• Content count

  • Joined

  • Last visited

Community Reputation

898 Good

About Prozak

  • Rank
  1. There is no specific reasons for doing it from scratch for company reasons, except that if I have the knowledge to do it from scratch, then I might adapt it into lower tier hardware we have lying around. None the less we already got a couple of 1080s on order.   Thank you for your site, do you know of any books that might help here?   Again, thank you for your time.
  2. Hi guys,   could someone point me to a book resource, or similar, regarding the implementation of CNNs for identification of features in images, such as letters, and then more complex features, like animals?   I would like to understand the underlying theory, and then do a CUDA implementation. Yes, I know there are libraries out there that already do this quite efficiently, but I would like to do this from scratch for academic, and then internal company reasons.   Thanks everyone and anyone for any help on this.
  3. Ugly StretchBlt

    Thx ApochPiQ,   I've done exactly that, capture a 400x400 area and reduce to 100x100, but the results are simply not acceptable.   What alternative are there outside of the outdated win32 gdi api?   Thx!
  4. Hi all,   just coded a small demo that captures the desktop window, and then tries to scale it down and draw it to my dialog' hDC.   Problem is, I'm using StretchBlt, and the final results are quite ugly. I've read several stuff that instructs us to use SetStretchBltMode and SetBrushOrgEx. The end result just isn't up to scratch.   Is there some solution, even if it involves GDI+/ OpenGL/ DirectX/ DirecDraw that can draw into my dialog and resize my original bitmap perfectly? Also, I'm looking for a somewhat fast solution. I wouldn't want this to take more than 1sec to draw...   Thanks for any tips on this. Have a great week.
  5. Backpropagation XOR issue

    Hmm, even tough I've validated my BP algorithm with JavaNNS, a neural net doesn't always achieve a result when training XOR. Is this a documented symptom? If I place the same value on all weights, it doesn't seem to reach a solution. Seemingly, I would venture, the net enjoys more success when filled with random weights that favor a certain result. I'm off for holidays, but I will post more results here when I have them. Thanks all, and to all happy holidays with your families.
  6. Backpropagation XOR issue

    I think I found something wrong with my code, I'm using JavaNNS to validate it right now, if I still have issues, I'll post them here soon.   For now thank you for your patience and assistance. 
  7. Backpropagation XOR issue

    Sure, ok, I think I can do that. Gimme a day and I'll post the results... brb....
  8. Backpropagation XOR issue

    So, you believe I should run the back-propagation always, period. Ok, I will change the code to do that.   Do you think sharing the weights after a run, and after, say, 100 runs, here, would be useful?   Thank you!
  9. Backpropagation XOR issue

    Sorry for that lapse.   I initialize the weights with a random number between -1 and +1.   All transfer functions are simple sigmoid. I have two training methods: * Linear - go through each example, if error above max_allowed_error, then backpropagate * Stochastic - select a random training example, if error above max_allowed_error, then backpropagate   My current max_allowed_error is 0.1   Btw, is there any online resource where I can copy an initial topology and weight setup and cycle through 1.000 iterations (for example), and then check the resulting weights, a sort of Roseta Stone to make sure the math behind my implementation of the learning algorithm is valid and correctly implemented?   Thank you.
  10. Backpropagation XOR issue

    This specific class is a 2 hidden layer class.   So the topology I use is 2-2-1-1 (2 inputs, 1 output, 2 layer 1 hidden neurons, 1 layer 2 hidden neuron).   It might reach solution in as little as 5000 iterations, using a constant learning rate of 1.0, or not reach it at all after 50 million iterations. I'm using doubles. Also, after inspection, some values in the weights seem to grow a lot, or be extremely small (near zero).   Thank you for any inputs you might have.
  11. Backpropagation XOR issue

    I placed this question here, because, making sure that I start with the same network topology, the only thing changing being a randomization of the starting weights and bias, sometimes the network will achieve solution, and other times not (by that I mean multiple million iterations not reaching solution).   So, in an example where the learning rate remained untouched throughout, one network achieves solution, and another does not, due to starting weights conditions.   Could that be symptomatic of a bug, or is this something that sometimes just happens due to initial starting conditions? (thank you for your replies!)
  12. Hi all,   would just like to know if, for a given problem, in this case the classic XOR binary problem,  the network will eventually fluctuate to the solution, if given infinite time to do so, or, if due to learning rate issues, it can get stuck in over-shooting behaviors.   If the learning rate should pose no problem, can it be mathematically said that the network should travel towards solution-space?   Thank you.
  13. Hi all,   has anyone here worked on something that would render astral bodies, like stars, constellations, galaxies, planets, taking into account the latitude and longitude of the user?   I'm tasked with doing something like this, in a browser. If there was some REST API I could consume and then render the horizon and the astral bodies, that would be awesome, apart from that I would require the actual equations for calculating their position above or bellow the horizon...   Thanks for any tips on this, Hugo Ferreira
  14. Hi, I'm coding a simple webGL engine, and would like to import stereographic panoramas (taken from an iPad) into the engine as cubemaps.   Is there any free utility i can download that would accomplish this?   Thank you in advance!