# Classification and Regression with H2O Deep Learning

## Introduction

This tutorial shows how a H2O Deep Learning model can be used to do supervised classification and regression. A great tutorial about Deep Learning is given by Quoc Le here and here. This tutorial covers usage of H2O from R. A python version of this tutorial will be available as well in a separate document. This file is available in plain R, R markdown and regular markdown formats, and the plots are available as PDF files. All documents are available on Github.

If run from plain R, execute R in the directory of this script. If run from RStudio, be sure to setwd() to the location of this script. h2o.init() starts H2O in R’s current working directory. h2o.importFile() looks for files from the perspective of where H2O was started.

More examples and explanations can be found in our H2O Deep Learning booklet and on our H2O Github Repository. The PDF slide deck can be found on Github.

### Start H2O

Start up a 1-node H2O server on your local machine, and allow it to use all CPU cores and up to 2GB of memory:

The h2o.deeplearning function fits H2O’s Deep Learning models from within R.
We can run the example from the man page using the example function, or run a longer demonstration from the h2o package using the demo function:

While H2O Deep Learning has many parameters, it was designed to be just as easy to use as the other supervised training methods in H2O. Early stopping, automatic data standardization and handling of categorical variables and missing values and adaptive learning rates (per weight) reduce the amount of parameters the user has to specify. Often, it’s just the number and sizes of hidden layers, the number of epochs and the activation function and maybe some regularization techniques.

### Let’s have some fun first: Decision Boundaries

We start with a small dataset representing red and black dots on a plane, arranged in the shape of two nested spirals. Then we task H2O’s machine learning methods to separate the red and black dots, i.e., recognize each spiral as such by assigning each point in the plane to one of the two spirals.

We visualize the nature of H2O Deep Learning (DL), H2O’s tree methods (GBM/DRF) and H2O’s generalized linear modeling (GLM) by plotting the decision boundary between the red and black spirals:

We build a few different models:

Let’s investigate some more Deep Learning models. First, we explore the evolution over training time (number of passes over the data), and we use checkpointing to continue training the same model:

You can see how the network learns the structure of the spirals with enough training time. We explore different network architectures next:

It is clear that different configurations can achieve similar performance, and that tuning will be required for optimal performance. Next, we compare between different activation functions, including one with 50% dropout regularization in the hidden layers:

Clearly, the dropout rate was too high or the number of epochs was too low for the last configuration, which often ends up performing the best on larger datasets where generalization is important.

## Cover Type Dataset

We important the full cover type dataset (581k rows, 13 columns, 10 numerical, 3 categorical).
We also split the data 3 ways: 60% for training, 20% for validation (hyper parameter tuning) and 20% for final testing.

Here’s a scalable way to do scatter plots via binning (works for categorical and numeric columns) to get more familiar with the dataset.

### First Run of H2O Deep Learning

Let’s run our first Deep Learning model on the covtype dataset.
We want to predict the Cover_Type column, a categorical feature with 7 levels, and the Deep Learning model will be tasked to perform (multi-class) classification. It uses the other 12 predictors of the dataset, of which 10 are numerical, and 2 are categorical with a total of 44 levels. We can expect the Deep Learning model to have 56 input neurons (after automatic one-hot encoding).

To keep it fast, we only run for one epoch (one pass over the training data).

Inspect the model in Flow for more information about model building etc. by issuing a cell with the content getModel "dl_model_first", and pressing Ctrl-Enter.

### Variable Importances

Variable importances for Neural Network models are notoriously difficult to compute, and there are many pitfalls. H2O Deep Learning has implemented the method of Gedeon, and returns relative variable importances in descending order of importance.

### Early Stopping

Now we run another, smaller network, and we let it stop automatically once the misclassification rate converges (specifically, if the moving average of length 2 does not improve by at least 1% for 2 consecutive scoring events). We also sample the validation set to 10,000 rows for faster scoring.

By default, H2O Deep Learning uses an adaptive learning rate (ADADELTA) for its stochastic gradient descent optimization. There are only two tuning parameters for this method: rho and epsilon, which balance the global and local search efficiencies. rho is the similarity to prior weight updates (similar to momentum), and epsilon is a parameter that prevents the optimization to get stuck in local optima. Defaults are rho=0.99 and epsilon=1e-8. For cases where convergence speed is very important, it might make sense to perform a few runs to optimize these two parameters (e.g., with rho in c(0.9,0.95,0.99,0.999) and epsilon in c(1e-10,1e-8,1e-6,1e-4)). Of course, as always with grid searches, caution has to be applied when extrapolating grid search results to a different parameter regime (e.g., for more epochs or different layer topologies or activation functions, etc.).

If adaptive_rate is disabled, several manual learning rate parameters become important: rate, rate_annealing, rate_decay, momentum_start, momentum_ramp, momentum_stable and nesterov_accelerated_gradient, the discussion of which we leave to H2O Deep Learning booklet.

