Links
One 60x60 sample is processed by a 2x2 group of processors. Given that Parallella can have 4x4 or 8x8 grids, this allows 4 or 16 groups respectively work independently either on different samples, or on the same sample rotated (but that's for another post).
Now by layers:
Layer 0: each processor has 30x30 chunk of the image.
Layer 1: 16 5x5 filters are applied + 2x2 max-polling: output 14x14x16.
Layer 2: 32 5x5x16 filters + 2x2 max-polling: output 6x6x32.
Layer 3: 48 3x3x32 filters are applied + 2x2 max-polling: output 3x3x48 with overlap. The size of the full matrix at this stage is 5x5x48.
Layer 4: 64 3x3x48 convolutions. In the paper, this is unshared convolution, so that at each spot the weights are different. I am still not sure whether I will do it unshared, or replace by a regular convolution. TECHNICALLY the amount co computation on each core is unchanged, but the amount of data that has to be copied back and forth between ARM and Epiphany is ginormous... So regular convolution for now.
Also, since this layer a) doesn't have max-polling, b) is followed by fully-connected layer and c) the input is already so small in spatial dimensions, the partitioning of the output is a bit different now. Instead of each core applying the same filter to its chunk of data, it will now apply DIFFERENT filters to the FULL matrix. So, in the output each core will have a 3x3x16 chunk of a full 3x3x64 matrix.
Given that the output of every layer has to be preserved for the back-propagation, we have:
(30x30x1 + 14x14x16 + 6x6x32 + 3x3x48 + 3x3x16)*(4 bytes) = 22.5 kb
Out of 32 :(
Hope I have enough memory to do backpropagation without swapping out stuff :(
[ 1 comment ] ( 1854 views ) | permalink |
( 2.8 / 224 )