Expose custom attributes from C++ functions#1430

Merged

apaszke merged 5 commits intopytorch:masterfrom

lantiga:autograd-attributes

May 7, 2017

Contributor

lantiga commented May 2, 2017

This PR is a resubmission of #1405 rebased onto master after the autograd refactor merge.

This PR demonstrates a possible way to expose custom attributes from C++ functions.
Motivation: right now functions implemented in C++ do not expose function-specific attributes or methods (e.g. stride for ConvForward). This makes it impossible to get such attributes while traversing a computation graph.

Currently only convolution attributes have been exposed, with the aim of illustrating the approach. If the proposal gets a thumbs up, I'll proceed and expose attributes for the rest of the C++ functions.

lantiga added 2 commits

May 2, 2017 14:02


          Expose custom attributes from C++ functions

be37df8


          Restore original function names

11a3fc9

apaszke reviewed

View reviewed changes

Contributor

apaszke left a comment

Looks great now 👍

torch/csrc/autograd/functions/init.cpp Outdated

                 registerCppFunction(typeid(C), &type);
               }
+              template<typename T, typename M, typename P, M P::*ptr, typename V, PyObject* (*Convert)(V)>

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

torch/csrc/autograd/functions/init.cpp Outdated

+              PyObject* getTupleAttr(PyObject* obj, void* _unused)
+              {
+                THPCppFunction* self = (THPCppFunction*)obj;
+                auto& arr = std::static_pointer_cast<T>(self->cdata).get()->*ptr;

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

torch/csrc/autograd/functions/init.cpp

+                for (size_t i = 0; i < num_elems; ++i) {
+                  PyTuple_SET_ITEM(py_tuple.get(), i, Convert(arr[i]));
+                }
+                return py_tuple.release();

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view


          Rename tpl parameters. Simplify casting. Wrap in HANDLE_TH_ERRORS for…

2d24ece

… exception handling.

Contributor Author

lantiga commented May 3, 2017

Thanks for the review @apaszke! At this point I'll complete the PR with BatchNorm, the only other C++ function with parameters. BatchNorm has thpp::Tensor as param, so I'll write another getter in the process.

apaszke reviewed

View reviewed changes

torch/csrc/autograd/functions/init.cpp Outdated

+              {
+                HANDLE_TH_ERRORS
+                THPCppFunction* self = (THPCppFunction*)obj;
+                auto& arr = (T*)(self->cdata.get())->*ptr;

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

torch/csrc/autograd/functions/init.cpp

+                {(char*)"transposed", (getter)getValueAttr<ConvForward, bool, ConvParams,
+                                                       &ConvParams::transposed, long, PyBool_FromLong>, NULL, NULL, NULL},
+                {(char*)"output_padding", (getter)getTupleAttr<ConvForward, std::vector<int>, ConvParams,
+                                                       &ConvParams::output_padding, long, PyInt_FromLong>, NULL, NULL, NULL},

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

lantiga added 2 commits

May 5, 2017 09:46


          Fix parens to clarify cast precedence

35ec4d4


          Add batch normalization custom attributes

f885c05

apaszke approved these changes

View reviewed changes

Contributor

apaszke commented May 7, 2017

@pytorchbot test this plesae

Contributor

apaszke commented May 7, 2017

@pytorchbot test this please

apaszke merged commit 23b556e into pytorch:master

Contributor

apaszke commented May 7, 2017

Thanks Luca!

caogang added a commit to caogang/pytorch that referenced this pull request


          Merge branch 'master' into develop

d56a703

* master:
  Add F.normalize (pytorch#1467)
  Expose custom attributes from C++ functions (pytorch#1430)
  Add high order gradient support for Sigmoid (pytorch#1496)

lantiga deleted the autograd-attributes branch

May 9, 2017 20:31

Jiaming-Liu pushed a commit to Jiaming-Liu/pytorch that referenced this pull request


          Expose custom attributes from C++ functions (pytorch#1430)

4f7a0eb

ezyang added the open source label

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels