search

hughperkins / DeepCL

OpenCL library to train deep convolutional neural networks
Mozilla Public License 2.0
865 stars 199 forks source link

Is it possible to use a rectangular matrix as input #81

Closed NKUCodingCat closed 8 years ago

NKUCodingCat commented 8 years ago

I read NeuralNet.pyx and it seems that I can use square image only, so is it possible to use an rectangular image ( like an image in 100*25 ) as input?

Meanwhile, what does netdef string like rt2-8c5z-relu-mp2-16c5z-relu-mp3-150n-tanh-10n means? I worked out that mp is for pooling layer, 2 is the size and tanh is an activation layer.....and so on. However, it is hard to remember each abbreviation, so are there any document or memorandum about it, or they are just an convention I don't know?

Anyway, thank you for answering my question.

NKUCodingCat commented 8 years ago

Well I found the meanings in Commandline.md...... How can I miss this doc=.=

hughperkins commented 8 years ago

As far as rectangular inputs ... you'd need to pad them to be square, for now. There is a convolutional kernel implemented that handles rectangular inputs though, so it's not out of the bounds of possibility to modify deepcl to handle them, but I think joining a few images together, padded, into a single square image, might be quite a lot easier, for now?

NKUCodingCat commented 8 years ago

Well, padding to a square sounds reasonable ... but it may let performance down, whatever, I gonna to try both of it, Thank you for your help.

hughperkins commented 8 years ago

Ok :-)