;; test the (fast) jblas implementation on mnist
user> (require '[boltzmann.mnist :as mnist]
           '[boltzmann.visualize :as v]
           '[boltzmann.core :refer :all]
           '[boltzmann.jblas :refer :all]
           '[clojure.core.matrix :refer [matrix] :as mat])
nil
;; ensure we are using clatrix (jblas)
user> (mat/current-implementation)
:clatrix
;; load the images, this can take up to a few minutes, so you can use a future
user> (future (def images (mnist/read-images "resources/train-images-idx3-ubyte"))
          (def batches (doall (map matrix (partition 10 images)))))
#<core$future_call$reify__6320@4c5681f1: :pending>
;; read label data (fast)
user> (def labels (mnist/read-labels "resources/train-labels-idx1-ubyte"))
#'user/labels
;; unify images and labels in one data vector
user> (def labeled-images
    (map #(float-array (concat %1 %2)) (map bin-label labels) images))
#'user/labeled-images
;; we use mini-batches of size 100 for a speedup
user> (def labeled-batches
    (doall (map matrix (partition 100 labeled-images))))
#'user/labeled-batches
;; train with 100 hidden units
user> (def trained
    (let [rbm (create-jblas-rbm 794 100)]
      (time (train-cd rbm labeled-batches :epochs 4 :init-learning-rate 0.01 :k 1))))
#'user/trained
;; build test-images with zeroed labels
user> (def test-images (map (comp float-array concat) (repeat (repeat 10 0)) images))
#'user/test-images
;; take a peek at the reconstructions of the first 100 images
user> (->> test-images
       (take 100)
       (map (comp matrix vector))
       (map #(->> %
                  (probs-hs-given-vs [(:restricted-weights trained)
                                      (:h-biases trained)])
                  (probs-vs-given-hs [(:restricted-weights trained)
                                      (:v-biases trained)])))
       (map (partial drop 10))
       (map #(partition 28 %))
       (v/tile 10)
       v/render-grayscale-float-matrix
       v/view)
#<JFrame javax.swing.JFrame[frame0,10,62,280x280,layout=java.awt.BorderLayout,title=MNIST Digit,normal,defaultCloseOperation=HIDE_ON_CLOSE,rootPane=javax.swing.JRootPane[,0,0,280x280,layout=javax.swing.JRootPane$RootLayout,alignmentX=0.0,alignmentY=0.0,border=,flags=16777673,maximumSize=,minimumSize=,preferredSize=],rootPaneCheckingEnabled=true]>
;; classifier the first 10% of training data
user> (def classified
    (->> test-images
         (take 6000)
         (map (comp matrix vector))
         (map #(->> %
                    (probs-hs-given-vs [(:restricted-weights trained)
                                        (:h-biases trained)])
                    (probs-vs-given-hs [(:restricted-weights trained)
                                        (:v-biases trained)])))
         (map (partial take 10))
         (map max-index)))
#'user/classified
;; check how good we model our training data (validation data gives better approximation)
user> (float (classification-rate labels classified))
0.697
;; 70% of the labels with the highest probability are correct
;; now let's load a pretrained rbm with 500 hidden units, trained for 40 epochs with CD-1
user> (require '[clojure.edn :as edn])
nil
user> (def trained
    (edn/read-string {:readers {'boltzmann.jblas/JBlasRBM load-jblas-rbm}}
                     (slurp "resources/mnist-labeled.edn")))
#'user/trained
;; the reconstructions are much better and fairly sharp
user> (->> test-images
       (take 100)
       (map (comp matrix vector))
       (map #(->> %
                  (probs-hs-given-vs [(:restricted-weights trained)
                                      (:h-biases trained)])
                  (probs-vs-given-hs [(:restricted-weights trained)
                                      (:v-biases trained)])))
       (map (partial drop 10))
       (map #(partition 28 %))
       (v/tile 10)
       v/render-grayscale-float-matrix
       v/view)
#<JFrame javax.swing.JFrame[frame2,0,24,280x280,layout=java.awt.BorderLayout,title=MNIST Digit,normal,defaultCloseOperation=HIDE_ON_CLOSE,rootPane=javax.swing.JRootPane[,0,0,280x280,layout=javax.swing.JRootPane$RootLayout,alignmentX=0.0,alignmentY=0.0,border=,flags=16777673,maximumSize=,minimumSize=,preferredSize=],rootPaneCheckingEnabled=true]>
;; reclassify
user> (def classified
    (->> test-images
         (take 6000)
         (map (comp matrix vector))
         (map #(->> %
                    (probs-hs-given-vs [(:restricted-weights trained)
                                        (:h-biases trained)])
                    (probs-vs-given-hs [(:restricted-weights trained)
                                        (:v-biases trained)])))
         (map (partial take 10))
         (map max-index)))
#'user/classified
;; now we get 93.6% of the labels right
user> (float (classification-rate labels classified))
0.93583333