yamins81
commented
10 years ago Time breakdowns that I've observed are basically:
1) gpu computation itself:
for large models = < 2s per batch
for smaller models (e.g. Krizhevsky) = ~ .3s per batch2) loading the data:
(from written batches) > 10s per batch
with optimizations = < 1s per batch3) img cropping, when img_flip is True = ~ 1s per batch
So I think doing the optimizations is probably a good idea even with the larger models. On Sep 30, 2014 10:54 AM, "Diego Ardila" notifications@github.com wrote:
@yamins81 https://github.com/yamins81 have you checked how much the optimizations you've done for loading data onto the GPU have an effect when the network is larger (say, the NYU network?). I have a suspicion that these optimizations are less relevant when the GPU compute time is higher since in theory the CPU should be able to get the next batch ready while the GPU runs, so I was wondering if you've checked this.
— Reply to this email directly or view it on GitHub https://github.com/dicarlolab/archconvnets/issues/29.
ardila
@yamins81 have you checked how much the optimizations you've done for loading data onto the GPU have an effect when the network is larger (say, the NYU network?). I have a suspicion that these optimizations are less relevant when the GPU compute time is higher since in theory the CPU should be able to get the next batch ready while the GPU runs, so I was wondering if you've checked this.