vwsetup: Create Vowpal Wabbit model, setup model parameters and data
In ivan-pavlov/rvwgsoc: R Interface for Vowpal Wabbit

Description Usage Arguments Value Examples

Sets up VW model together with parameters and data

vwsetup(algorithm = c("sgd", "bfgs", "ftrl", "pistol", "ksvm",
  "OjaNewton", "svrg"), general_params = list(),
  feature_params = list(), optimization_params = list(),
  dir = tempdir(), model = NULL, params_str = NULL, option = c("",
  "binary", "oaa", "ect", "csoaa", "wap", "log_multi", "recall_tree",
  "lda", "multilabel_oaa", "classweight", "new_mf", "lrq", "stage_poly",
  "bootstrap", "autolink", "replay", "explore_eval", "cb", "cb_explore",
  "cbify", "multiworld_test_check", "nn", "topk", "search", "boosting",
  "marginal"), ...)

`algorithm`	[string] Optimzation algorithm `sgd` - adaptive, normalized, invariant stochastic gradient descent `bfgs` - Limited-memory Broyden-Fletcher-Goldfarb-Shanno optimization algorithm `ftrl` - FTRL: Follow the Regularized Leader optimization algorithm `pistol` - FTRL: Parameter-free Stochastic Learning `ksvm` - Kernel svm `OjaNewton` - Online Newton with Oja's Sketch `svrg` - Stochastic Variance Reduced Gradient
`general_params`	List of parameters: `random_seed` [int] - Seed random number generator (default: 0) `ring_size` [int] - Size of example ring `holdout_off` [bool] - No holdout data in multiple passes (default: FALSE) `holdout_period` [int] - Holdout period for test only (default: 10) `holdout_after` [int] - Holdout after n training examples, default off (disables holdout_period) (default: 0) `early_terminate` [int] - Specify the number of passes tolerated when holdout loss doesn't decrease before early termination (default: 3) `loss_function` [string] - Specify the loss function to be used, uses squared by default. Currently available ones are: squared, classic, hinge, logistic, quantile and poisson. (default: squared) `link` [string] - Specify the link function: identity, logistic, glf1 or poisson. (default: identity) `quantile_tau` [real] - Parameter "tau" associated with Quantileloss. (default: 0.5)
`feature_params`	List of parameters: More information about "interactions" option (also "quadratic", "cubic") avaliable here https://github.com/VowpalWabbit/vowpal_wabbit/wiki/Command-line-arguments#example-manipulation-options `bit_precision` [int] - Number of bits in the feature table (default: 18) `quadratic` [string] - Create and use quadratic features (Specify 2 namespaces) `cubic` [string] - Create and use cubic features (Specify 3 namespaces) `interactions` [string] - Create feature interactions of any level between namespaces (Specify several namespaces) `permutations` [bool] - Use permutations instead of combinations for feature interactions of same namespace (default: FALSE) `leave_duplicate_interactions` [bool] - Don't remove interactions with duplicate combinations of namespaces. For ex. this is a duplicate: 'quadratic="ab", quadratic="ba"' and a lot more in 'quadratic="::"'. (default: FALSE) `noconstant` [bool] - Don't add a constant feature (default: FALSE) `feature_limit` [string] - limit to N features. To apply to a single namespace 'foo', arg should be "fN" `ngram` [string] - Generate N grams. To generate N grams for a single namespace 'foo', arg should be "fN". `skips` [string] - Use second derivative in line searchGenerate skips in N grams. This in conjunction with the ngram tag can be used to generate generalized n-skip-k-gram. To generate n-skips for a single namespace 'foo', arg should be "fN". `hash` [string] - How to hash the features. Available options: "strings", "all" (default: "strings") `affix` [string] - Generate prefixes/suffixes of features; argument "+2a,-3b,+1" means generate 2-char prefixes for namespace a, 3-char suffixes for b and 1 char prefixes for default namespace `spelling` [string] - Compute spelling features for a given namespace (use '_' for default namespace) `interact` [string] - Put weights on feature products from namespaces <n1> and <n2>