### Tuning

With some tuning, it is possible to obtain less than 10% test set error rate in about one minute. Error rates of below 5% are possible with larger models. Note that deep tree methods can be more effective for this dataset than Deep Learning, as they directly partition the space into sectors, which seems to be needed here.

Let’s compare the training error with the validation and test set errors

To confirm that the reported confusion matrix on the validation set (here, the test set) was correct, we make a prediction on the test set and compare the confusion matrices explicitly:

### Hyper-parameter Tuning with Grid Search

Since there are a lot of parameters that can impact model accuracy, hyper-parameter tuning is especially important for Deep Learning:

For speed, we will only train on the first 10,000 rows of the training dataset:

The simplest hyperparameter search method is a brute-force scan of the full Cartesian product of all combinations specified by a grid search:

Let’s see which model had the lowest validation error:

### Random Hyper-Parameter Search

Often, hyper-parameter search for more than 4 parameters can be done more efficiently with random parameter search than with grid search. Basically, chances are good to find one of many good models in less time than performing an exhaustive grid search. We simply build up to max_models models with parameters drawn randomly from user-specified distributions (here, uniform). For this example, we use the adaptive learning rate and focus on tuning the network architecture and the regularization parameters. We also let the grid search stop automatically once the performance at the top of the leaderboard doesn’t change much anymore, i.e., once the search has converged.

Let’s look at the model with the lowest validation misclassification rate:

### Checkpointing

Let’s continue training the manually tuned model from before, for 2 more epochs. Note that since many important parameters such as epochs, l1, l2, max_w2, score_interval, train_samples_per_iteration, input_dropout_ratio, hidden_dropout_ratios, score_duty_cycle, classification_stop, regression_stop, variable_importances, force_load_balance can be modified between checkpoint restarts, it is best to specify as many parameters as possible explicitly.

Once we are satisfied with the results, we can save the model to disk (on the cluster). In this example, we store the model in a directory called mybest_deeplearning_covtype_model, which will be created for us since force=TRUE.

It can be loaded later with the following command:

N-fold cross-validation is especially useful with early stopping, as the main model will pick the ideal number of epochs from the convergence behavior of the cross-validation models.

## Regression and Binary Classification

Assume we want to turn the multi-class problem above into a binary classification problem. We create a binary response as follows:

Let’s build a quick model and inspect the model:

Instead of a binary classification model, we find a regression model (H2ORegressionModel) that contains only 1 output neuron (instead of 2). The reason is that the response was a numerical feature (ordinal numbers 0 and 1), and H2O Deep Learning was run with distribution=AUTO, which defaulted to a Gaussian regression problem for a real-valued response.
H2O Deep Learning supports regression for distributions other than Gaussian such as Poisson, Gamma, Tweedie, Laplace. It also supports Huber loss and per-row offsets specified via an offset_column. We refer to our H2O Deep Learning regression code examples for more information.

To perform classification, the response must first be turned into a categorical (factor) feature:

Now the model performs (binary) classification, and has multiple (2) output neurons.

## Unsupervised Anomaly detection

For instructions on how to build unsupervised models with H2O Deep Learning, we refer to our previous Tutorial on Anomaly Detection with H2O Deep Learning and our MNIST Anomaly detection code example, as well as our Stacked AutoEncoder R code example and another one for
Unsupervised Pretraining with an AutoEncoder R code example.

## H2O Deep Learning Tips & Tricks

#### Performance Tuning

The Definitive H2O Deep Learning Performance Tuning blog post covers many of the following points that affect the computational efficiency, so it’s highly recommended.

#### Activation Functions

While sigmoids have been used historically for neural networks, H2O Deep Learning implements Tanh, a scaled and shifted variant of the sigmoid which is symmetric around 0. Since its output values are bounded by -1..1, the stability of the neural network is rarely endangered. However, the derivative of the tanh function is always non-zero and back-propagation (training) of the weights is more computationally expensive than for rectified linear units, or Rectifier, which is max(0,x) and has vanishing gradient for x&lt;=0, leading to much faster training speed for large networks and is often the fastest path to accuracy on larger problems. In case you encounter instabilities with the Rectifier (in which case model building is automatically aborted), try a limited value to re-scale the weights: max_w2=10. The Maxout activation function is computationally more expensive, but can lead to higher accuracy. It is a generalized version of the Rectifier with two non-zero channels. In practice, the Rectifier (and RectifierWithDropout, see below) is the most versatile and performant option for most problems.

#### Generalization Techniques

L1 and L2 penalties can be applied by specifying the l1 and l2 parameters. Intuition: L1 lets only strong weights survive (constant pulling force towards zero), while L2 prevents any single weight from getting too big. Dropout has recently been introduced as a powerful generalization technique, and is available as a parameter per layer, including the input layer. input_dropout_ratio controls the amount of input layer neurons that are randomly dropped (set to zero), while hidden_dropout_ratios are specified for each hidden layer. The former controls overfitting with respect to the input data (useful for high-dimensional noisy data), while the latter controls overfitting of the learned features. Note that hidden_dropout_ratios require the activation function to end with ...WithDropout.

