LispNet/network.lisp at main · Woodi-dev/LispNet · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167

(in-package :common-lisp-user)
(defpackage #:lispnet.network
  (:use
   #:common-lisp
   #:petalisp
   #:lispnet.layers
   #:lispnet.initializers
   #:lispnet.trainable-parameter
   #:lispnet.loss
   #:lispnet.optimizer)
  (:export
   #:train-test
   #:predict
   #:fit
   #:network-weights
   #:network-weights-size))

(in-package #:lispnet.network)

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;;
;;; Training - Validation


(defun train-test (network output-training-data
                   &rest training-data-plist
                   &key loss optimizer (mode "train") &allow-other-keys)
  (let* ((trainable-parameters
           (remove-if-not #'trainable-parameter-p (network-parameters network)))
         (output-parameters
           (loop for output in (network-outputs network)
                 collect (make-instance 'parameter
                           :element-type (element-type output)
                           :shape (array-shape output))))
         (lossfunc  (loop for output in (network-outputs network)
                          for output-parameter in output-parameters
                          collect
                          (funcall loss output-parameter output )))
         (loss-network
           (apply #'make-network lossfunc))
         (loss-network-gradient (list (lazy-reshape 1.0 (~))))
         (gradient (differentiator (network-outputs loss-network) loss-network-gradient))


         (training-network
           (apply #'make-network
                  (nconc (loop for trainable-parameter in trainable-parameters
                               collect
                               (funcall gradient trainable-parameter)) lossfunc)))
         (n nil)
         (losses '())
         (gradients (loop for i below (list-length trainable-parameters) collect (lazy-reshape 0.0 (~)))))
    ;; Determine the training data size.
    (dolist (data output-training-data)
      (if (null n)
          (setf n (range-size (first (shape-ranges (array-shape data)))))
          (assert (= n (range-size (first (shape-ranges (array-shape data))))))))
    (alexandria:doplist (parameter data training-data-plist) ()
      (unless (symbolp parameter)
        (assert (= n (range-size (first (shape-ranges (array-shape data))))))))
    ;; Iterate over the training data.
    (loop for index below n do
      ;; Assemble the arguments.
      (let ((args '()))
        ;; Inputs.
        (alexandria:doplist (parameter data training-data-plist)
          (unless (symbolp parameter)
            (push parameter args)
            ;; (push (lazy-slice data index) args)))
            (push (lazy-drop-axes (lazy-slices data (range index (+ index 1)) ) 0) args)))
        ;; Outputs.
        (loop for data in output-training-data
              for output-parameter in output-parameters do
                (push output-parameter args)
                (push  (lazy-drop-axes (lazy-slices data (range index (+ index 1)) ) 0) args))
        ;; Trainable parameters.
        (dolist (trainable-parameter trainable-parameters)
          (push trainable-parameter args)
          (push (trainable-parameter-value trainable-parameter) args))
        ;; Forward + backward pass
        (if (string-equal mode "train")
            ;;train
            (let* ((net-out-values
                     (first (multiple-value-list
                             (apply #'call-network training-network (reverse args)))))
                   (loss-out (compute(first(last net-out-values))))
                   (new-gradients (butlast net-out-values 1)))
              (loop for i below (list-length new-gradients) do
                (setf (nth i gradients) (lazy #'+ (nth i new-gradients)
                                              (nth i gradients))))
              (push loss-out losses))
            ;;test
            (let* ((net-out-values
                     (first (multiple-value-list
                             (apply #'call-network loss-network (reverse args)))))
                   (loss-out (compute(first net-out-values)))
                   )
              (push loss-out losses)))))
    (when (string-equal mode "train")
      ;;Average gradients
      (loop for i below (length gradients) do
        (setf (nth i gradients) (lazy #'/ (nth i gradients) n)))
      ;;Update weights
      (update-weights optimizer :weights trainable-parameters :gradients gradients))
    ;; Return the trained network and batch loss.
    (values network (/ (reduce #'+ losses) (length losses)) )))


(defun fit (network input train-input-data train-label-data val-input-data val-label-data &key (epochs 10) (batch-size 100) (loss #'mse) optimizer)
  (let ((train-input-data-length (array-dimension train-input-data 0))
        (train-label-data-length (array-dimension train-label-data 0))
        (val-input-data-length (array-dimension val-input-data 0))
        (val-label-data-length (array-dimension val-label-data 0)))
    (assert (= train-input-data-length train-label-data-length))
    (assert (= val-input-data-length val-label-data-length))
    (format t "Train on ~d samples~%" train-input-data-length)
    (loop for epoch from 1 to epochs do
      (format t "Epoch ~d/~d~%" epoch epochs)
      (let ((batch-train-losses '())
            (batch-val-losses '())
            (time-start (get-internal-real-time)))
        ;;Training
        (loop for offset below train-input-data-length by batch-size do
          (let* ((batch-range (range offset (min train-input-data-length (+ offset batch-size))))
                 (batch-data (lazy-slices train-input-data batch-range))
                 (batch-labels (lazy-slices train-label-data batch-range))
                 (batch-input (compute (lazy-collapse batch-data)))
                 (batch-output (compute (lazy-collapse batch-labels))))
            (multiple-value-bind (trained-net batch-loss)
                (train-test network (list batch-output)
                            :loss loss :optimizer optimizer :mode "train"
                            input batch-input)
              (setq batch-train-losses (append batch-train-losses (list batch-loss))))))
        ;;Validation
        (loop for offset below val-input-data-length by batch-size do
          (let* ((batch-range (range offset (min val-input-data-length (+ offset batch-size))))
                 (batch-data (lazy-slices val-input-data batch-range))
                 (batch-labels (lazy-slices val-label-data batch-range))
                 (batch-input (compute (lazy-collapse batch-data)))
                 (batch-output (compute (lazy-collapse batch-labels))))
            (multiple-value-bind (net batch-loss)
                (train-test network (list batch-output)
                            :loss loss :optimizer optimizer :mode "test"
                            input batch-input)
              (setq batch-val-losses (append  batch-val-losses (list batch-loss))))))
        (format t "~Ss train_loss: ~S - val_loss: ~S ~%"  (/ (- (get-internal-real-time) time-start) 1000.0)
                (/ (reduce #'+ batch-train-losses) (length batch-train-losses))
                (/ (reduce #'+ batch-val-losses) (length batch-val-losses))))))
  network)

;;;;; Prediction
(defun predict (network input)
  (values-list (apply #'call-network
                      network
                      (loop for parameter in (network-parameters network)
                            collect parameter
                            collect (if (trainable-parameter-p parameter)
                                        (trainable-parameter-value parameter)
                                        input)))))

(defun network-weights (network)
  (remove-if-not #'trainable-parameter-p (network-parameters network)))

(defun network-weights-size (network)
  (reduce #'+ (loop for weight in (network-weights network)
                    collect (shape-size(array-shape weight)))))