# autoencoder
# Enable auto-encoder for model building.

# use_all_factor_levels
# Logical. Use all factor levels of categorical variance. Otherwise the first factor level is omittted (without loss of accuracy). Useful for variable imporotances and auto-enabled for autoencoder.

# 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); or -2 auto-tuning (default)

# seed
# Seed for random numbers (affects sampling) - Note: only reproducible when running single threaded

# adaptive_rate
# Logical. Adaptive learning rate (ADAELTA)

# rho
# Adaptive learning rate time decay factor (similarity to prior updates)

# epsilon
# Adaptive learning rate parameter, similar to learn rate annealing during initial training phase. Typical values are between 1.0e-10 and 1.0e-4

# rate
# Learning rate (higher => less stable, lower => slower convergence)

# rate_annealing
# Learning rate annealing: (rate)/(1 + rate_annealing*samples)

# rate_decay
# Learning rate decay factor between layers (N-th layer: rate*α^(N-1))

# momentum_start
# Initial momentum at the beginning of traning (try 0.5)

# momentum_ramp
# Number of training samples for which momentum increases

# momentum_stable
# Final momentum after ther amp is over (try 0.99)

# nesterov_accelerated_gradient
# Logical. Use Nesterov accelerated gradient (recommended)

# input_dropout_ratio
# A fraction of the features for each training row to be omitted from training in order to improve generalization (dimension sampling).

# hidden_dropout_ratios
# Input layer dropout ration (can improve generalization) specify one value per hidden layer, defaults to 0.5

# l1
# L1 regularization (can add stability and improve generalization, cause many weights to become 0)

# l2
# L2 regularization (can add stability and improve generalization, causes many weights to be small)

# max_w2
# Constraint for squared sum of incoming weights per unit (e.g. Rectifier)

# initial_weight_distribution
# Can be "Uniform", "UniformAdaptive", or "Normal"

# initial_weight_scale
# Unifrom: -value ... value, Normal: stddev

# loss
# Loss function: Automatic, CrossEntropy (for classification only), Quadratic, Absolute (experimental) or Huber (experimental)

# score_interval
# Shortest time interval (in secs) between model scoring

# score_training_samples
# Number of training set samples for scoring (0 for all)

# score_validation_samples
# Number of validation set samples for scoring (0 for all)

# score_duty_cycle
# Maximum duty cycle fraction for scoring (lower: more training, higher: more scoring)

# classification_stop
# Stopping criterion for classification error fraction on training data (-1 to disable)

# regression_stop
# Stopping criterion for regression error (MSE) on training data (-1 to disable)

# quiet_mode
# Enable quiet mode for less output to standard output

# max_confusion_matrix_size
# Max. size (number of classes) for confusion matrices to be shown

# max_hit_ratio_k
# Max number (top K) of predictions to use for hit ration computation(for multi-class only, 0 to disable)

# balance_classes
# Balance training data class counts via over/under-sampling (for imbalanced data)

# 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.

# max_after_balance_size
# Maximum relative size of the training data after balancing class counts (can be less than 1.0)

# score_validation_sampling
# Method used to sample validation dataset for scoring

# diagnostics
# Enable diagnostics for hidden layers

# variable_importances
# Compute variable importances for input features (Gedeon method) - can be slow for large networks)

# fast_mode
# Enable fast mode (minor approximations in back-propagation)

# ignore_const_cols
# Ignore constant columns (no information can be gained anwyay)

# force_load_balance
# Force extra load balancing to increase training speed for small datasets (to keep all cores busy)

# replicate_training_data
# Replicate the entire training dataset onto every node for faster training

# single_node_mode
# Run on a single node for fine-tuning of model parameters

# shuffle_training_data
# Enable shuffling of training data (recommended if training data is replicated and train_samples_per_iteration is close to numRows*numNodes

# col_major
# Use a column major weight matrix for input layer. Can speed up forward proagation, but might slow down backpropagation (Experimental)

