Merge branch 'main' of github.com:scicloj/noj

scicloj · Dec 23, 2024 · 624c522 · 624c522
2 parents 3388ec3 + 354127b
commit 624c522
Show file tree

Hide file tree

Showing 3 changed files with 59 additions and 45 deletions.
diff --git a/deps.edn b/deps.edn
@@ -13,9 +13,6 @@
         org.scicloj/scicloj.ml.xgboost            {:mvn/version "6.2.0"}
 
         org.scicloj/scicloj.ml.tribuo             {:mvn/version "0.1.4"}
-        techascent/tech.ml.dataset                {:mvn/version "7.032"}   ;; no JVM crash
-        ;;techascent/tech.ml.dataset                {:mvn/version "7.033"}  ;; JVM crash
-        ;;techascent/tech.ml.dataset                {:mvn/version "7.034"} ;; JVM crash
 
         org.tribuo/tribuo-regression-liblinear    {:mvn/version "4.3.1"}
         org.tribuo/tribuo-regression-libsvm       {:mvn/version "4.3.1"}
@@ -30,7 +27,7 @@
         org.tribuo/tribuo-classification-xgboost   {:mvn/version "4.3.1"}
 
 
-        clj-python/libpython-clj                  {:mvn/version "2.025"}
+        clj-python/libpython-clj                  {:mvn/version "2.026"}
         org.scicloj/kind-pyplot                   {:mvn/version "1-beta2.1"}
         scicloj/clojisr                           {:mvn/version "1.0.0"}
 

diff --git a/notebooks/noj_book/automl.clj b/notebooks/noj_book/automl.clj
@@ -9,8 +9,12 @@
 (ns noj-book.automl
   (:require [noj-book.ml-basic :as ml-basic]
             [scicloj.kindly.v4.kind :as kind]
-            [scicloj.kindly.v4.api :as kindly]
-            [scicloj.metamorph.ml :as ml]))
+            [scicloj.metamorph.ml :as ml]
+            [tablecloth.api :as tc]
+            [scicloj.metamorph.ml.loss :as loss]
+            [scicloj.metamorph.core :as mm]
+            [scicloj.metamorph.ml.gridsearch :as gs]
+            [tech.v3.dataset.modelling :as ds-mod]))
 
 ;; ## The metamorph pipeline abstraction
 ;; When doing automl, it is very useful to be able to manage
@@ -278,6 +282,7 @@ ctx-after-train
          '[scicloj.metamorph.core :as mm]
          '[scicloj.ml.tribuo]
          '[scicloj.ml.xgboost]
+         '[scicloj.ml.smile.classification]
          '[scicloj.sklearn-clj.ml])
 
 
@@ -289,7 +294,7 @@ ctx-after-train
 
 ;;  the following will find the best model across:
 ;;
-;;  * 4 different model classes
+;;  * 4 different model classes with different hyper params
 ;;
 ;;  * 6 different selections of used features
 ;;
@@ -308,29 +313,47 @@ ctx-after-train
 (def titanic-k-fold (tc/split->seq ml-basic/numeric-titanic-data :kfold {:seed 12345}))
 
