Reuse of nn.Sequential()

1

If I define my net like this:

class Net(): 
   def __init__
      self.block = nn.Sequential(...)
   def forward
      x0 = self.block(inputs[0])
      x1 = self.block(inputs[1])

Will the learnables of self.block be shared among input 0 and 1, or will each call to self.block in the forward function automatically upscale the number of learnables?

2

self.block is reused in your approach and so will be its trainable parameters. PyTorch’s Autograd creates a computation graph capturing both operations, won’t create new parameters, but will just reuse them.