Hello,
I am trying to create my own custom linear layer.
In my custom backward function I perform some calculation and then I need to call the function used in the backward pass
when I use the pytorch implementation I am able to do this:
device = "cuda"
x = torch.randn(16, 16, requires_grad=True, device=device)
t = torch.randn(16, 16, requires_grad=True, device=device)
m = torch.nn.Linear(16, 16, device=device)
o = m(x)
loss = MSELoss()
l = loss(o,t)
l.backward()
print(o.grad_fn)
#==><AddmmBackward0 at 0x7ee7ebf5e8c0>
From what I understand, the function used to compute the backward pass in a linear layer is called addmm_backward (or something close to that).
I perform the same kind of research for convolution layer and batch norm layer and I have been able to find that the function called in the backward are (respectively) convolution_backward and batch_norm_backward .
They are listed inside the file native_functions.yaml and are accessible through python/c++ API.
But there is no function called addmm_backward (or something close to that).
So , my question is there another functions listed in native_functions.yaml to perform the backward pass for a linear layer or another way to call the function addmm_backward ?
Thanks in advance for your answers.