Which dimensions of Matrix are multiplied?

143 · 2019 年12 月 12 日 23:27

I want to multiply A(2, 6) with B(6, 1) to obtain a Matrix with shape (2, 1).
If I use A= ti.Matrix(1, dt=ti.f32, shape=(2, 6), needs_grad=True). It will initialize a Matrix with shape (2, 6, 1, 1).
If I use A = ti.Matrix(2, 6, dt=ti.f32, shape=1, needs_grad=True). It will initialize a Matrix with shape (1, 2, 6).
I want to ask which two dimensions are multiplied when I implement multiplication another matrix.
I also want to ask which is the correct way to initialize a Matrix.

yuanming · 2019 年12 月 13 日 03:39

I’m working on a documentation page to clarify this. Please expect that to be online in 45 minutes.
A quick answer for now:
A = ti.Matrix(1, 4, dt=ti.f32, shape=(2, 6), needs_grad=True) allocates a 2x6 tensor of 1x4 matrices. You need to first get the matrix by tensor indexing, e.g. A[1, 5], and then do your matrice multiplication using the @ operator.

yuanming · 2019 年12 月 13 日 04:17

It seems to me that the correct way for your case is to use
A = ti.Matrix(2, 6, dt=ti.f32, shape=(), needs_grad=True), which allocates you a 0D tensor of 2x6 matrices.

yuanming · 2019 年12 月 13 日 04:36

More doc on tensor of matrices: https://taichi.readthedocs.io/en/latest/tensor_matrix.html

143 · 2019 年12 月 13 日 04:55

Thank you so much! The documents are very clear now.

143 · 2019 年12 月 13 日 05:06

Is there a way to create a matrix with dynamic shape? Like a universal data holder?

yuanming · 2019 年12 月 13 日 05:10

Do you mean something like std::vector<Matrix>?

143 · 2019 年12 月 13 日 05:11

In C++, yes. In python, a standard torch tensor may have dynamic batch size.

143 · 2019 年12 月 13 日 05:12

Also, another question is how to pass data = torch.zeros((2, 6)) to ti_data_holder = ti.Matrix(2, 6, dt=ti.f32, shape=(), needs_grad=True). It says list index out of range.

yuanming · 2019 年12 月 13 日 05:31

Sorry v0.2.4 does not support from/to_torch for 0D tensors yet. I just now pushed v0.2.5 with this feature. It will be available in 20 minutes. Example:

github.com

yuanming-hu/taichi/blob/master/tests/python/test_torch_io.py#L129


  for i in range(n):
    for j in range(n):
      for k in range(2):
        for l in range(3):
          assert x2[i, j][k, l] == 2


  t3 = x2.to_torch()
  assert (t2 == t3).all()


@ti.host_arch
def test_io_simple():
  if not ti.has_pytorch():
    return
  import torch
  mat = ti.Matrix(2, 6, dt=ti.f32, shape=(), needs_grad=True)
  zeros = torch.zeros((2, 6))
  zeros[1, 2] = 3
  mat.from_torch(zeros + 1)
  
  assert mat[None][1, 2] == 4

yuanming · 2019 年12 月 13 日 05:37

For something like std::vector<Matrix>, see here: https://github.com/yuanming-hu/taichi/blob/4c12a6b7c6a488478ca620aa5d7f40272eaf0a0b/tests/python/test_dynamic.py#L56

Although a dynamic batch size is doable, that would require a deeper understanding into Taichi. I would suggest using a fixed batch size for now.