I think that nn.conv2d() will have the best improvement when the
inner-product dimension is large, i.e. when filter_x * filter_y *
n_channels is large. For 3x3 filters with 1 channel, it may just be too
small for the im2col/CorrMM version to show any improvement.
On Monday, October 17, 2016 at 10:30:13 PM UTC-7, Bogdan Opanchuk wrote:
>
> The difference in performance between nnet.conv2d() and nnet.conv.conv2d()
> seems to be about the same for 100x100 matrices.
>
> The profile is as follows:
>
> Apply
> ------
> <% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
> 88.0% 88.0% 18.591s 1.84e-01s 101 5 CorrMM{valid,
> (1, 1), (1, 1)}(InplaceDimShuffle{x,x,0,1}.0, Subtensor{::, ::, ::int64,
> ::int64}.0)
> 5.5% 93.5% 1.170s 1.16e-02s 101 11
> IncSubtensor{Set;int64:int64:, int64:int64:}(u, Reshape{2}.0, Constant{1},
> Constant{-1}, Constant{1}, Constant{-1})
> 4.8% 98.3% 1.022s 1.01e-02s 101 12
> Elemwise{Composite{sqr((i0 - i1))}}(IncSubtensor{Set;int64:int64:,
> int64:int64:}.0, u)
> 1.7% 100.0% 0.352s 3.49e-03s 101 13
> Sum{acc_dtype=float64}(Elemwise{Composite{sqr((i0 - i1))}}.0)
> ...
>
> Making `u` shared does not change the timings much, I expect it will be
> more important if I use the GPU backend.
>
--
---
You received this message because you are subscribed to the Google Groups
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
For more options, visit https://groups.google.com/d/optout.
