anokland / dfa-torch Goto Github PK

How do I get the MNIST data in the format that can be used by the code, i.e. the mnist-train.t7 and mnist-test.t7 files? The script from https://github.com/soumith/cifar.torch.git only retrieves CIFAR data.

Thank you for providing good simulation code.

But When I run DFA with relu activation function it has very low accuracy.

Is it the problem of code? or the limits of algorithm DFA?

Hi, I think some updates in the unsup package broke the whitening procedure in your code. The problem should be the zca_whiten function.

The code gets stuck for an indefinite amount of time there. Then, at some point:

/home/iacolippo/torch/install/bin/luajit: /home/iacolippo/torch/install/share/lua/5.1/trepl/init.lua:389: 
Lapack Error syev : 3070 off-diagonal elements didn't converge to zero at 
/home/iacolippo/torch/pkg/torch/lib/TH/generic/THTensorLapack.c:384
stack traceback:
	[C]: in function 'error'
	/home/iacolippo/torch/install/share/lua/5.1/trepl/init.lua:389: in function 'require'
	Main.lua:83: in main chunk
	[C]: in function 'dofile'
	...ippo/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk
	[C]: at 0x00405d50

Tried with th Main.lua -dataset Cifar10 -network conv.lua -LR 5e-5 -whiten

I suppose feedback alignment and similarity feedback could be related:
https://discourse.numenta.org/t/similarity-alignment-a-missing-link-between-structure-function-and-algorithms/3683/4
And that would be the end of the mystery.

To simulate FA, I wrote following command,
"th Main.lua -dataset Cifar10 -network conv.lua -LR 9e-6 -whiten -gradient fa -epoch 101"

But It stopped working because of the error. (THNN.lua:110: bad argument #8 tp 'v')

Do you know why does it not work?