Function Differences with torch.diag

torch.diag

torch.diag(
    input,
    diagonal=0,
    out=None
)

For more information, see torch.diag.

mindspore.nn.MatrixDiag

class mindspore.nn.MatrixDiag()(x)

For more information, see mindspore.nn.MatrixDiag.

Differences

PyTorch: Only 1D and 2D are supported. If the input is a 1D Tensor, a 2D diagonal matrix will be returned, and all elements in the returned matrix are set to 0 except the diagonals. If the input is a 2D Tensor, the value on the diagonal of the matrix will be returned. It also supports diagonal offsets specified by parameter diagonal.

MindSpore：Returns a diagonal matrix based on the given value, and k+1 dimensional diagonal matrix for k dimensional input.

Code Example

import mindspore
from mindspore import Tensor, nn
import torch
import numpy as np

x1 = np.random.randn(2)
x2 = np.random.randn(2, 3)
x3 = np.random.randn(2, 3, 4)

# In MindSpore, for the given k-dimension input, a k+1 dimension diagonal matrix will be returned.
matrix_diag = nn.MatrixDiag()
for n, x in enumerate([x1, x2, x3]):
    try:
        input_x = Tensor(x, mindspore.float32)
        output = matrix_diag(input_x)
        print('input shape: {}; output size: {}'.format(
            str(n + 1), str(output.shape)
        ))
    except Exception as e:
        print('ERROR: ' + str(e))
# Out:
# input shape: 1; output size: (2, 2)
# input shape: 2; output size: (2, 3, 3)
# input shape: 3; output size: (2, 3, 4, 4)

# In torch, output for 1-dimension and 2-dimension input will be returned based on different rules.
# If the dimension of the input is greater than 2, it will raise error.
for n, x in enumerate([x1, x2, x3]):
    try:
        input_x = torch.tensor(x)
        output = torch.diag(input_x)
        print('input shape: {}; output size: {}'.format(
            str(n + 1), str(output.shape)
        ))
    except Exception as e:
        print('ERROR: ' + str(e))
# Out：
# input shape: 1; output size: torch.Size([2, 2])
# input shape: 2; output size: torch.Size([2])
# ERROR: matrix or a vector expected