cannot do matmul for high-dim tensors about singa HOT 7 CLOSED

Hi @joddiy , for tensor mul >2d, please use tensor.tensordot, it depends on which axis to perform the tensordot

    def test_4d_tensor_dot(self):
        for dev in [cpu_dev, gpu_dev]:
            x1 = np.random.randn(1, 12, 256, 64).astype(np.float32)
            x2 = np.random.randn(1, 12, 64, 256).astype(np.float32)
            x1 = tensor.from_numpy(x1)
            x1.to_device(dev)
            x2 = tensor.from_numpy(x2)
            x2.to_device(dev)

            y = tensor.tensordot(x1, x2, axes=([3],[2]))
            print(y.shape)

            y = tensor.tensordot(x1, x2, axes=([2,3],[3,2]))
            print(y.shape)

output:

(1, 12, 256, 1, 12, 256)
(1, 12, 1, 12)
(1, 12, 256, 1, 12, 256)
(1, 12, 1, 12)
...........................
----------------------------------------------------------------------
Ran 27 tests in 0.121s

OK

reference: https://docs.scipy.org/doc/numpy/reference/generated/numpy.tensordot.html

from singa.

Hi, shicong, according to the np's definition, the matmul for high-dim matrix should be:

x1 = np.random.randn(1, 12, 256, 64).astype(np.float32)
x2 = np.random.randn(1, 12, 64, 256).astype(np.float32)
print(np.matmul(x1, x2).shape)

Output:

(1, 12, 256, 256)

Actually, it does the matmul for (256, 64) * (64, 256) for all cells of the first two axes (1, 12).

from singa.

fixed in #639

from singa.

Hi, @dcslin , thanks for your help, this PR works fine for basic test cases. However, for some special cases in onnx, it has an incorrect result, I've extracted this test case for you, please use this:

for dev in [cpu_dev, gpu_dev]:

      X = np.random.random((1, 256, 12, 64)).astype(np.float32)
      x = tensor.from_numpy(X)
      x.to_device(dev)

      W = np.random.random((1, 256, 12, 64)).astype(np.float32)
      w = tensor.from_numpy(W)
      w.to_device(dev)

      X = np.transpose(X, (0, 2, 1, 3))
      W = np.transpose(W, (0, 2, 1, 3))
      W = np.transpose(W, (0, 1, 3, 2))
      Y = np.matmul(X, W)

      x = autograd.transpose(x, (0, 2, 1, 3))
      w = autograd.transpose(w, (0, 2, 1, 3))
      w = autograd.transpose(w, (0, 1, 3, 2))
      y = autograd.matmul(x, w)

      np.testing.assert_array_almost_equal(tensor.to_numpy(x), X)
      np.testing.assert_array_almost_equal(tensor.to_numpy(w), W)
      np.testing.assert_array_almost_equal(tensor.to_numpy(y), Y)

This test case reports:

Traceback (most recent call last):
File "../../test/python/test_tensor.py", line 389, in test_matmul
np.testing.assert_array_almost_equal(tensor.to_numpy(y), Y)
File "/usr/local/lib/python3.5/dist-packages/numpy/testing/_private/utils.py", line 1015, in assert_array_almost_equal
precision=decimal)
File "/usr/local/lib/python3.5/dist-packages/numpy/testing/_private/utils.py", line 827, in assert_array_compare
raise AssertionError(msg)
AssertionError:
Arrays are not almost equal to 6 decimals

Mismatch: 100%
Max absolute difference: 11.377065
Max relative difference: 1.148434
x: array([[[[14.145736, 13.506734, 13.252323, ..., 14.569139, 15.795746,
15.196916],
[18.76216 , 17.117498, 14.830437, ..., 17.131226, 17.974703,...
y: array([[[[16.710552, 17.515999, 15.49446 , ..., 15.438944, 18.269606,
15.611665],
[14.75556 , 14.552402, 14.04308 , ..., 14.70639 , 15.604133,...

from singa.

And please check the subgraph for this test case:

from singa.

Hi @joddiy , Thank you for pointing out the issue, kindly help to check the lastest code
PR:#639
your test is added:

from singa.

Hi @dcslin , it works now, thanks for your hlep.

from singa.