Restore allow_unused functionality #6553
Conversation
ssnl
requested review from
apaszke,
colesbury,
ezyang,
gchanan,
soumith and
zdevito
as code owners
| if (!grad_fn) { | ||
| output_edges.emplace_back(); | ||
| } else { | ||
| THPUtils_assert(grad_fn, |
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
apaszke
left a comment
apaszke
left a comment
There was a problem hiding this comment.
Two minor things, and it's good to merge.
I actually can't remember why I changed that, sorry to hear that it caused those problems. I know for sure that I wanted to remove only_inputs, but I can't think of any reasons why allow_unused got cut.
It might have been that I've heard many comments saying that returning None is what most people expect, and this is also what other libraries do, but I think it's ok to be strict in this case. This often happens when taking higher order derivatives.
| unsigned char keep_graph = 0; | ||
| unsigned char create_graph = 0; | ||
| PyObject *inputs = nullptr; | ||
| unsigned char allow_unreachable = 1; |
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
| THPObjectPtr py_outputs {PyTuple_New(num_inputs)}; | ||
| if (!py_outputs) return nullptr; | ||
| for (int i = 0; i < num_inputs; i++) { | ||
| THPUtils_assert(allow_unreachable || outputs[i].defined(), "One of the differentiated " |
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
test/test_jit.py
Outdated
| l1_ge, recording_inputs, create_graph=True, allow_unused=True) | ||
| l2_ge = (allSum(grads_ge) * l1_ge) | ||
| grads2_ge = torch.autograd.grad(l2_ge, recording_inputs) | ||
| grads2_ge = torch.autograd.grad(l2_ge, recording_inputs, allow_unused=True) |
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
|
Hi @apaszke , What is the reason for deprecating the I think it was useful to do partial backpropagations while still filling the gradient buffers on the way. Without it, is there another way to halt the backpropagation flow at a given node when calling Thanks! |
4 participants
Addresses #6535
I'll do a Variable -> Tensor codemod in a separate PR.