Large GPU RAM consumption

[winstywang]

Hi, all
I have tried both cuda-convnet and caffe on large imagenet model. It seems that Caffe will consume significantly larger RAM than cuda-convnet, which makes it impossible to train a even larger model (say using stride 2 in the first conv layer). Any idea on this issue?

[shelhamer]

Our standard convolution is computed by dense matrix multiplications by
unrolling image patches into columns. This design choice wins simplicity,
efficiency by GPU BLAS, and is general w.r.t. filter dimensions but
comes at a memory cost.

We are exploring alternative convolution strategies, so that our
convolution layers can be tuned to different regimes (such as small filter,
low stride for dense, early layers). The convolution layer parameter could
include a switch to select the back end, such as FFT or tuned explicit
convolution for common filter / input dimensionalities.

Expect a family of convolutions to appear this summer.

We welcome efforts to this end!

Le mercredi 28 mai 2014, winstywang notifications@github.com a écrit :

Hi, all
I have tried both cuda-convnet and caffe on large imagenet model. It seems
that Caffe will consume significantly larger RAM than cuda-convnet, which
makes it impossible to train a even larger model (say using stride 2 in the
first conv layer). Any idea on this issue?

—
Reply to this email directly or view it on GitHub
#463.

Evan Shelhamer

[sguada]

@winstywang follow #520 for memory reduction. I have been able to train bigger models, using stride 2 in conv1 in a Titan or K40.

[Coderx7]

@shelhamer : Caffe still takes much more memory compared to e.g torch, did you impelment what you intended to back in 2014?
Is there yet a way to reduce the memory consumption in caffe ? ( I have a GTX980 with 4GIg vram and yet still I'm not able to train alex net! )
Thanks for your attention in advance