Sparse tensor printing; add NotImplemented autograd fn by ssnl · Pull Request #10181 · pytorch/pytorch

ssnl · 2018-08-02T21:24:47Z

Commits:

Add autograd function NotImplemented (subclass of Error) so python grad_fn prints nicer. Since Error is used in DelayedError to implement @oncedifferentiable, I can't just change its name. cc @colesbury
Add printing for sparse tensors. Fixes grad_fn of sparse tensor not displayed correctly #9412 . cc @weiyangfb @li-roy .
Add tests for sparse printing

Examples:

  In [2]: x = torch.sparse.FloatTensor(torch.arange(4).view(2,2), torch.randn(2, 2), [10, 10, 2])
  
  In [3]: x
  Out[3]:
- torch.sparse.FloatTensor of size (10,10,2) with indices:
- tensor([[0, 1],
-         [2, 3]])
- and values:
- tensor([[-1.1832, -0.5927],
-         [ 0.0831,  0.2511]])
+ tensor(indices=tensor([[0, 1],
+                        [2, 3]]),
+        values=tensor([[ 1.5081,  0.3451],
+                       [-0.0392,  0.4776]]),
+        size=(10, 10, 2), nnz=2, layout=torch.sparse_coo)
  
  In [4]: x.requires_grad_()
  Out[4]:
- torch.sparse.FloatTensor of size (10,10,2) with indices:
- tensor([[0, 1],
-         [2, 3]], grad_fn=<Error>)
- and values:
- tensor([[-1.1832, -0.5927],
-         [ 0.0831,  0.2511]], grad_fn=<Error>)
+ tensor(indices=tensor([[0, 1],
+                        [2, 3]]),
+        values=tensor([[ 1.5081,  0.3451],
+                       [-0.0392,  0.4776]]),
+        size=(10, 10, 2), nnz=2, layout=torch.sparse_coo, requires_grad=True)
  
  In [5]: x + x
  Out[5]:
- torch.sparse.FloatTensor of size (10,10,2) with indices:
- tensor([[0, 1],
-         [2, 3]], grad_fn=<Error>)
- and values:
- tensor([[-2.3664, -1.1855],
-         [ 0.1662,  0.5021]], grad_fn=<Error>)
+ tensor(indices=tensor([[0, 1],
+                        [2, 3]]),
+        values=tensor([[ 3.0162,  0.6902],
+                       [-0.0785,  0.9553]]),
+        size=(10, 10, 2), nnz=2, layout=torch.sparse_coo, grad_fn=<AddBackward0>)
  
  In [6]: x.double()
  Out[6]:
- torch.sparse.DoubleTensor of size (10,10,2) with indices:
- tensor([[0, 1],
-         [2, 3]], grad_fn=<Error>)
- and values:
- tensor([[-1.1832, -0.5927],
-         [ 0.0831,  0.2511]], dtype=torch.float64, grad_fn=<Error>)
+ tensor(indices=tensor([[0, 1],
+                        [2, 3]]),
+        values=tensor([[ 1.5081,  0.3451],
+                       [-0.0392,  0.4776]]),
+        size=(10, 10, 2), nnz=2, dtype=torch.float64, layout=torch.sparse_coo,
+        grad_fn=<NotImplemented>)
  
  In [7]: x = torch.sparse.FloatTensor(torch.ones(0, 2, dtype=torch.long), torch.randn(2, 0), [0])
  
  In [8]: x
  Out[8]:
- torch.sparse.FloatTensor of size (0,) with indices:
- tensor([], size=(0, 2), dtype=torch.int64)
- and values:
- tensor([], size=(2, 0))
+ tensor(indices=tensor([], size=(0, 2)),
+        values=tensor([], size=(2, 0)),
+        size=(0,), nnz=2, layout=torch.sparse_coo)
  
  In [9]: x = torch.sparse.FloatTensor(torch.ones(0, 2, dtype=torch.long), torch.randn(2), [])
  
  In [10]: x
  Out[10]:
- torch.sparse.FloatTensor of size () with indices:
- tensor([], size=(0, 2), dtype=torch.int64)
- and values:
- tensor([-0.0064,  0.8518])
+ tensor(indices=tensor([], size=(0, 2)),
+        values=tensor([ 0.9800, -0.5978]),
+        size=(), nnz=2, layout=torch.sparse_coo)

facebook-github-bot

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

ssnl · 2018-08-03T03:05:09Z

@pytorchbot retest this please

facebook-github-bot

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

torch/_tensor_str.py

test/test_sparse.py

weiyangfb

Very nice!

facebook-github-bot

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

gchanan

Could you include before and after comparisons?

