Build a Deep Learning model using multiple native GPU backends

Source: R/deepwater.R

Builds a deep neural network on an H2OFrame containing various data sources.

h2o.deepwater(x, y, training_frame, model_id = NULL, checkpoint = NULL,
  autoencoder = FALSE, validation_frame = NULL, nfolds = 0,
  balance_classes = FALSE, max_after_balance_size = 5,
  class_sampling_factors = NULL, keep_cross_validation_predictions = FALSE,
  keep_cross_validation_fold_assignment = FALSE, fold_assignment = c("AUTO",
  "Random", "Modulo", "Stratified"), fold_column = NULL,
  offset_column = NULL, weights_column = NULL,
  score_each_iteration = FALSE, categorical_encoding = c("AUTO", "Enum",
  "OneHotInternal", "OneHotExplicit", "Binary", "Eigen", "LabelEncoder",
  "SortByResponse", "EnumLimited"), overwrite_with_best_model = TRUE,
  epochs = 10, train_samples_per_iteration = -2,
  target_ratio_comm_to_comp = 0.05, seed = -1, standardize = TRUE,
  learning_rate = 0.001, learning_rate_annealing = 1e-06,
  momentum_start = 0.9, momentum_ramp = 10000, momentum_stable = 0.9,
  distribution = c("AUTO", "bernoulli", "multinomial", "gaussian", "poisson",
  "gamma", "tweedie", "laplace", "quantile", "huber"), score_interval = 5,
  score_training_samples = 10000, score_validation_samples = 0,
  score_duty_cycle = 0.1, classification_stop = 0, regression_stop = 0,
  stopping_rounds = 5, stopping_metric = c("AUTO", "deviance", "logloss",
  "MSE", "RMSE", "MAE", "RMSLE", "AUC", "lift_top_group", "misclassification",
  "mean_per_class_error"), stopping_tolerance = 0, max_runtime_secs = 0,
  ignore_const_cols = TRUE, shuffle_training_data = TRUE,
  mini_batch_size = 32, clip_gradient = 10, network = c("auto", "user",
  "lenet", "alexnet", "vgg", "googlenet", "inception_bn", "resnet"),
  backend = c("mxnet", "caffe", "tensorflow"), image_shape = c(0, 0),
  channels = 3, sparse = FALSE, gpu = TRUE, device_id = c(0),
  cache_data = TRUE, network_definition_file = NULL,
  network_parameters_file = NULL, mean_image_file = NULL,
  export_native_parameters_prefix = NULL, activation = c("Rectifier",
  "Tanh"), hidden = NULL, input_dropout_ratio = 0,
  hidden_dropout_ratios = NULL, problem_type = c("auto", "image",
  "dataset"))

