Arie, I now see that my comment regarding a Hook was irrelevant (if not just wrong) because the lha of the amend gerund supplies the m in m}. So to get different strides is just a matter of supplying a different set of array indices for the lha From. So, all that's needed for different strides is a separate verb which produces such array indices for the lha.
Thanks, again, On Wed, Feb 20, 2019 at 10:57 AM Brian Schott <[email protected]> wrote: > Aii, > > (Is it Ari?) > Yes, that is a nice extension that absolutely anticipates my expanded > needs. I played with increasing the dimensions of i and w a little and it > works beautifully. > > [snip] > > I've convinced myself that your code is mostly a Hook with a "dummy" lha > because notice in the next demonstration I have inserted a [ on the left to > get the same result as the immediately previous demonstration. Not shown is > the case where I instead inserted a ], so that's why I am calling it a > dummy lha. > > ;/($ $"1 ([(0 0;0 _1;_1 0 ;_1 _1){($ w),;._3 i. $ i)(,w)"_`[`] }"1 ,) > i=.4 4 $ Ga [ w=. 3 3 $ ka > ┌────┬────┬────┬────┐ > │abcP│Mabc│MNOP│MNOP│ > │defT│Qdef│abcT│Qabc│ > │ghiX│Ughi│defX│Udef│ > │YZAB│YZAB│ghiB│Yghi│ > └────┴────┴────┴────┘ > > Let me refer to the w as a weight filter and the i as an input field. This > terminology is somewhat compatible with the application I am studying: > convolutional neural nets. I am trying to work through some examples of > convnet's and am hoping such the construct being developed in this thread > can be used in convnets. > > So, my reason for exploring the dummy lefthand side is that I am thinking > that your script might be further expanded to accommodate more than just > the corner filter positions to positions defined by a compatible "stride" > which is like the movement vector in ;._3 (the 0{x in tesselation > discussion in the Dictionary). I don't need that at the moment, but it > would be considerably more general in the convnet context. > > > -- > (B=) > ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
