[theano-users] Grouped convolution not behaving as expected

[Michael Oliver]

with newest Theano, libgpuarray, cudnn 6, the following code produces
unexpected results:

x = np.random.randn(100,10,60,60)

k1 = np.random.randn(10,1,5,5)

k2 = np.concatenate([k,k], axis=0)

o1 = T.nnet.conv2d(x,k, num_groups=10)

o1 = o1.eval()

o2 = T.nnet.conv2d(x,k2, num_groups=10)

o2 = o2.eval()

o1[0,:,0,0]
array([ 3.22360683, -1.20146486, -0.47498998, 0.24816176, -5.08138128,
-6.36815384, -4.16210719, -9.07320647, 6.09663779, 4.29091563])

o2[0,:,0,0]
array([ 3.22360683, 2.03843043, -0.65284486, -3.08015858,
2.71291378, -5.86320533, 1.53297356, 3.75655768,
-1.23243404, 4.76971381, -2.61702975, 7.87672154,
-5.74371147, 8.65516924, 10.45707448, 0.67422908,
8.08504753, -0.83615408, -7.9871845 , 4.29091563])

Should not o2 be just o1 repeated twice? Given that the first and last
values match, but nothing else does, unless I'm misunderstanding something
it seems like some kind of bug. Thanks!

[Michael Oliver]

Sorry these places where its just k should be k1:

k2 = np.concatenate([k1,k1], axis=0)

o1 = T.nnet.conv2d(x,k1, num_groups=10)

Post by Michael Oliver
with newest Theano, libgpuarray, cudnn 6, the following code produces
x = np.random.randn(100,10,60,60)
k1 = np.random.randn(10,1,5,5)
k2 = np.concatenate([k,k], axis=0)
o1 = T.nnet.conv2d(x,k, num_groups=10)
o1 = o1.eval()
o2 = T.nnet.conv2d(x,k2, num_groups=10)
o2 = o2.eval()
o1[0,:,0,0]
array([ 3.22360683, -1.20146486, -0.47498998, 0.24816176, -5.08138128,
-6.36815384, -4.16210719, -9.07320647, 6.09663779, 4.29091563])
o2[0,:,0,0]
array([ 3.22360683, 2.03843043, -0.65284486, -3.08015858,
2.71291378, -5.86320533, 1.53297356, 3.75655768,
-1.23243404, 4.76971381, -2.61702975, 7.87672154,
-5.74371147, 8.65516924, 10.45707448, 0.67422908,
8.08504753, -0.83615408, -7.9871845 , 4.29091563])
Should not o2 be just o1 repeated twice? Given that the first and last
values match, but nothing else does, unless I'm misunderstanding something
it seems like some kind of bug. Thanks!

[Michael Oliver]

Nevermind: https://github.com/Theano/Theano/issues/6297

Post by Michael Oliver
k2 = np.concatenate([k1,k1], axis=0)
o1 = T.nnet.conv2d(x,k1, num_groups=10)

Post by Michael Oliver
with newest Theano, libgpuarray, cudnn 6, the following code produces
x = np.random.randn(100,10,60,60)
k1 = np.random.randn(10,1,5,5)
k2 = np.concatenate([k,k], axis=0)
o1 = T.nnet.conv2d(x,k, num_groups=10)
o1 = o1.eval()
o2 = T.nnet.conv2d(x,k2, num_groups=10)
o2 = o2.eval()
o1[0,:,0,0]
array([ 3.22360683, -1.20146486, -0.47498998, 0.24816176, -5.08138128,
-6.36815384, -4.16210719, -9.07320647, 6.09663779, 4.29091563])
o2[0,:,0,0]
array([ 3.22360683, 2.03843043, -0.65284486, -3.08015858,
2.71291378, -5.86320533, 1.53297356, 3.75655768,
-1.23243404, 4.76971381, -2.61702975, 7.87672154,
-5.74371147, 8.65516924, 10.45707448, 0.67422908,
8.08504753, -0.83615408, -7.9871845 , 4.29091563])
Should not o2 be just o1 repeated twice? Given that the first and last
values match, but nothing else does, unless I'm misunderstanding something
it seems like some kind of bug. Thanks!