removed :pre-metric-standarisation-fn

Author: behrica

deps.edn

@@ -2,7 +2,7 @@
  :deps {org.clojure/clojure                     {:mvn/version "1.12.4"}
         org.scicloj/metamorph.ml
         {:git/url "https://github.com/scicloj/metamorph.ml"
-         :git/sha "2cac3eff15792462aaf07ebf34f758e694214d45"}
+         :git/sha "86f5c18b4c9eb12ca187eae275357db53155d88b"}
         ;{:mvn/version "1.3.2"}
 
         cheshire/cheshire                       {:mvn/version "6.1.0"}
@@ -20,6 +20,7 @@
                 same/ish                              {:mvn/version "0.1.7"}
                 ;;  the unit tests needs this
                 org.tribuo/tribuo-regression-sgd      {:mvn/version "4.3.2"}
+                org.tribuo/tribuo-classification-libsvm      {:mvn/version "4.3.2"}
                 org.tribuo/tribuo-classification-tree {:mvn/version "4.3.2"}}}
 
 

src/scicloj/ml/tribuo.clj

@@ -231,40 +231,21 @@
 
 
 (defn pre-metric-standardise 
-  "converts prediction result and the trueth into either
-   seq of 
-   :discrete     keyword,string,intXX,...
-   :continous   double, float
-   or fails.
-
-   `prediction-ds` and `thrueth-ds` are tabular data, 
-   usualy of type tech.v3.dataset
-
-   returns map of 
-   :prediction (seq)
-   :trueth     (seq)
-
-   I case of :discrete the discrete values in :predicion and :trueth
-   should have semantically identical meaning, as they might get
-   compared via '=' later
-   "
   [prediction-ds trueth-ds discrete-or-continous]
-  
   {:prediction (do-harmonize-prediction prediction-ds discrete-or-continous)
    :trueth (do-harmonize-trueth trueth-ds discrete-or-continous)}
   )
 
 
 (defn- score 
-  ([model scoring-ds options]
+  ([prediction-ds trueth-ds metric-fn]
    ;; classificatioon only
-   (let [prediction (ml/predict (cf/feature scoring-ds) model)
-         trueth (cf/target scoring-ds)
-         standardised (pre-metric-standardise prediction trueth :discrete)
+   (let [
+         standardised (pre-metric-standardise prediction-ds trueth-ds :discrete)
          prediction-values (:prediction standardised)
          trueth-values (:trueth standardised)
          _ (safety-first! prediction-values trueth-values)]
-     (loss/classification-accuracy prediction-values trueth-values)))
+     (metric-fn prediction-values trueth-values)))
   ([model scoring-ds](score model scoring-ds nil))
   )
 
@@ -275,7 +256,6 @@
   {:glance-fn glance-fn-regression
    :augment-fn augment-fn-regression
    :thaw-fn thaw
-   :pre-metric-standarisation-fn pre-metric-standardise
    })
 
 
@@ -285,7 +265,6 @@
   predict-classification
   {:thaw-fn thaw
    :score-fn score
-   :pre-metric-standarisation-fn pre-metric-standardise
    })
 
 

test/scicloj/ml/linear_regression_test.clj

@@ -96,44 +96,3 @@
   )
 
 
-(comment 
-
-  (def tribuo-linear-sdg
-    (ml/train
-     diabetes
-     {:model-type :scicloj.ml.tribuo/regression
-      :tribuo-components [{:name "squared"
-                           :type "org.tribuo.regression.sgd.objectives.SquaredLoss"}
-                          {:name "trainer"
-                           :type "org.tribuo.regression.sgd.linear.LinearSGDTrainer"
-                           :properties  {:epochs "100"
-                                         :minibatchSize "1"
-                                         :objective "squared"}}]
-      :tribuo-trainer-name "trainer"}))
-  
-
-
-  (def prediction (ml/predict diabetes tribuo-linear-sdg))
-  
-
-
-;1. classical
-  (loss/mae (:disease-progression prediction) 
-            (:disease-progression (cf/target diabetes)))
-  ;;=> 54.83060340549487
-  
-
-
-;2. use standardized
-  (def standardised
-    ((:pre-metric-standarisation-fn
-      (ml/options->model-def {:model-type :scicloj.ml.tribuo/regression}))
-     prediction
-     diabetes :continous))
-  (loss/mae (:prediction standardised) (:trueth standardised))
-;;=> 54.83060340549487  
-  
-  
-  )
-
-

test/scicloj/ml/tribuo_test.clj

@@ -8,11 +8,9 @@
    [scicloj.ml.tribuo]
    [tablecloth.api :as tc]
    [tech.v3.dataset :as ds]
-   [tech.v3.dataset.categorical :as dscat]
-   [tech.v3.dataset.modelling :as ds-mod]
+   [tech.v3.dataset.categorical :as ds-cat]
    [tech.v3.dataset.column-filters :as cf]
-   [tech.v3.libs.tribuo :as tribuo]))
-
+   [tech.v3.dataset.modelling :as ds-mod]))
 
 (def iris-target-raw
   (->>
@@ -58,24 +56,19 @@
         model (ml/train (:train-ds split) options)
         predictions (->  (ml/predict (:test-ds split) model))
 
-        standardise-fn (:pre-metric-standarisation-fn (ml/options->model-def options))
-
-
 
         accuracy (loss/classification-accuracy (-> split :test-ds
-                                                   dscat/reverse-map-categorical-xforms
+                                                   ds-cat/reverse-map-categorical-xforms
                                                    :species)
                                                (-> predictions
-                                                   dscat/reverse-map-categorical-xforms
+                                                   ds-cat/reverse-map-categorical-xforms
                                                    :species))
-        score-fn (:score-fn (ml/options->model-def options))
-        score (score-fn model (:test-ds split))
-        standardised (standardise-fn predictions (:test-ds split) :discrete)
-        accuracy-from-standardised (loss/classification-accuracy
-                                    (-> standardised :prediction)
-                                    (-> standardised :trueth))]
-
-    (t/is (< expected-accuracy accuracy-from-standardised))
+        score (ml/score model 
+                        (cf/prediction predictions)
+                        (cf/target (:test-ds split))
+                        loss/classification-accuracy
+                        )]
+
     (t/is (< expected-accuracy score))
     (t/is (< expected-accuracy accuracy))
     (t/is (= expected-target-val
@@ -343,6 +336,7 @@
   (validate-target-symetry :float64))
 
 
+
 (t/deftest xxx
   (let [iris
         (make-iris-ds
@@ -355,24 +349,37 @@
          :int)
         split (ds-mod/train-test-split iris {:seed 123})
         options {:model-type :scicloj.ml.tribuo/classification
-                 :tribuo-components [{:name "trainer"
-                                      :type "org.tribuo.classification.dtree.CARTClassificationTrainer"}]
+                 :tribuo-components [{:name "nu"
+                                      :type "org.tribuo.classification.libsvm.SVMClassificationType"
+                                      :properties {:type "NU_SVC"}}
+                                     {:name "trainer"
+                                      :type "org.tribuo.classification.libsvm.LibSVMClassificationTrainer"
+                                      :properties {:svmType "nu"
+                                                   :probability "true"
+                                                   }}]
                  :tribuo-trainer-name "trainer"}
         model (ml/train (:train-ds split)
                         options)
 
-        p (ml/predict (:test-ds split) model)]
-    (t/is (= (-> p cf/prediction ds/column-names) [:species]))
-    (t/is (= (-> p cf/probability-distribution ds/column-names) ["virginica" "setosa" "versicolor"]))
+        prediction (ml/predict (:test-ds split) model)]
+    
+
+    (t/is (= 1.0
+             (ml/score model
+                       prediction
+                       (:test-ds split)
+                       loss/classification-accuracy)))
+
+
+    (t/is (= (-> prediction cf/prediction ds/column-names) [:species]))
+    (t/is (= (->> prediction cf/probability-distribution ds/column-names (into #{})) #{"virginica" "setosa" "versicolor"}))
 
     (t/is (.equals
            {:name :species,
             :datatype :int32,
             :n-elems 45,
             :column-type :prediction,
             :categorical-map {:lookup-table {:versicolor 0, :setosa 1, :virginica 2}, :src-column :species, :result-datatype :int}}
-           (-> p (cf/prediction) :species meta)))))
-
-
+           (-> prediction (cf/prediction) :species meta)))))