases but I guess that is not the problem.

• Jason Brownlee December 16, 2020 at 7:43 am #

If your approach gives you good results on your dataset, and knowledge of test data is not leaking into training, then go for it!

8. Minnie March 28, 2021 at 3:51 am #

Hi Jason.

I am finding output of Batch Normalisation on Keras.
have posted my ques on Stackoverflow:

https://stackoverflow.com/questions/66832395/is-output-of-batch-normalization-in-keras-dependent-on-number-of-epochs

Can you check it?

9. Richa May 30, 2021 at 8:50 pm #

Hi Jason

Thanks for sharing the knowledge on ensemble modeling. Would you please share the link or explain how to clone a subclass model (custom model).

• Jason Brownlee May 31, 2021 at 5:48 am #

Sorry, I don’t have a tutorial on that topic.