#### Early stopping and optimizing for lowest validation error

By default, Deep Learning training stops when the stopping_metric does not improve by at least stopping_tolerance (0.01 means 1% improvement) for stopping_rounds consecutive scoring events on the training (or validation) data. By default, overwrite_with_best_model is enabled and the model returned after training for the specified number of epochs (or after stopping early due to convergence) is the model that has the best training set error (according to the metric specified by stopping_metric), or, if a validation set is provided, the lowest validation set error. Note that the training or validation set errors can be based on a subset of the training or validation data, depending on the values for score_validation_samples or score_training_samples, see below. For early stopping on a predefined error rate on the training data (accuracy for classification or MSE for regression), specify classification_stop or regression_stop.

#### Training Samples per MapReduce Iteration

The parameter train_samples_per_iteration matters especially in multi-node operation. It controls the number of rows trained on for each MapReduce iteration. Depending on the value selected, one MapReduce pass can sample observations, and multiple such passes are needed to train for one epoch. All H2O compute nodes then communicate to agree on the best model coefficients (weights/biases) so far, and the model may then be scored (controlled by other parameters below). The default value of -2 indicates auto-tuning, which attemps to keep the communication overhead at 5% of the total runtime. The parameter target_ratio_comm_to_comp controls this ratio. This parameter is explained in more detail in the H2O Deep Learning booklet,

#### Categorical Data

For categorical data, a feature with K factor levels is automatically one-hot encoded (horizontalized) into K-1 input neurons. Hence, the input neuron layer can grow substantially for datasets with high factor counts. In these cases, it might make sense to reduce the number of hidden neurons in the first hidden layer, such that large numbers of factor levels can be handled. In the limit of 1 neuron in the first hidden layer, the resulting model is similar to logistic regression with stochastic gradient descent, except that for classification problems, there’s still a softmax output layer, and that the activation function is not necessarily a sigmoid (Tanh). If variable importances are computed, it is recommended to turn on use_all_factor_levels (K input neurons for K levels). The experimental option max_categorical_features uses feature hashing to reduce the number of input neurons via the hash trick at the expense of hash collisions and reduced accuracy. Another way to reduce the dimensionality of the (categorical) features is to use h2o.glrm(), we refer to the GLRM tutorial for more details.

####Sparse Data
If the input data is sparse (many zeros), then it might make sense to enable the sparse option. This will result in the input not being standardized (0 mean, 1 variance), but only de-scaled (1 variance) and 0 values remain 0, leading to more efficient back-propagation. Sparsity is also a reason why CPU implementations can be faster than GPU implementations, because they can take advantage of if/else statements more effectively.

#### Missing Values

H2O Deep Learning automatically does mean imputation for missing values during training (leaving the input layer activation at 0 after standardizing the values). For testing, missing test set values are also treated the same way by default. See the h2o.impute function to do your own mean imputation.

#### Loss functions, Distributions, Offsets, Observation Weights

H2O Deep Learning supports advanced statistical features such as multiple loss functions, non-Gaussian distributions, per-row offsets and observation weights.
In addition to Gaussian distributions and Squared loss, H2O Deep Learning supports Poisson, Gamma, Tweedie and Laplace distributions. It also supports Absolute and Huber loss and per-row offsets specified via an offset_column. Observation weights are supported via a user-specified weights_column.

We refer to our H2O Deep Learning R test code examples for more information.

#### Exporting Weights and Biases

The model parameters (weights connecting two adjacent layers and per-neuron bias terms) can be stored as H2O Frames (like a dataset) by enabling export_weights_and_biases, and they can be accessed as follows:

#### Reproducibility

Every run of DeepLearning results in different results since multithreading is done via Hogwild! that benefits from intentional lock-free race conditions between threads. To get reproducible results for small datasets and testing purposes, set reproducible=T and set seed=1337 (pick any integer). This will not work for big data for technical reasons, and is probably also not desired because of the significant slowdown (runs on 1 core only).

#### Scoring on Training/Validation Sets During Training

The training and/or validation set errors can be based on a subset of the training or validation data, depending on the values for score_validation_samples (defaults to 0: all) or score_training_samples (defaults to 10,000 rows, since the training error is only used for early stopping and monitoring). For large datasets, Deep Learning can automatically sample the validation set to avoid spending too much time in scoring during training, especially since scoring results are not currently displayed in the model returned to R.

Note that the default value of score_duty_cycle=0.1 limits the amount of time spent in scoring to 10%, so a large number of scoring samples won’t slow down overall training progress too much, but it will always score once after the first MapReduce iteration, and once at the end of training.

Stratified sampling of the validation dataset can help with scoring on datasets with class imbalance. Note that this option also requires balance_classes to be enabled (used to over/under-sample the training dataset, based on the max. relative size of the resulting training dataset, max_after_balance_size):