`optimization_params`	List of parameters: `learning_rate` [real] - Set initial learning Rate (default: 0.5) `initial_pass_length` [int] - Initial number of examples per pass `l1` [real] - L1 regularization (default: 0) `l2` [real] - L2 regularization (default: 0) `no_bias_regularization` [string] - no bias in regularization (Available options: "on", "off") `feature_mask` [string] - Use existing regressor to determine which parameters may be updated. If no initial_regressor given, also used for initial weights. `decay_learning_rate` [real] - Set Decay factor for learning_rate between passes (default: 1) `initial_t` [real] - initial t value (default: 0) `power_t` [real] - t power value (default: 0.5) `initial_weight` [int] - Set all weights to an initial value of arg (default: 0) `random_weights` [string] - Make initial weights random (Available options: "on", "off") (default: "off") `normal_weights` [string] - Make initial weights normal (Available options: "on", "off") (default: "off") `truncated_normal_weights` [string] - Make initial weights truncated normal (Available options: "on", "off") (default: "off") `sparse_weights` [bool] - Use a sparse datastructure for weights. `input_feature_regularizer` [string] - Per feature regularization input file. Additional parameters depending on `algorithm` choice: `sgd`: `adaptive` [bool] - Use adaptive, individual learning rates (default: TRUE) `normalized` [bool] - Use per feature normalized updates (default: TRUE) `invariant` [bool] - Use safe/importance aware updates (default: TRUE) `adax` [bool] - Use adaptive learning rates with x^2 instead of g^2x^2 (default: FALSE) `sparse_l2` [real] - use per feature normalized updates (default: 0) `l1_state` [real] - use per feature normalized updates (default: 0) `l2_state` [real] - use per feature normalized updates (default: 1) `bfgs`: `conjugate_gradient` [bool] - Use conjugate gradient based optimization (default: FALSE) `hessian_on` [bool] - Use second derivative in line search (default: FALSE) `mem` [int] - Memory in bfgs. (default: 15) `termination` [real] - Termination threshold. (default: 0.00100000005) `ftrl`: `ftrl_alpha` [real] - Learning rate for FTRL optimization (default: 0.005) `ftrl_beta` [real] - FTRL beta parameter (default: 0.1) `pistol`: `ftrl_alpha` [real] - Learning rate for FTRL optimization (default: 0.005) `ftrl_beta` [real] - FTRL beta parameter (default: 0.1) `ksvm`: `reprocess` [int] - number of reprocess steps for LASVM (default: 1) `kernel` [string] - type of kernel (rbf or linear) (default: "linear") `bandwidth` [real] - bandwidth of rbf kernel (default: 1.0) `degree` [int] - degree of poly kernel (default: 2) `lambda` [real] - saving regularization for test time (default: -1) `OjaNewton`: `sketch_size` [int] - size of sketch (default: 10) `epoch_size` [int] - size of epoch (default: 1) `alpha` [real] - multiplicative constant for identity (default: 1) `alpha_inverse` [real] - one over alpha, similar to learning rate `learning_rate_cnt` - constant for the learning rate 1/t (default: 2) `normalize` [string] - normalize the features or not (Available options: "on", "off") (default: "on") `random_init` [string] - randomize initialization of Oja or not (Available options: "on", "off") (default: "on") `svrg`: `stage_size` [int] - Number of passes per SVRG stage (default: 1)
`dir`	[string] Working directory path, default is tempdir()
`model`	[string] File name for model weights or path to existng model file.
`params_str`	[string] Pass cmd line parameters directly, bypassing the default approach. For compatibility, parameters from vwtrain,vwtest, predict.vw can't be used here and functions add_option, vwparams aren't supported.
`option`	[string] Add Learning algorithm / reduction option: `binary` - Reports loss as binary classification with -1,1 labels `oaa` - One-against-all multiclass learning with labels `ect` - Error correcting tournament with labels `csoaa` - One-against-all multiclass learning with costs `wap` - Weighted all-pairs multiclass learning with costs `multilabel_oaa` - One-against-all multilabel with multiple labels `log_multi` - Online (decision) trees for classes `classweight` - Importance weight classes `lda` - Latent Dirichlet Allocation `recall_tree` - Use online tree for multiclass `new_mf` - Matrix factorization mode `lrq` - Low rank quadratic features `stage_poly` - Stagewise polynomial features `bootstrap` - bootstrap with K rounds by online importance resampling `autolink` - Create link function with polynomial N `replay` - Experience Replay `explore_eval` - Explore evaluation `cb` - Contextual bandit learning `cb_explore` - Contextual Bandit Exploration `cbify` - Convert multiclass on K classes into a contextual bandit problem `multiworld_test` - Multiworld Testing `nn` - Sigmoidal feedforward network `topk` - Top K recommendation `struct_search` - Search-based structured prediction (SEARN or DAgger) `boosting` - Online boosting with weak learners `marginal` - Substitute marginal label estimates for ids
