Refactor Optional arg and Regression interface.

- Renamed spec type to usage.
- Created special constructors for regression models to hide the
  optional arguments. Each type is now defined only in that module and
  does not pollute the main Regression module.
- Changed to using polymorphic variants to express regularizers.

src/lib/classify.ml

@@ -60,7 +60,7 @@ module type Classifier_intf = sig
   val eval : t -> feature -> clas probabilities
 
   type samples = (clas * feature) list
-  val estimate : ?spec:spec -> ?classes:clas list -> samples -> t
+  val estimate : ?opt:opt -> ?classes:clas list -> samples -> t
 end
 
 module type Generative_intf = sig
@@ -136,7 +136,7 @@ let estimate_naive_bayes modulename (type c) init update incorporate
 module BinomialNaiveBayes(Data: Dummy_encoded_data_intf)
   : (Generative_intf with type feature = Data.feature
                      and type clas = Data.clas
-                     and type spec = binomial_spec)
+                     and type opt = binomial_spec)
   = struct
 
   type feature = Data.feature
@@ -174,12 +174,12 @@ module BinomialNaiveBayes(Data: Dummy_encoded_data_intf)
     in
     eval_naive_bayes ~to_prior ~to_likelihood nb.table
 
-  type spec = binomial_spec
+  type opt = binomial_spec
   let default = { smoothing = 0.0; bernoulli = false }
 
   module Cm = Map.Make(struct type t = clas let compare = compare end)
 
-  let estimate ?(spec=default) ?classes data =
+  let estimate ?(opt=default) ?classes data =
     let aa = Data.size + 1 in
     let init _cls = Array.make aa 0 in
     let update arr ftr =
@@ -190,7 +190,7 @@ module BinomialNaiveBayes(Data: Dummy_encoded_data_intf)
       arr
     in
     let incorporate all num_classes totalf =
-      let to_prob = smoothing_to_prob spec.smoothing in
+      let to_prob = smoothing_to_prob opt.smoothing in
       List.map all ~f:(fun (cl, attr_count) ->
         let prior_count = float attr_count.(Data.size) in
         let likelihood =
@@ -206,7 +206,7 @@ module BinomialNaiveBayes(Data: Dummy_encoded_data_intf)
       estimate_naive_bayes "BinomialNaiveBayes"
         init update incorporate (module Cm) ?classes data
     in
-    {table ; e_bernoulli = spec.bernoulli}
+    {table ; e_bernoulli = opt.bernoulli}
 
   let class_probabilities nb cls =
     let arr = List.assoc cls nb.table in
@@ -235,7 +235,7 @@ type smoothing = float
 module CategoricalNaiveBayes(Data: Category_encoded_data_intf)
   : (Generative_intf with type feature = Data.feature
                      and type clas = Data.clas
-                     and type spec = smoothing)
+                     and type opt = smoothing)
 
   = struct
 
@@ -274,20 +274,20 @@ module CategoricalNaiveBayes(Data: Category_encoded_data_intf)
     in
     eval_naive_bayes ~to_prior ~to_likelihood table
 
-  type spec = smoothing
+  type opt = smoothing
   let default = 0.0
 
   module Cm = Map.Make(struct type t = clas let compare = compare end)
 
-  let estimate ?(spec=default) =
+  let estimate ?(opt=default) =
     let init _ = (0, Array.map (fun i -> Array.make i 0) Data.encoding_sizes) in
     let update (c, arr) ftr =
       let ftr_arr = safe_encoding ftr in
       Array.iteri (fun i j -> arr.(i).(j) <- arr.(i).(j) + 1) ftr_arr;
       (c + 1, arr)
     in
     let incorporate all num_classes totalf =
-      let to_prob = smoothing_to_prob spec in
+      let to_prob = smoothing_to_prob opt in
       List.map all ~f:(fun (cl, (class_count, attr_count)) ->
         let prior      = to_prob (float class_count) totalf (float num_classes) in
         let likelihood =
@@ -322,7 +322,7 @@ let to_safe_encoding_size_checked interfacename size encoding f =
 module GaussianNaiveBayes(Data: Continuous_encoded_data_intf)
   : (Generative_intf with type feature = Data.feature
                      and type clas = Data.clas
-                     and type spec = unit)
+                     and type opt = unit)
 
   = struct
 
@@ -353,13 +353,13 @@ module GaussianNaiveBayes(Data: Continuous_encoded_data_intf)
     in
     eval_naive_bayes ~to_prior ~to_likelihood table
 
-  type spec = unit
+  type opt = unit
   let default = ()
 
   module Cm = Map.Make(struct type t = clas let compare = compare end)
 
-  let estimate ?(spec=default) =
-    ignore spec;
+  let estimate ?(opt=default) =
+    ignore opt;
     let init _c = (0, Array.make Data.size Running.empty) in
     let update (c, rs_arr) ftr =
       let attr = safe_encoding ftr in
@@ -396,7 +396,7 @@ module LrCommon(Data: Continuous_encoded_data_intf) = struct
 
   type samples = (clas * feature) list
 
-  type spec = log_reg_spec
+  type opt = log_reg_spec
   let default = { lambda = 1e-4
                 ; tolerance = 1e4
                 }
@@ -408,7 +408,7 @@ module LrCommon(Data: Continuous_encoded_data_intf) = struct
   let copy1 arr = Array.init (Data.size + 1) (function | 0 -> 1. | i -> arr.(i - 1))
 
   (* map classes to [1;2 ... 3], convert features to matrix and run Softmax *)
-  let estimate ~method_name ~class_bound ~to_t ?(spec=default) ?(classes=[]) data =
+  let estimate ~method_name ~class_bound ~to_t ?(opt=default) ?(classes=[]) data =
     let class_bound =
       match class_bound with
       | None   -> fun n -> n
@@ -444,8 +444,8 @@ module LrCommon(Data: Continuous_encoded_data_intf) = struct
         in
         let weights =
           Softmax_regression.regress
-            ~lambda:spec.lambda
-            ~tolerance:spec.tolerance
+            ~lambda:opt.lambda
+            ~tolerance:opt.tolerance
             ftrs classes
         in
         let sortedc =
@@ -461,7 +461,7 @@ module LogisticRegression(Data: Continuous_encoded_data_intf)
   : sig
     include Classifier_intf with type feature = Data.feature
                             and type clas = Data.clas
-                            and type spec = log_reg_spec
+                            and type opt = log_reg_spec
 
     val coefficients : t -> float array
 
@@ -504,7 +504,7 @@ module MulticlassLogisticRegression(Data: Continuous_encoded_data_intf)
   : sig
     include Classifier_intf with type feature = Data.feature
                             and type clas = Data.clas
-                            and type spec = log_reg_spec
+                            and type opt = log_reg_spec
 
     val coefficients : t -> float array array
 

src/lib/classify.mli

@@ -109,22 +109,22 @@ module type Classifier_intf = sig
   (** Representing training data. *)
   type samples = (clas * feature) list
 
-  (** [estimate spec classes samples] estimates a classifier based upon the
+  (** [estimate opt classes samples] estimates a classifier based upon the
       training [samples].
 
       [classes] is an optional argument to specify ahead of time the possible
       classes to train on (defaults to the ones found in the training data).
       This is useful for models where we know the population domain but may
       not see an example of a training datum for rare cases.
 
-      [spec] are the optional classifier dependent estimation/evaluation
+      [opt] are the optional classifier dependent estimation/evaluation
       arguments.
 
       @raise Invalid_argument if [classes] are specified and new ones are
       found in the training [samples].
       @raise Invalid_argument if [samples] is empty.
   *)
-  val estimate : ?spec:spec -> ?classes:clas list -> samples -> t
+  val estimate : ?opt:opt -> ?classes:clas list -> samples -> t
 end
 
 (** A generative classifier builds models of the form
@@ -172,7 +172,7 @@ type binomial_spec =
   classifier on data encoded using
   {{!modtype:Dummy_encoded_data_intf}Dummy variables.} *)
 module BinomialNaiveBayes(D: Dummy_encoded_data_intf) :
-  Generative_intf with type spec = binomial_spec
+  Generative_intf with type opt = binomial_spec
                   and type feature = D.feature
                   and type clas = D.clas
 
@@ -181,7 +181,7 @@ module BinomialNaiveBayes(D: Dummy_encoded_data_intf) :
   classifier on data encoded using
   {{!modtype:Category_encoded_data_intf}Categorical variables.} *)
 module CategoricalNaiveBayes(D: Category_encoded_data_intf) :
-  Generative_intf with type spec = smoothing
+  Generative_intf with type opt = smoothing
                   and type feature = D.feature
                   and type clas = D.clas
 
@@ -191,7 +191,7 @@ module CategoricalNaiveBayes(D: Category_encoded_data_intf) :
   for each of the quantitative features in the
   {{!modtype:Continuous_encoded_data_intf}encoded data}. *)
 module GaussianNaiveBayes(D: Continuous_encoded_data_intf) :
-  Generative_intf with type spec = unit
+  Generative_intf with type opt = unit
                   and type feature = D.feature
                   and type clas = D.clas
 
@@ -230,7 +230,7 @@ type log_reg_spec =
 *)
 module LogisticRegression(D: Continuous_encoded_data_intf) :
   sig
-    include Classifier_intf with type spec = log_reg_spec
+    include Classifier_intf with type opt = log_reg_spec
                             and type feature = D.feature
                             and type clas = D.clas
 
@@ -259,7 +259,7 @@ module LogisticRegression(D: Continuous_encoded_data_intf) :
 *)
 module MulticlassLogisticRegression(D: Continuous_encoded_data_intf) :
   sig
-    include Classifier_intf with type spec = log_reg_spec
+    include Classifier_intf with type opt = log_reg_spec
                              and type feature = D.feature
                              and type clas = D.clas
 

src/lib/classify.mlt

@@ -58,7 +58,7 @@ let () =
           let size = 5
         end)
       in
-      let naiveb = NB.estimate ~spec:{NB.default with bernoulli = true } data in
+      let naiveb = NB.estimate ~opt:{NB.default with bernoulli = true } data in
       let sample = [ `shortbread ; `whiskey; `porridge  ] in
       let result = NB.eval naiveb sample in
       let expect =