Arguments

x	(Optional) A vector containing the names or indices of the predictor variables to use in building the model. If x is missing, then all columns except y are used.
y	The name or column index of the response variable in the data. The response must be either a numeric or a categorical/factor variable. If the response is numeric, then a regression model will be trained, otherwise it will train a classification model.
training_frame	Id of the training data frame.
model_id	Destination id for this model; auto-generated if not specified.
checkpoint	Model checkpoint to resume training with.
autoencoder	Logical. Auto-Encoder. Defaults to FALSE.
validation_frame	Id of the validation data frame.
nfolds	Number of folds for K-fold cross-validation (0 to disable or >= 2). Defaults to 0.
balance_classes	Logical. Balance training data class counts via over/under-sampling (for imbalanced data). Defaults to FALSE.
max_after_balance_size	Maximum relative size of the training data after balancing class counts (can be less than 1.0). Requires balance_classes. Defaults to 5.0.
class_sampling_factors	Desired over/under-sampling ratios per class (in lexicographic order). If not specified, sampling factors will be automatically computed to obtain class balance during training. Requires balance_classes.
keep_cross_validation_predictions	Logical. Whether to keep the predictions of the cross-validation models. Defaults to FALSE.
keep_cross_validation_fold_assignment	Logical. Whether to keep the cross-validation fold assignment. Defaults to FALSE.
fold_assignment	Cross-validation fold assignment scheme, if fold_column is not specified. The 'Stratified' option will stratify the folds based on the response variable, for classification problems. Must be one of: "AUTO", "Random", "Modulo", "Stratified". Defaults to AUTO.
fold_column	Column with cross-validation fold index assignment per observation.
offset_column	Offset column. This will be added to the combination of columns before applying the link function.
weights_column	Column with observation weights. Giving some observation a weight of zero is equivalent to excluding it from the dataset; giving an observation a relative weight of 2 is equivalent to repeating that row twice. Negative weights are not allowed. Note: Weights are per-row observation weights and do not increase the size of the data frame. This is typically the number of times a row is repeated, but non-integer values are supported as well. During training, rows with higher weights matter more, due to the larger loss function pre-factor.
score_each_iteration	Logical. Whether to score during each iteration of model training. Defaults to FALSE.
categorical_encoding	Encoding scheme for categorical features Must be one of: "AUTO", "Enum", "OneHotInternal", "OneHotExplicit", "Binary", "Eigen", "LabelEncoder", "SortByResponse", "EnumLimited". Defaults to AUTO.
overwrite_with_best_model	Logical. If enabled, override the final model with the best model found during training. Defaults to TRUE.
epochs	How many times the dataset should be iterated (streamed), can be fractional. Defaults to 10.
train_samples_per_iteration	Number of training samples (globally) per MapReduce iteration. Special values are 0: one epoch, -1: all available data (e.g., replicated training data), -2: automatic. Defaults to -2.
target_ratio_comm_to_comp	Target ratio of communication overhead to computation. Only for multi-node operation and train_samples_per_iteration = -2 (auto-tuning). Defaults to 0.05.
seed	Seed for random numbers (affects certain parts of the algo that are stochastic and those might or might not be enabled by default) Note: only reproducible when running single threaded. Defaults to -1 (time-based random number).
standardize	Logical. If enabled, automatically standardize the data. If disabled, the user must provide properly scaled input data. Defaults to TRUE.
learning_rate	Learning rate (higher => less stable, lower => slower convergence). Defaults to 0.001.
learning_rate_annealing	Learning rate annealing: rate / (1 + rate_annealing * samples). Defaults to 1e-06.
momentum_start	Initial momentum at the beginning of training (try 0.5). Defaults to 0.9.
momentum_ramp	Number of training samples for which momentum increases. Defaults to 10000.
momentum_stable	Final momentum after the ramp is over (try 0.99). Defaults to 0.9.
distribution	Distribution function Must be one of: "AUTO", "bernoulli", "multinomial", "gaussian", "poisson", "gamma", "tweedie", "laplace", "quantile", "huber". Defaults to AUTO.
score_interval	Shortest time interval (in seconds) between model scoring. Defaults to 5.
score_training_samples	Number of training set samples for scoring (0 for all). Defaults to 10000.
score_validation_samples	Number of validation set samples for scoring (0 for all). Defaults to 0.
score_duty_cycle	Maximum duty cycle fraction for scoring (lower: more training, higher: more scoring). Defaults to 0.1.
classification_stop	Stopping criterion for classification error fraction on training data (-1 to disable). Defaults to 0.
regression_stop	Stopping criterion for regression error (MSE) on training data (-1 to disable). Defaults to 0.
stopping_rounds	Early stopping based on convergence of stopping_metric. Stop if simple moving average of length k of the stopping_metric does not improve for k:=stopping_rounds scoring events (0 to disable) Defaults to 5.
stopping_metric	Metric to use for early stopping (AUTO: logloss for classification, deviance for regression) Must be one of: "AUTO", "deviance", "logloss", "MSE", "RMSE", "MAE", "RMSLE", "AUC", "lift_top_group", "misclassification", "mean_per_class_error". Defaults to AUTO.
stopping_tolerance	Relative tolerance for metric-based stopping criterion (stop if relative improvement is not at least this much) Defaults to 0.
max_runtime_secs	Maximum allowed runtime in seconds for model training. Use 0 to disable. Defaults to 0.
ignore_const_cols	Logical. Ignore constant columns. Defaults to TRUE.
shuffle_training_data	Logical. Enable global shuffling of training data. Defaults to TRUE.
mini_batch_size	Mini-batch size (smaller leads to better fit, larger can speed up and generalize better). Defaults to 32.
clip_gradient	Clip gradients once their absolute value is larger than this value. Defaults to 10.
network	Network architecture. Must be one of: "auto", "user", "lenet", "alexnet", "vgg", "googlenet", "inception_bn", "resnet". Defaults to auto.
backend	Deep Learning Backend. Must be one of: "mxnet", "caffe", "tensorflow". Defaults to mxnet.
image_shape	Width and height of image. Defaults to [0, 0].
channels	Number of (color) channels. Defaults to 3.
sparse	Logical. Sparse data handling (more efficient for data with lots of 0 values). Defaults to FALSE.
gpu	Logical. Whether to use a GPU (if available). Defaults to TRUE.
device_id	Device IDs (which GPUs to use). Defaults to [0].
cache_data	Logical. Whether to cache the data in memory (automatically disabled if data size is too large). Defaults to TRUE.
network_definition_file	Path of file containing network definition (graph, architecture).
network_parameters_file	Path of file containing network (initial) parameters (weights, biases).
mean_image_file	Path of file containing the mean image data for data normalization.
export_native_parameters_prefix	Path (prefix) where to export the native model parameters after every iteration.
activation	Activation function. Only used if no user-defined network architecture file is provided, and only for problem_type=dataset. Must be one of: "Rectifier", "Tanh".
hidden	Hidden layer sizes (e.g. [200, 200]). Only used if no user-defined network architecture file is provided, and only for problem_type=dataset.
input_dropout_ratio	Input layer dropout ratio (can improve generalization, try 0.1 or 0.2). Defaults to 0.
hidden_dropout_ratios	Hidden layer dropout ratios (can improve generalization), specify one value per hidden layer, defaults to 0.5.
problem_type	Problem type, auto-detected by default. If set to image, the H2OFrame must contain a string column containing the path (URI or URL) to the images in the first column. If set to text, the H2OFrame must contain a string column containing the text in the first column. If set to dataset, Deep Water behaves just like any other H2O Model and builds a model on the provided H2OFrame (non-String columns). Must be one of: "auto", "image", "dataset". Defaults to auto.