Without checking the before-and-after closely, this looks good to me (after the nits are addressed). I have some concerns about how this looks post #9279 -- for example, you never print the shape of the indices tensor, but your print can be wrong in the numel() == 0 case, because the shape is really part of the tensor print (because you can't deduce the shape from the print).

test/test_sparse.py

torch/csrc/autograd/functions/basic_ops.h

torch/_tensor_str.py

fmassa · 2018-08-28T16:46:31Z

ping @ssnl on Greg's comments. Also, this now needs a rebase

aten/src/ATen/native/sparse/SparseUtils.h

facebook-github-bot

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

ezyang · 2018-09-04T16:21:22Z

@pytorchbot retest this please

ssnl · 2018-09-04T16:37:14Z

@pytorchbot retest this please

facebook-github-bot

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

ssnl · 2018-09-04T21:35:06Z

@pytorchbot retest this please

…ad_fn prints nicer

facebook-github-bot

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

ssnl · 2018-09-05T05:29:25Z

@pytorchbot retest this please

torch/csrc/Size.cpp

+  for (Py_ssize_t i = 0; i < PyTuple_Size((PyObject*)self); ++i) {
+    numel *= PyLong_AsLong(PyTuple_GET_ITEM(self, i));
+  }
+  return THPUtils_packUInt64(numel);


test/test_sparse.py

+        ]
+
+        printed = []
+        for shape, dimI, nnz in shape_dimI_nnz:


test/test_sparse.py

+            for d in range(dimI):
+                indices[d].clamp_(max=shape[d])  # make it valid index
+            if indices.numel() > 0:
+                indices[:, -1] = indices[:, 0]  # make it uncoalesced


torch/_tensor_str.py

+        if not has_default_dtype:
+            suffixes.append('dtype=' + str(self.dtype))
+        indices_prefix = 'indices=tensor('
+        indices = self._indices().detach()


facebook-github-bot

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

Summary: Commits: 1. Add autograd function `NotImplemented` (subclass of `Error`) so python `grad_fn` prints nicer. Since `Error` is used in `DelayedError` to implement `oncedifferentiable`, I can't just change its name. cc colesbury 2. Add printing for sparse tensors. Fixes pytorch/pytorch#9412 . cc weiyangfb The controller you requested could not be found. . 3. Add tests for sparse printing Examples: ```diff In [2]: x = torch.sparse.FloatTensor(torch.arange(4).view(2,2), torch.randn(2, 2), [10, 10, 2]) In [3]: x Out[3]: - torch.sparse.FloatTensor of size (10,10,2) with indices: - tensor([[0, 1], - [2, 3]]) - and values: - tensor([[-1.1832, -0.5927], - [ 0.0831, 0.2511]]) + tensor(indices=tensor([[0, 1], + [2, 3]]), + values=tensor([[ 1.5081, 0.3451], + [-0.0392, 0.4776]]), + size=(10, 10, 2), nnz=2, layout=torch.sparse_coo) In [4]: x.requires_grad_() Out[4]: - torch.sparse.FloatTensor of size (10,10,2) with indices: - tensor([[0, 1], - [2, 3]], grad_fn=<Error>) - and values: - tensor([[-1.1832, -0.5927], - [ 0.0831, 0.2511]], grad_fn=<Error>) + tensor(indices=tensor([[0, 1], + [2, 3]]), + values=tensor([[ 1.5081, 0.3451], + [-0.0392, 0.4776]]), + size=(10, 10, 2), nnz=2, layout=torch.sparse_coo, requires_grad=True) In [5]: x + x Out[5]: - torch.sparse.FloatTensor of size (10,10,2) with indices: - tensor([[0, 1], - [2, 3]], grad_fn=<Error>) - and values: - tensor([[-2.3664, -1.1855], - [ 0.1662, 0.5021]], grad_fn=<Error>) + tensor(indices=tensor([[0, 1], + [2, 3]]), + values=tensor([[ 3.0162, 0.6902], + [-0.0785, 0.9553]]), + size=(10, 10, 2), nnz=2, layout=torch.sparse_coo, grad_fn=<AddBackward0>) In [6]: x.double() Out[6]: - torch.sparse.DoubleTensor of size (10,10,2) with indices: - tensor([[0, 1], - [2, 3]], grad_fn=<Error>) - and values: - tensor([[-1.1832, -0.5927], - [ 0.0831, 0.2511]], dtype=torch.float64, grad_fn=<Error>) + tensor(indices=tensor([[0, 1], + [2, 3]]), + values=tensor([[ 1.5081, 0.3451], + [-0.0392, 0.4776]]), + size=(10, 10, 2), nnz=2, dtype=torch.float64, layout=torch.sparse_coo, + grad_fn=<NotImplemented>) In [7]: x = torch.sparse.FloatTensor(torch.ones(0, 2, dtype=torch.long), torch.randn(2, 0), [0]) In [8]: x Out[8]: - torch.sparse.FloatTensor of size (0,) with indices: - tensor([], size=(0, 2), dtype=torch.int64) - and values: - tensor([], size=(2, 0)) + tensor(indices=tensor([], size=(0, 2)), + values=tensor([], size=(2, 0)), + size=(0,), nnz=2, layout=torch.sparse_coo) In [9]: x = torch.sparse.FloatTensor(torch.ones(0, 2, dtype=torch.long), torch.randn(2), []) In [10]: x Out[10]: - torch.sparse.FloatTensor of size () with indices: - tensor([], size=(0, 2), dtype=torch.int64) - and values: - tensor([-0.0064, 0.8518]) + tensor(indices=tensor([], size=(0, 2)), + values=tensor([ 0.9800, -0.5978]), + size=(), nnz=2, layout=torch.sparse_coo) ``` Pull Request resolved: pytorch/pytorch#10181 Differential Revision: D9139845 Pulled By: SsnL fbshipit-source-id: 353eebd55fac4049ed9bf85f8b0ee2c1418a744e

