Common Lispでゼロから作るディープラーニング （1）行列演算とニューラルネットのフォワード計算 ref: https://qiita.com/masatoi0/items/b3201d44fe13e94afc19

file0.txt

export PATH=/usr/local/cuda-8.0/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-8.0/lib64:$LD_LIBRARY_PATH
export C_INCLUDE_PATH=/usr/local/cuda-8.0/include:$C_INCLUDE_PATH

file1.txt

(ql:quickload :mgl-mat)

file10.txt

vb = alpha * va + vb

file11.txt

(defparameter va (make-mat '(3 1) :initial-contents '((1) (2) (3))))
(defparameter vb (make-mat '(3 1) :initial-contents '((10) (20) (30))))

(axpy! 1.0 va vb) ; => #<MAT 3x1 ABF #2A((11.0) (22.0) (33.0))>
vb                ; => #<MAT 3x1 ABF #2A((11.0) (22.0) (33.0))>

file12.txt

;; Sigmoid関数
(defun sigmoid! (v)
  (.logistic! v))

;; 3.4 3層ニューラルネットワークの実装

(defparameter x  (make-mat '(2 1) :initial-contents '((1.0) (0.5))))
(defparameter W1 (make-mat '(3 2) :initial-contents '((0.1 0.2) (0.3 0.4) (0.5 0.6))))
(defparameter b1 (make-mat '(3 1) :initial-contents '((0.1) (0.2) (0.3))))
(defparameter z1 (make-mat '(3 1) :initial-element 0.0))

(gemm! 1.0 W1 x 0.0 z1)
(axpy! 1.0 b1 z1) ; => #<MAT 3x1 AF #2A((0.3) (0.7) (1.1))>
(sigmoid! z1)     ; => #<MAT 3x1 ABF #2A((0.5744425) (0.66818774) (0.7502601))>

(defparameter W2 (make-mat '(2 3) :initial-contents '((0.1 0.2 0.3) (0.4 0.5 0.6))))
(defparameter b2 (make-mat '(2 1) :initial-contents '((0.1) (0.2))))
(defparameter z2 (make-mat '(2 1) :initial-element 0.0))

(gemm! 1.0 W2 z1 0.0 z2)
(axpy! 1.0 b2 z2)
(sigmoid! z2)

(defparameter W3 (make-mat '(2 2) :initial-contents '((0.1 0.2) (0.3 0.4))))
(defparameter b3 (make-mat '(2 1) :initial-contents '((0.1) (0.2))))
(defparameter z3 (make-mat '(2 1) :initial-element 0.0))

(gemm! 1.0 W3 z2 0.0 z3)
(axpy! 1.0 b3 z3) ; => #<MAT 2x1 AF #2A((0.3168271) (0.6962791))>

file13.txt

(time (progn
        (gemm! 1.0 W1 x 0.0 z1)
        (axpy! 1.0 b1 z1)
        (sigmoid! z1)
        (gemm! 1.0 W2 z1 0.0 z2)
        (axpy! 1.0 b2 z2)
        (sigmoid! z2)
        (gemm! 1.0 W3 z2 0.0 z3)
        (axpy! 1.0 b3 z3)))

;; Evaluation took:
;;   0.000 seconds of real time
;;   0.000000 seconds of total run time (0.000000 user, 0.000000 system)
;;   100.00% CPU
;;   691,968 processor cycles
;;   0 bytes consed

file2.txt

ros install cl-cuda melisgl/mgl-mat

file3.txt

(defpackage cl-zerodl
  (:use :cl :mgl-mat)
  (:nicknames :zerodl))

(in-package :cl-zerodl)

file4.txt

(setf *default-mat-ctype* :float)
(setf *cuda-enabled* t)

file5.txt

(defparameter ma (make-mat '(2 2) :initial-contents '((1 -2) (-3 4))))
(defparameter mb (make-mat '(2 2) :initial-contents '((5 6) (7 8))))
(defparameter mc (make-mat '(2 2) :initial-element 0.0))

ma ; => #<MAT 2x2 AB #2A((1.0 -2.0) (-3.0 4.0))>
mb ; => #<MAT 2x2 BF #2A((5.0 6.0) (7.0 8.0))>
mc ; => #<MAT 2x2 - #2A((0.0 0.0) (0.0 0.0))>

file6.txt

mc = alpha * ma * mb + beta * mc

file7.txt

(gemm! 1.0 ma mb 0.0 mc)
; => #<MAT 2x2 F #2A((-9.0 -10.0) (13.0 14.0))>

mc ; => #<MAT 2x2 AF #2A((-9.0 -10.0) (13.0 14.0))>

file8.txt

(gemm! 1.0 ma mb 0.0 mb)

file9.txt

;; 巨大な配列を表示してもREPLが落ちないように表示範囲を制限
(setf *print-length* 10
      *print-level* 10)

(defparameter ma (make-mat '(10000 10000)))
(defparameter mb (make-mat '(10000 10000)))
(defparameter mc (make-mat '(10000 10000)))

;; 一様乱数で初期化
(uniform-random! ma)
(uniform-random! mb)

;; OpenBLAS
(time (gemm! 1.0 ma mb 0.0 mc))

;; Evaluation took:
;;   6.539 seconds of real time
;;   26.092000 seconds of total run time (25.744000 user, 0.348000 system)
;;   399.02% CPU
;;   22,180,377,236 processor cycles
;;   0 bytes consed

;; CUBLAS
(with-cuda* ()
  (time (gemm! 1.0 ma mb 0.0 mc)))

;; Evaluation took:
;;   0.427 seconds of real time
;;   0.424000 seconds of total run time (0.424000 user, 0.000000 system)
;;   99.30% CPU
;;   1,447,343,752 processor cycles
;;   0 bytes consed