Logsigmoid as chaining operation

[Bartolo1024]

🚀 Feature

I would like to have a chained log sigmoid operation in torch.Tensor object.

Motivation

I prefer to use chained versions of the operations like x.sum(), x.mean(), etc. instead of torch.sum() torch.mean(), etc.
This prevents me from the formulation of hardly readable long operations.
For example:
x.pow(2).sum(dim=1).mean()
is much cleaner than:
torch.mean(torch.sum(torch.pow(x, 2), dim=1))

Currently, the operation is available only in torch.nn or as torch module method or can be implemented as the combination of sigmoid and logarithm, which is less numerically stable.

Pitch

Add log sigmoid to tensor object methods.

import torch
x = torch.rand(2, 3, 10, 10)
log_sigmoid_from_x = x.logsigmoid()

[malfet]

Hmm, sigmoid is method of Tensor, so why not logsigmoid couldn't be one.
@ezyang, what do you think?

[ezyang]

...suuuuuure? it's one darn slippery slope, but logsigmoid might deserve it. cc @gchanan

[vadimkantorov]

softmax/logsumexp/log_softmax is also very-very common, may deserve migration from torch.nn.functional to torch, but it's a slippery slope indeed...

[Bartolo1024]

@vadimkantorov softmax/logsumexp/log_softmax are already available as torch module and torch.Tensor methods?

[vadimkantorov]

@Bartolo1024 I think this is currently unintetional...

[firstprayer]

Only place I found so far related to how is Tensor.xxx() implemented is this

pytorch/torch/csrc/jit/runtime/symbolic_script.cpp

Lines 783 to 788 in 6162343

	def sigmoid(self):
	result = torch.sigmoid(self)
	def backward(grad_output):
	return (1 - result) * result * grad_output
	return result, backward

Looks like it wraps it within a helper function but still calling torch.xxx() under the hook>