* upstream/master: (26 commits) cudnn 7 upgrade with spatialBN fix (pytorch#11291) Ignore FuseGraph Call on Windows (pytorch#11015) defer resolution of mkl to a cmake wrapper library (pytorch#11298) Cleanup dependency of distributed flags (pytorch#11221) Move minimal wrapdim functionality to core, remove THTensor include i… (pytorch#11283) Change includes from ATen/Storage.h to ATen/core/Storage.h (pytorch#11217) Fix scalar tensor assert in fusion compiler (pytorch#10952) Add dead code elimination pass (pytorch#10101) Distributed Data Parallel CPU module for C10D (pytorch#11168) Back out "[pt1][tensor] Add strides to caffe2::Tensor" Fix conv gradient conversion (pytorch#11312) Bag of clang tidy fixes for torch/csrc/ and torch/csrc/autograd (pytorch#11050) Sparse tensor printing; add NotImplemented autograd fn (pytorch#10181) Add convertToCaffe2Proto to python API fix doc for functional.dropout* (pytorch#10417) typo fix Tranpose2D -> Transpose2D (pytorch#11281) Remove THFinalizer Forward declarations of needed curand functions (pytorch#10911) nomnigraph - simplify core graph API and test (pytorch#11256) Small fixes to cppdocs for sync script (pytorch#11300) ...

Summary: Commits: 1. Add autograd function `NotImplemented` (subclass of `Error`) so python `grad_fn` prints nicer. Since `Error` is used in `DelayedError` to implement `oncedifferentiable`, I can't just change its name. cc colesbury 2. Add printing for sparse tensors. Fixes pytorch#9412 . cc weiyangfb The controller you requested could not be found. . 3. Add tests for sparse printing Examples: ```diff In [2]: x = torch.sparse.FloatTensor(torch.arange(4).view(2,2), torch.randn(2, 2), [10, 10, 2]) In [3]: x Out[3]: - torch.sparse.FloatTensor of size (10,10,2) with indices: - tensor([[0, 1], - [2, 3]]) - and values: - tensor([[-1.1832, -0.5927], - [ 0.0831, 0.2511]]) + tensor(indices=tensor([[0, 1], + [2, 3]]), + values=tensor([[ 1.5081, 0.3451], + [-0.0392, 0.4776]]), + size=(10, 10, 2), nnz=2, layout=torch.sparse_coo) In [4]: x.requires_grad_() Out[4]: - torch.sparse.FloatTensor of size (10,10,2) with indices: - tensor([[0, 1], - [2, 3]], grad_fn=<Error>) - and values: - tensor([[-1.1832, -0.5927], - [ 0.0831, 0.2511]], grad_fn=<Error>) + tensor(indices=tensor([[0, 1], + [2, 3]]), + values=tensor([[ 1.5081, 0.3451], + [-0.0392, 0.4776]]), + size=(10, 10, 2), nnz=2, layout=torch.sparse_coo, requires_grad=True) In [5]: x + x Out[5]: - torch.sparse.FloatTensor of size (10,10,2) with indices: - tensor([[0, 1], - [2, 3]], grad_fn=<Error>) - and values: - tensor([[-2.3664, -1.1855], - [ 0.1662, 0.5021]], grad_fn=<Error>) + tensor(indices=tensor([[0, 1], + [2, 3]]), + values=tensor([[ 3.0162, 0.6902], + [-0.0785, 0.9553]]), + size=(10, 10, 2), nnz=2, layout=torch.sparse_coo, grad_fn=<AddBackward0>) In [6]: x.double() Out[6]: - torch.sparse.DoubleTensor of size (10,10,2) with indices: - tensor([[0, 1], - [2, 3]], grad_fn=<Error>) - and values: - tensor([[-1.1832, -0.5927], - [ 0.0831, 0.2511]], dtype=torch.float64, grad_fn=<Error>) + tensor(indices=tensor([[0, 1], + [2, 3]]), + values=tensor([[ 1.5081, 0.3451], + [-0.0392, 0.4776]]), + size=(10, 10, 2), nnz=2, dtype=torch.float64, layout=torch.sparse_coo, + grad_fn=<NotImplemented>) In [7]: x = torch.sparse.FloatTensor(torch.ones(0, 2, dtype=torch.long), torch.randn(2, 0), [0]) In [8]: x Out[8]: - torch.sparse.FloatTensor of size (0,) with indices: - tensor([], size=(0, 2), dtype=torch.int64) - and values: - tensor([], size=(2, 0)) + tensor(indices=tensor([], size=(0, 2)), + values=tensor([], size=(2, 0)), + size=(0,), nnz=2, layout=torch.sparse_coo) In [9]: x = torch.sparse.FloatTensor(torch.ones(0, 2, dtype=torch.long), torch.randn(2), []) In [10]: x Out[10]: - torch.sparse.FloatTensor of size () with indices: - tensor([], size=(0, 2), dtype=torch.int64) - and values: - tensor([-0.0064, 0.8518]) + tensor(indices=tensor([], size=(0, 2)), + values=tensor([ 0.9800, -0.5978]), + size=(), nnz=2, layout=torch.sparse_coo) ``` Pull Request resolved: pytorch#10181 Differential Revision: D9139845 Pulled By: SsnL fbshipit-source-id: 353eebd55fac4049ed9bf85f8b0ee2c1418a744e

