Hi, I am new in PyTorch.
When I created the weight tensors by calling torch.nn.Conv2d, I saw that its weights are initialized by some way. its values are not similar to non-initialized version. (see the captured image)
Could you explain how these weights are initialized? I could not find any hint in docs file…
So if you use torch.Tensor it does not initialize it but puts some garbage. If you want some sort of random initialization, you should use torch.rand
>>> torch.rand(4,4)
tensor([[0.8693, 0.5824, 0.3661, 0.1016],
[0.4629, 0.7107, 0.1525, 0.9696],
[0.0603, 0.8134, 0.3207, 0.8813],
[0.0799, 0.1383, 0.3611, 0.3585]])
>>> torch.Tensor(4,4)
tensor([[-0.0000, 0.0000, -0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000, 0.0000],
[-0.0000, 0.0000, -0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000, 1.3564]])
>>> torch.Tensor(4,4)
tensor([[-0.0000, 0.0000, -0.0000, 0.0000],
[ 0.0000, 0.0000, -0.0000, 0.0000],
[-0.0000, 0.0000, 0.0000, 0.0000],
[ 0.0799, 0.1383, 0.3611, 0.3585]])
>>> torch.Tensor(4,4)
tensor([[-0.0000, 0.0000, -0.0000, 0.0000],
[ 0.0000, 0.0000, -0.0000, 0.0000],
[-0.0000, 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000, 1.3564]])
>>> torch.rand(4,4)
tensor([[0.9820, 0.4046, 0.1679, 0.9906],
[0.3627, 0.8029, 0.2323, 0.5281],
[0.2631, 0.0574, 0.6574, 0.3545],
[0.8812, 0.9527, 0.6162, 0.3369]])
But I can feel why you are confused because it looks as if it was randomly initialized.
This blog has a good writeup:
https://jhui.github.io/2018/02/09/PyTorch-Basic-operations/
The docs usually don’t mention the initialization method, but if you look at PyTorch’s source code (https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/conv.py#L41), you can see the weights are initialized with Kaiming uniform initialization.
Your answer is exactly what i was wondering . Thanks!!!
