dense.to_sparse() re: #8853 #12171

realdoug · 2018-09-28T15:37:43Z

Here is my stab at dense.to_sparse

torch/_tensor_docs.py

+to_sparse() -> Tensor
+
+Returns a sparse copy of the tensor.  Torch supports sparse tensors
+in coordinate format. See https://pytorch.org/docs/stable/sparse.html.


torch/_tensor_docs.py

@@ -2384,6 +2384,14 @@ def callable(a, b) -> number
    0.012766935862600803
 """)

+add_docstr_all('to_sparse',


aten/src/ATen/native/sparse/SparseTensor.cpp

+
+  LongTensor indices = self.nonzero().transpose(0, 1);
+  std::vector<int64_t> maskShape;
+  if(sparseDims != dims){


weiyangfb · 2018-10-05T17:41:35Z

CIs failed look related

realdoug · 2018-10-11T20:11:12Z

@weiyangfb ok ready for review again

test/test_sparse.py

+                expected, _, _ = self._gen_sparse(sparse_dims, nnz, shape)
+                d = expected.to_dense()
+                result = d.to_sparse(sparse_dims)
+                self.assertEqual(expected.to_dense(), result.to_dense())  # == not implemented for sparse tensors yet


test/test_sparse.py

+        shape = [10, 5, 19, 8]
+        for dim, dim_sz in enumerate(shape):
+            sparse_dims = dim + 1
+            for nnz in range(1, dim_sz):


test/test_sparse.py

+                d = expected.to_dense()
+                result = d.to_sparse(sparse_dims)
+                self.assertEqual(expected.to_dense(), result.to_dense())  # == not implemented for sparse tensors yet
+                self.assertEqual(sparse_dims, result._sparseDims())


test/test_sparse.py

+                self.assertEqual(sparse_dims, result._sparseDims())
+
+        sp, _, _ = self._gen_sparse(2, 10, [3, 3, 3])
+        self.assertRaises(RuntimeError, lambda: sp.to_sparse())


aten/src/ATen/native/sparse/SparseTensor.cpp

+  }else{
+    Tensor i = nz.narrow(0, 0, sparseDims);
+    std::tie(indices, std::ignore) = _unique_dim(i, 1);
+    indices = indices.clone();


torch/_tensor_docs.py

+
+Returns a sparse copy of the tensor.  PyTorch supports sparse tensors in
+:ref:`coordinate format <sparse-docs>`.
+""")


weiyangfb

LGTM, there are some minor nits

realdoug · 2018-10-22T22:02:01Z

@weiyangfb sounds good, done & done!

realdoug · 2018-10-23T16:27:19Z

build failures were due to #12984. Retriggered the build via rebase

weiyangfb · 2018-10-24T17:21:27Z

@realdoug please rebase to fix the CI failures

aten/src/ATen/native/native_functions.yaml

@@ -2238,6 +2238,18 @@
  variants: function, method
  device_guard: False

+- func: to_sparse(Tensor self, int64_t sparseDims) -> Tensor


aten/src/ATen/native/sparse/SparseTensor.cpp

+  int64_t dims = self.dim();
+  AT_CHECK(sparse_dim > 0, "sparse_dim must be >0");
+  AT_CHECK(sparse_dim <= dims, 
+    "sparseDims must be less than or equal to self.dim()");


aten/src/ATen/native/sparse/SparseTensor.cpp

+
+  Tensor nz = self.nonzero().transpose(0, 1);
+  if(nz.numel() == 0){
+    return new_with_dims_sparse(sparse_dim, dims-sparse_dim, sizes, sparse_options);


torch/_tensor_docs.py

+    >>> d.to_sparse()
+    tensor(indices=tensor([[1, 1],
+                           [0, 2]]),
+        values=tensor([ 9, 10]),


torch/_tensor_docs.py

+        size=(3, 3), nnz=2, layout=torch.sparse_coo)
+    >>> d.to_sparse(1)
+    tensor(indices=tensor([[1]]),
+        values=tensor([[ 9,  0, 10]]),


aten/src/ATen/native/sparse/SparseTensor.cpp

+  LongTensor indices;
+  if(sparse_dim == dims){
+    indices = nz.clone();
+  }else{


aten/src/ATen/native/sparse/SparseTensor.cpp

+  std::vector<int64_t> sizes = self.sizes().vec();
+
+  Tensor nz = self.nonzero().transpose(0, 1);
+  if(nz.numel() == 0){


facebook-github-bot

weiyangfb has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

weiyangfb · 2018-10-26T03:24:46Z

need a rebase :(

facebook-github-bot

weiyangfb has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

Summary: Here is my stab at ```dense.to_sparse``` Pull Request resolved: pytorch/pytorch#12171 Differential Revision: D10859078 Pulled By: weiyangfb fbshipit-source-id: 5df72f72ba4f8f10e283402ff7731fd535682664

weiyangfb · 2018-10-27T07:26:37Z

landed, Thanks @realdoug !

realdoug requested review from apaszke, colesbury, ezyang, gchanan, soumith and zdevito as code owners September 28, 2018 15:37

ssnl reviewed Sep 28, 2018

View reviewed changes

torch/_tensor_docs.py Outdated

to_sparse() -> Tensor

Returns a sparse copy of the tensor. Torch supports sparse tensors

in coordinate format. See https://pytorch.org/docs/stable/sparse.html.

This comment was marked as off-topic.

Sign in to view

vishwakftw reviewed Sep 28, 2018

View reviewed changes

realdoug requested a review from zou3519 as a code owner October 1, 2018 03:42

realdoug mentioned this pull request Oct 3, 2018

Todo functions and autograd supports for Sparse Tensor #8853

Open

14 tasks

weiyangfb reviewed Oct 4, 2018

View reviewed changes

realdoug force-pushed the dense2sparse branch 2 times, most recently from facfac6 to 4b8027e Compare October 5, 2018 04:46

djsaunde mentioned this pull request Oct 7, 2018

Sparse tensor use cases #10043

Open

realdoug force-pushed the dense2sparse branch 5 times, most recently from 0823819 to 840624c Compare October 11, 2018 18:40

weiyangfb reviewed Oct 11, 2018

View reviewed changes

test/test_sparse.py Outdated

d = expected.to_dense()

result = d.to_sparse(sparse_dims)

self.assertEqual(expected.to_dense(), result.to_dense()) # == not implemented for sparse tensors yet

self.assertEqual(sparse_dims, result._sparseDims())

This comment was marked as off-topic.

Sign in to view

weiyangfb reviewed Oct 11, 2018

View reviewed changes

torch/_tensor_docs.py Outdated

Returns a sparse copy of the tensor. PyTorch supports sparse tensors in

:ref:`coordinate format <sparse-docs>`.

""")

This comment was marked as off-topic.

Sign in to view

realdoug mentioned this pull request Oct 14, 2018

odd behavior w/ add_out_dense_sparse_cuda #12633

Closed

realdoug force-pushed the dense2sparse branch 2 times, most recently from f30d038 to e7a624a Compare October 17, 2018 21:30

weiyangfb approved these changes Oct 22, 2018

View reviewed changes

realdoug force-pushed the dense2sparse branch 2 times, most recently from 6ff7c27 to dbff197 Compare October 23, 2018 16:22

weiyangfb self-assigned this Oct 23, 2018

realdoug force-pushed the dense2sparse branch from dbff197 to 5ac4ecb Compare October 24, 2018 18:26