 (-> titanic-k-fold count)
+
+;; We add as well 10 hyper-parameter variants for logistic regression
+;; obtained via Sobol search over the hyper parameter space of the model.
+(def hyper-params 
+  (->>
+   (ml/hyperparameters :smile.classification/logistic-regression)
+   (gs/sobol-gridsearch)
+   (take 10)))
+hyper-params
+
+(def logistic-regression-specs
+  (map
+   #(assoc %
+           :model-type :smile.classification/logistic-regression)
+   hyper-params))
+logistic-regression-specs
+
+
 ;; The list of the model types we want to try:
-(def models [{ :model-type :xgboost/classification
-               :round 10}
-             {:model-type :sklearn.classification/decision-tree-classifier}
-             {:model-type :sklearn.classification/logistic-regression}
-             {:model-type :sklearn.classification/random-forest-classifier}
-             {:model-type :metamorph.ml/dummy-classifier}
-             {:model-type :scicloj.ml.tribuo/classification
-              :tribuo-components [{:name "logistic"
-                                   :type "org.tribuo.classification.sgd.linear.LinearSGDTrainer"}]
-              :tribuo-trainer-name "logistic"}
-             {:model-type :scicloj.ml.tribuo/classification
-              :tribuo-components [{:name "random-forest"
-                                   :type "org.tribuo.classification.dtree.CARTClassificationTrainer"
-                                   :properties {:maxDepth "8"
-                                                :useRandomSplitPoints "false"
-                                                :fractionFeaturesInSplit "0.5"}}]
-              :tribuo-trainer-name "random-forest"}])
-
-
-;;  This uses models from Smile and Tribuo, but could be any
-;;  metamorph.ml compliant model ( library  `sklearn-clj` wraps all python sklearn
-;;  models, for example)
+(def models-specs 
+  (concat logistic-regression-specs
+          [{:model-type :xgboost/classification :round 10}
+           {:model-type :sklearn.classification/decision-tree-classifier}
+           {:model-type :sklearn.classification/logistic-regression}
+           {:model-type :sklearn.classification/random-forest-classifier}
+           {:model-type :metamorph.ml/dummy-classifier}
+           {:model-type :scicloj.ml.tribuo/classification
+            :tribuo-components [{:name "logistic"
+                                 :type "org.tribuo.classification.sgd.linear.LinearSGDTrainer"}]
+            :tribuo-trainer-name "logistic"}
+           {:model-type :scicloj.ml.tribuo/classification
+            :tribuo-components [{:name "random-forest"
+                                 :type "org.tribuo.classification.dtree.CARTClassificationTrainer"
+                                 :properties {:maxDepth "8"
+                                              :useRandomSplitPoints "false"
+                                              :fractionFeaturesInSplit "0.5"}}]
+            :tribuo-trainer-name "random-forest"}]))
+
+
+;;  This uses models from Smile, Tribuo and sklearn but could be any
+;;  metamorph.ml compliant model 
 
 ;;  The list of feature combinations to try for each model:
 (def feature-combinations
@@ -341,11 +364,11 @@ ctx-after-train
    [:sex :embarked]
    [:sex :pclass]])
 
-;; generate 42 pipeline functions:
+;; generate 102 pipeline functions:
 (def pipe-fns
-  (for [model models
+  (for [model-spec models-specs
         feature-combination feature-combinations]
-    (make-pipe-fn model feature-combination)))
+    (make-pipe-fn model-spec feature-combination)))
 
 (count pipe-fns)
 ;; Execute all pipelines for all splits in the  cross-validations
@@ -383,7 +406,7 @@ ctx-after-train
 
 
 ;; In total it creates and evaluates
-;; 7 models * 6 feature configurations * 5 CV = 210 models
+;; 17 models (incl. hyper parameters variations) * 6 feature configurations * 5 CV = 510 models
 (->  evaluation-results-all flatten count)
 
 ;;  We can find the best as well by hand, it's the first from the list,
@@ -392,7 +415,7 @@ ctx-after-train
     (tc/unique-by)
     (tc/order-by [:mean-accuracy] :desc)
     (tc/head 20)
-    (kind/table))
+    (kind/dataset))
 
 
 ;; ## Best practices for data transformation steps in or outside pipeline

diff --git a/notebooks/noj_book/smile_regression.clj b/notebooks/noj_book/smile_regression.clj
@@ -1,19 +1,13 @@
 ;; # Smile regression models reference - DRAFT 🛠
 
+^:kindly/hide-code
+(require '[scicloj.ml.smile.regression]
+         '[scicloj.ml.tribuo])
+
 (ns noj-book.smile-regression 
   (:require
    [noj-book.utils.render-tools :refer [render-key-info]]
-   [scicloj.kindly.v4.kind :as kind]
-   [scicloj.metamorph.core :as mm]
-   [scicloj.metamorph.ml :as ml]
-   [scicloj.metamorph.ml.toydata :as datasets]
-   [tablecloth.api :as tc]
-   [tech.v3.dataset :as ds]
-   [tech.v3.dataset.metamorph :as ds-mm]
-   [tech.v3.datatype.functional :as dtf]))
-
-^:kindly/hide-code
-(require '[scicloj.ml.smile.regression])
+   [scicloj.ml.tribuo]))
 
 
 ^:kindly/hide-code