`...`	Additional options for a learning algorithm / reduction `oaa` or `ect`: `num_classes` [int] - Number of classes `oaa_subsample` [int] - Subsample this number of negative examples when learning `multilabel_oaa`: `num_labels` [int] - Number of labels `csoaa` or `wap`: `num_classes` [int] - Number of classes `csoaa_ldf` or `wap_ldf` - `singleline` (Default) or `multiline` label dependent features `log_multi`: `num_classes` [int] - Number of classes `no_progress` [bool] - Disable progressive validation (default: FALSE) `swap_resistance` [int] - Higher = more resistance to swap, (default: 4) `classweight`: `class_multiplier` [real] - importance weight multiplier for class `recall_tree`: `num_classes` [int] - Number of classes `max_candidates` [int] - Maximum number of labels per leaf in the tree `bern_hyper` [real] - Recall tree depth penalty (default: 1) `max_depth` [int] - Maximum depth of the tree, (default: log_2(number of classes) ) `node_only` [string] - Only use node features, not full path (Available options: "on", "off") (default: "off") `randomized_routing` [string] - Randomized routing (Available options: "on", "off") (default: "off") `lda`: `num_topics` [int] - Number of topics `lda_alpha` [real] - Prior on sparsity of per-document topic weights (default: 0.100000001) `lda_rho` [real] - Prior on sparsity of topic distributions (default: 0.100000001) `lda_D` [int] - Number of documents (default: 10000) `lda_epsilon` [real] - Loop convergence threshold (default: 0.00100000005) `math-mode` [string] - Math mode: simd, accuracy, fast-approx `minibatch` [int] - Minibatch size (default: 1) `metrics` [string] - Compute metrics (Available options: "on", "off") (default: "off") `new_mf`: `rank` [int] - rank for matrix factorization `lrq`: `features` [string] - low rank quadratic features `lrqdropout` [bool] - use dropout training for low rank quadratic features (default: FALSE) `stage_poly`: `sched_exponent` [real] - exponent controlling quantity of included features (default: 1.0) `batch_sz` [int] - multiplier on batch size before including more features (default: 1000) `batch_sz_no_doubling` [bool] - batch_sz does not double (default: TRUE) `bootstrap`: `num_rounds` [int] - number of rounds `bs_type` [string] - the bootstrap mode: 'mean' or 'vote' (default: "mean") `autolink`: `degree` [int] - polynomial degree (default: 2) `replay`: `level` [string] - Use experience replay at a specified level (b=classification/regression, m=multiclass, c=cost sensitive) `buffer` [int] - Buffer size (default: 100) `count` [int] - how many times (in expectation) should each example be played (default: 1 = permuting) `explore_eval`: `multiplier` [real] - Multiplier used to make all rejection sample probabilities <= 1 `cb`: `num_costs` [int] - number of num_costs If costs=0, contextual bandit learning with multiline action dependent features (ADF) is triggered ("–cb_adf"). `cb_type` [string] - contextual bandit method to use in ips,dm,dr, mtr (for ADF) (default: "dr") `eval` [bool] - Evaluate a policy rather than optimizing (default: FALSE) `rank_all` [bool] - Return actions sorted by score order. (for ADF) (default: FALSE) `no_predict` [bool] - Do not do a prediction when training. (for ADF) (default: FALSE) `cb_explore`: `num_actions` [bool] - number of actions in online explore-exploit for a <k> action contextual bandit problem. If num_actions=0, online explore-exploit for a contextual bandit problem with multiline action dependent features (ADF) is triggered ("–cb_explore_adf"). `explore_type` [string] - Type of exploration to use: "epsilon" (epsilon-greedy exploration) (default), "first" (tau-first exploration), "bag" (bagging-based exploration), "cover" (Online cover based exploration), "softmax" (softmax exploration), "regcb" (RegCB-elim exploration), "regcbopt" (RegCB optimistic exploration). "softmax", "regcb" and "regcbopt" types are only avaliable for exploration with ADF. (default: "epsilon") `explore_arg` [real] - Parameter for exploration algorithm. Applicable for "epsilon", "first", "bag" and "cover" types of exploration. (default: 0.05) `psi` [real] - Disagreement parameter for "cover" algorithm. (default: 1) `nounif` [bool] - Do not explore uniformly on zero-probability actions in "cover" algorithm. (default: FALSE) `mellowness` [real] - "RegCB" mellowness parameter c_0. (default: 0.1) `greedify` [bool] - Always update first policy once in "bag" (default: FALSE) `lambda` [real] - Parameter for "softmax". (default: -1) `cb_min_cost` [real] - Lower bound on cost. (default: 0) For ADF only `cb_max_cost` [real] - Upper bound on cost. (default: 1) For ADF only `first_only` [bool] - Only explore the first action in a tie-breaking event. For ADF only (default: FALSE) `cbify`: `num_classes` [int] - number of classes `cbify_cs` [bool] - consume cost-sensitive classification examples instead of multiclass (default: FALSE) `loss0` [real] - loss for correct label (default: 0) `loss1` [real] - loss for incorrect label (default: 1) `multiworld_test`: `features` [string] - Evaluate features as a policies `learn` [int] - Do Contextual Bandit learning on <n> classes. `num_classes` [bool] - Discard mwt policy features before learning (default: FALSE) `nn`: `num_hidden` [int] - number of hidden units `inpass` [bool] - Train or test sigmoidal feedforward network with input passthrough (default: FALSE) `multitask` [bool] - Share hidden layer across all reduced tasks (default: FALSE) `dropout` [bool] - Train or test sigmoidal feedforward network using dropout (default: FALSE) `meanfield` [bool] - Train or test sigmoidal feedforward network using mean field (default: FALSE) `topk`: `num_k` [int] - number of top k recomendations `struct_search`: `id` [int] - maximum action id or 0 for LDF `search_task` [string] - search task: sequence, sequencespan, sequence_ctg, argmax, sequence_demoldf, multiclasstask, dep_parser, entity_relation, hook, graph `search_interpolation` [string] - at what level should interpolation happen? (data or policy) `search_rollout` [string] - how should rollouts be executed? (policy, oracle, mix_per_state, mix_per_roll, none) `search_rollin` [string] - how should past trajectories be generated? (policy, oracle, mix_per_state, mix_per_roll) `search_passes_per_policy` [int] - number of passes per policy (only valid for search_interpolation=policy). (default: 1) `search_beta` [real] - interpolation rate for policies (only valid for search_interpolation=policy). (default: 0.5) `search_alpha` [real] - annealed beta = 1-(1-alpha)^t (only valid for search_interpolation=data). (default: 1e-10) `search_total_nb_policies` [int] - if we are going to train the policies through multiple separate calls to vw, we need to specify this parameter and tell vw how many policies are eventually going to be trained `search_trained_nb_policies` [int] - the number of trained policies in a file `search_allowed_transitions` [string] - read file of allowed transitions. default: all transitions are allowed `search_subsample_time` [real] - instead of training at all timesteps, use a subset. if value in (0,1), train on a random v `search_neighbor_features` [string] - copy features from neighboring lines. argument looks like: '-1:a,+2' meaning copy previous line from namespace "a" and next line from namespace "unnamed", where ',' separates them `search_rollout_num_steps` [int] - how many calls of "loss" before we stop really predicting on rollouts and switch to oracle (default means "infinite") `search_history_length` [int] - some tasks allow you to specify how much history their depend on; specify that here. (default: 1) `search_no_caching` [bool] - turn off the built-in caching ability (makes things slower, but technically more safe) (default: FALSE) `search_xv` [bool] - train two separate policies, alternating prediction/learning. (default: FALSE) `search_perturb_oracle` [real] - perturb the oracle on rollin with this probability. (default: 0) `search_linear_ordering` [bool] - insist on generating examples in linear order. (default: FALSE and using hoopla permutation) `search_active_verify` [real] - verify that active learning is doing the right thing (arg = multiplier, should be = cost_range * range_c) `search_save_every_k_runs` [int] - save model every k runs `boosting`: `num_learners` [int] - number of weak learners `gamma` [real] - weak learner's edge (=0.1), used only by online BBM (default: 0.100000001) `alg` - specify the boosting algorithm: BBM (default), logistic (AdaBoost.OL.W), adaptive (AdaBoost.OL) (default: "BBM") `marginal`: `ids` [string] - Substitute marginal label estimates for ids `initial_denominator` [real] - Initial denominator (default: 1) `initial_numerator` [real] - Initial numerator (default: 0.5) `compete` [bool] - Enable competition with marginal features (default: FALSE) `update_before_learn` [string] - Update marginal values before learning (Available options: "on", "off") (default: "off") `unweighted_marginals` [string] - Ignore importance weights when computing marginals (Available options: "on", "off") (default: "off") `decay` [real] - Decay multiplier per event (1e-3 for example) (default=0)

vwmodel list class

vwsetup(
 dir = tempdir(),
 model = "pk_mdl.vw",
 general_params = list(loss_function="logistic", link="logistic"),
 optimization_params = list(adaptive=FALSE),
 option = "binary"
)