dChan
Anonymous ID: df3b57 March 4, 2019, 5:29 p.m. No.5510509   🗄️.is 🔗kun

I was gone for Q's recent short stringer comms posts. Can i get a TL;DR? Do we know what that was all about?

Anonymous ID: df3b57 March 4, 2019, 5:38 p.m. No.5510669   🗄️.is 🔗kun   >>0686

(defclass-now ->conv (lump)

((x :initarg :x :reader x)

(weights

:initform (->weight :dimensions '(1 1))

:type ->weight :initarg :weights :reader weights

:documentation "A ->WEIGHT lump.")

(transpose-weights-p

:initform nil :initarg :transpose-weights-p

:reader transpose-weights-p

:documentation "Determines whether the input is multiplied by

WEIGHTS or its transpose.")

(kernel-size

:initform nil

:initarg :kernel-size

:reader kernel-size)

(num-filters

:initform nil

:initarg :num-filters

:reader num-filters)

(stride

:initform nil

:initarg :stride

:reader stride)

(out-size

:initform 1

:accessor out-size))

(:documentation "Perform (CONV X WEIGHTS) where X (the input) is of

size M and WEIGHTS is a ->WEIGHT whose single stripe is taken to

be of dimensions NUM-FILTERS x KERNEL-SIZE"))

 

(defmaker (->conv :unkeyword-args (x weights)))

 

(defmethod initialize-instance :after ((lump ->conv) &key &allow-other-keys)

(setf (slot-value (weights lump) 'dimensions)

(if (transpose-weights-p lump)

(list (kernel-size lump)(num-filters lump))

(list (num-filters lump) (kernel-size lump))))

;; force reshaping

(setf (max-n-stripes (weights lump)) (max-n-stripes (weights lump)))

(setf (slot-value lump 'out-size) (1+ (floor (- (size (x lump)) 1) (stride lump)))))

 

(defmethod default-size ((lump ->conv))

;;default size is equal to the output of conv operation

(1+ (floor (- (size (x lump)) 1) (stride lump))))

 

(defmethod forward ((lump ->conv))

(let* ((x1 (x lump))

(x (nodes (x lump)))

(weights (nodes (weights lump)))

(y (nodes lump))

(stride (list 1 (stride lump)))

(start '(0 0))

(anchor '(0 0)))

;; FIXEXT:

(assert (stripedp x1))

(if (transpose-weights-p lump)

;; a = x *conv w'

(error "->conv not implemented for transposed weights")

;; a = conv(x, w)

(convolve! x weights y :start start :stride stride :anchor anchor))))

 

(defmethod set-input (instances (conv simple-conv))

(let ((input-nodes (nodes (find-clump 'input conv))))

(fill! 0 input-nodes)

(loop for instance in instances

for i upfrom 0

do

(loop for j upfrom 0

for digit in instance

do (setf (mref input-nodes i j) digit)))))

 

(defclass simple-conv (fnn) ())

 

(defun make-conv-fnn (&key(input-size 5)(num-filters 1)(kernel-size 3)(stride 2))

(build-fnn (:class 'simple-conv)

(input (->input :size input-size :name 'input))

(bias (->weight :size (1+ (floor (- input-size 1) stride)) :name 'bias))

(w (->weight :name 'filter :size num-filters))

(conv (->conv input w :kernel-size kernel-size :num-filters num-filters :stride stride))

(conv-act (->+ (list conv bias) :name 'conv-act))

(relu (->relu conv-act))))

Anonymous ID: df3b57 March 4, 2019, 5:41 p.m. No.5510738   🗄️.is 🔗kun

It seems to me these portions of the last crumbs go together:

 

"D Fire_Good

Location assist_Good"

 

"Did the fire in China cause GOOG to continue work from CA?"

Anonymous ID: df3b57 March 4, 2019, 5:55 p.m. No.5511035   🗄️.is 🔗kun   >>1182

https://github.com/melisgl/mgl/blob/master/doc/md/mgl-manual.md#x-28MGL-3A-40MGL-CODE-ORGANIZATION-20MGL-PAX-3ASECTION-29

 

"MGL is a Common Lisp machine learning library by Gábor Melis with some parts originally contributed by Ravenpack International. It mainly concentrates on various forms of neural networks (boltzmann machines, feed-forward and recurrent backprop nets). Most of MGL is built on top of MGL-MAT so it has BLAS and CUDA support.

 

In general, the focus is on power and performance not on ease of use. Perhaps one day there will be a cookie cutter interface with restricted functionality if a reasonable compromise is found between power and utility."