# average_activation
# Average activation for sparse auto-encoder (Experimental)

# sparsity_beta
# Sparsity regularization (Experimental)

# max_categorical_features
# Max. number of categorical features, enforced via hashing Experimental)

# reproducible
# Force reproducibility on small data (will be slow - only uses 1 thread)

# export_weights_and_biases
# Whether to export Neural Network weights and biases to H2O Frames"

# ...
# extra parameters to pass onto functions (not implemented)


# Details: https://leanpub.com/deeplearning/read

#' @export
makeRLearner.regr.h2o.deeplearning = function() {
  makeRLearnerRegr(
    cl = "regr.h2o.deeplearning",
    package = "h2o",
    par.set = makeParamSet(
      makeLogicalLearnerParam("autoencoder", default = FALSE),
      makeLogicalLearnerParam("use_all_factor_level", default = TRUE),
      makeDiscreteLearnerParam("activation", values = c("Rectifier", "Tanh",
        "TanhWithDropout", "RectifierWithDropout", "Maxout", "MaxoutWithDropout"),
      default = "Rectifier"),
      # FIXME: hidden can also be a list of integer vectors for grid search
      makeIntegerVectorLearnerParam("hidden", default = c(200L, 200L),
        len = NA_integer_, lower = 1L),
      makeNumericLearnerParam("epochs", default = 10L, lower = 1), # doc says can be fractional
      makeNumericLearnerParam("train_samples_per_iteration", default = -2, lower = -2),
      makeIntegerLearnerParam("seed", tunable = FALSE),
      makeLogicalLearnerParam("adaptive_rate", default = TRUE),
      makeNumericLearnerParam("rho", default = 0.99, lower = 0), # is there a upper limit for this?
      makeNumericLearnerParam("epsilon", default = 1e-08, lower = 1e-10, upper = 1e-4),
      makeNumericLearnerParam("rate", default = 0.005, lower = 0, upper = 1),
      makeNumericLearnerParam("rate_annealing", default = 1e-06, lower = 0),
      makeNumericLearnerParam("rate_decay", default = 1, lower = 0),
      makeNumericLearnerParam("momentum_start", default = 0),
      makeNumericLearnerParam("momentum_ramp", default = 1e+06),
      makeNumericLearnerParam("momentum_stable", default = 0),
      makeLogicalLearnerParam("nesterov_accelerated_gradient", default = TRUE),
      makeNumericLearnerParam("input_dropout_ratio", default = 0),
      makeNumericVectorLearnerParam("hidden_dropout_ratios", default = 0.5),
      makeNumericLearnerParam("l1", default = 0),
      makeNumericLearnerParam("l2", default = 0),
      makeNumericLearnerParam("max_w2", default = Inf, allow.inf = TRUE),
      makeDiscreteLearnerParam("initial_weight_distribution",
        values = c("UniformAdaptive", "Uniform", "Normal"), default = "UniformAdaptive"),
      makeNumericLearnerParam("initial_weight_scale", default = 1),
      makeDiscreteLearnerParam("loss", values = c("Automatic", "Quadratic",
        "Absolute", "Huber")),
      makeDiscreteLearnerParam("distribution", values = c("AUTO", "gaussian",
        "poisson", "gamma", "tweedie", "laplace",
        "huber", "quantile"), default = "AUTO"),
      makeNumericLearnerParam("quantile_alpha", default = 0.5, lower = 0, upper = 1,
        requires = quote(distribution == "quantile")),
      makeNumericLearnerParam("tweedie_power", default = 1.5, lower = 1, upper = 2,
        requires = quote(distribution == "tweedie")),
      makeNumericLearnerParam("score_interval", default = 5),
      makeIntegerLearnerParam("score_training_samples", default = 10000),
      makeIntegerLearnerParam("score_validation_samples", default = 0),
      makeNumericLearnerParam("score_duty_cycle", default = 0.1),
      makeNumericLearnerParam("regression_stop", default = 1e-6, lower = -1),
      makeIntegerLearnerParam("stopping_rounds", default = 5L, lower = 0L),
      makeDiscreteLearnerParam("stopping_metric", values = c("AUTO", "deviance",
        "mse", "rmse", "mae", "rmsle", "r2"), default = "AUTO",
      requires = quote(stopping_rounds > 0L)),
      makeNumericLearnerParam("stopping_tolerance", default = 0, lower = 0),
      makeNumericLearnerParam("max_runtime_secs", default = 0, lower = 0),
      makeLogicalLearnerParam("quiet_mode", tunable = FALSE),
      makeLogicalLearnerParam("balance_classes", default = FALSE),
      makeNumericLearnerParam("class_sampling_factors", requires = quote(balance_classes == TRUE)),
      makeNumericLearnerParam("max_after_balance_size", default = 5),
      makeDiscreteLearnerParam("score_validation_sampling", values = c("Uniform",
        "Stratified"), default = "Uniform"),
      makeLogicalLearnerParam("diagnostics", default = TRUE, tunable = FALSE),
      makeLogicalLearnerParam("variable_importances", default = TRUE, tunable = FALSE),
      makeLogicalLearnerParam("fast_mode", default = TRUE, tunable = FALSE),
      makeLogicalLearnerParam("ignore_const_cols", default = TRUE, tunable = FALSE),
      makeLogicalLearnerParam("force_load_balance", default = TRUE, tunable = FALSE),
      makeLogicalLearnerParam("replicate_training_data", default = TRUE, tunable = FALSE),
      makeLogicalLearnerParam("single_node_mode", default = FALSE, tunable = FALSE),
      makeLogicalLearnerParam("shuffle_training_data", tunable = FALSE),
      makeLogicalLearnerParam("sparse", default = FALSE, tunable = FALSE),
      makeLogicalLearnerParam("col_major", default = FALSE, tunable = FALSE),
      makeLogicalLearnerParam("average_activation", tunable = FALSE),
      # makeLogicalLearnerParam("sparsity_beta", tunable = FALSE),
      makeLogicalLearnerParam("reproducible", default = FALSE, tunable = FALSE),
      makeLogicalLearnerParam("export_weights_and_biases", default = FALSE, tunable = FALSE)
    ),
    properties = c("numerics", "factors", "weights", "missings"),
    name = "h2o.deeplearning",
    short.name = "h2o.dl",
    note = 'The default value of `missing_values_handling` is `"MeanImputation"`, so missing values are automatically mean-imputed.',
    callees = "h2o.deeplearning"
  )
}

#' @export
trainLearner.regr.h2o.deeplearning = function(.learner, .task, .subset, .weights = NULL, ...) {

  # check if h2o connection already exists, otherwise start one
  conn.up = tryCatch(h2o::h2o.getConnection(), error = function(err) {
    return(FALSE)
  })
  if (!inherits(conn.up, "H2OConnection")) {
    h2o::h2o.init()
  }
  y = getTaskTargetNames(.task)
  x = getTaskFeatureNames(.task)
  d = getTaskData(.task, subset = .subset)
  wcol = NULL
  if (!is.null(.weights)) {
    d$.mlr.weights = .weights
    wcol = ".mlr.weights"
  }
  h2of = h2o::as.h2o(d)
  h2o::h2o.deeplearning(y = y, x = x, training_frame = h2of, weights_column = wcol, ...)
}

#' @export
predictLearner.regr.h2o.deeplearning = function(.learner, .model, .newdata, ...) {
  m = .model$learner.model
  h2of = h2o::as.h2o(.newdata)
  p = h2o::h2o.predict(m, newdata = h2of, ...)
  p.df = as.data.frame(p)
  return(p.df$predict)
}