Problem in the Deformable conv speed

[ywangeq]

I am not sure, but i think the implement of the deformable conv is slower than offical version when i use it in the network

[DEKHTIARJonathan]

Can you give some reproducible code?

It is hardly possible to do anything with your statement.

[ywangeq]

I don't find a good implement based on tensorflow. And I check the speed in the MXnet(the offical version). It is similar to the speed which is told in the paper. But when i use the deformable conv in the tensorlayer with my current work. My speed decrease from 11ms to 1s. Now i cannot give my code which related some works.
I am quite sorry. If i found some good implement, i will remind you!
Thank for you work

[DEKHTIARJonathan]

Im sorry, we can't fix an issue if you are not able to provide a simple dummy example.

The issue may be in the TF lib, now t necesseraly in TL.

I'm closing this issue.

[DEKHTIARJonathan]

If you are able to provide a simple example without any business logic @ywangeq just a few layers and dummy random input, I'll be glad to give it a look @ywangeq

[WildChlamydia]

@DEKHTIARJonathan
I confirm that Deformable conv, for example:

offset1 = Conv2d(conv5, 18, (3, 3), (1, 1), act=None, padding='SAME'')
conv5 = DeformableConv2d(conv5, offset1, int(512 * filter_size_scale), (3, 3), act=tf.nn.leaky_relu, W_init=w_init, b_init=None'

is about 3+ times slower than simple conv:

conv5 = Conv2d(pool4, 512 * filter_size_scale, (3, 3), act=tf.nn.leaky_relu, W_init=w_init)

[DEKHTIARJonathan]

@undeadblow I reopen the issue. Please provide a simple working example to show your point.

Just a 1 layer network and make it comparable with a non deformable conv. I can't test your assumption with your code.

@zsdonghao and me, we'll have a look at this as soon as you provide some working example showing your point.

[WildChlamydia]

@DEKHTIARJonathan
arch.py.zip
No problem, here is not 1 layer, but simple Unet with the only difference - 1 layer in the end of encoder are conv and deform conv respectively (conv5 layer). Just one layer!
I got such results with that code:
Sunet time is 0.1247713565826416 seconds <-- plain convolution layers
Dunet time is 2.2409575939178467 seconds <-- one deform convolution layer in the middle
It's about 20 times slower.
Original paper states that there should be not very significant slowdown. Definitely not in 20 times.

It's very critical, but for now I can't understand why it is so slow.

[jianlong-yuan]

Any News?

[DEKHTIARJonathan]

@zsdonghao @lgarithm