ssnl requested review from apaszke, colesbury, ezyang, gchanan, soumith and zdevito as code owners August 2, 2018 21:24

facebook-github-bot reviewed Aug 2, 2018

View reviewed changes

ssnl force-pushed the sp_prt branch from b1fd9e5 to e0eb7dd Compare August 2, 2018 21:26

facebook-github-bot reviewed Aug 3, 2018

View reviewed changes

weiyangfb reviewed Aug 3, 2018

View reviewed changes

torch/_tensor_str.py Outdated

This comment was marked as off-topic.

Sign in to view

weiyangfb reviewed Aug 3, 2018

View reviewed changes

test/test_sparse.py Outdated

This comment was marked as off-topic.

Sign in to view

weiyangfb reviewed Aug 3, 2018

View reviewed changes

weiyangfb mentioned this pull request Aug 3, 2018

Fix grad_fn=<Error> in printing sparse Variable #9802

Closed

ssnl force-pushed the sp_prt branch from 7fc8126 to fd36f5f Compare August 3, 2018 14:47

facebook-github-bot reviewed Aug 3, 2018

View reviewed changes

gchanan reviewed Aug 6, 2018

View reviewed changes

test/test_sparse.py Outdated

This comment was marked as off-topic.

Sign in to view

torch/csrc/autograd/functions/basic_ops.h Outdated

This comment was marked as off-topic.

Sign in to view

torch/_tensor_str.py Outdated

This comment was marked as off-topic.

Sign in to view

ssnl mentioned this pull request Aug 7, 2018

[READY] Preserve sparse tensor shape and dim invariants, and add scalar tensor support #9279

Closed

ssnl force-pushed the sp_prt branch 5 times, most recently from 436e030 to c3f00e6 Compare September 4, 2018 15:39

ssnl commented Sep 4, 2018

View reviewed changes

aten/src/ATen/native/sparse/SparseUtils.h Outdated

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

ssnl force-pushed the sp_prt branch from c3f00e6 to 0bc65ea Compare September 4, 2018 15:50

facebook-github-bot reviewed Sep 4, 2018

View reviewed changes

ssnl added 7 commits September 4, 2018 19:13

Add autograd function NotImplemented (subclass of Error) so python gr…

833b4c8

…ad_fn prints nicer

Update sparse tensor printing

f8b0621

add sparse test_print

b77772d

update print; address comments

565badf

oncedifferentiable should use Error

5bf2d91

adress comments

9dad3fe

Fix for rocm

d8969e5

ssnl force-pushed the sp_prt branch from b0d301c to d8969e5 Compare September 4, 2018 23:13

facebook-github-bot reviewed Sep 4, 2018

View reviewed changes

ssnl mentioned this pull request Sep 4, 2018

[sparse] Autograd get_indices/values and sparse_coo ctor #11253

Closed

weiyangfb mentioned this pull request Sep 5, 2018

torch.zeros for sparse matrices #4204

Closed

gchanan approved these changes Sep 5, 2018

View reviewed changes

address comments

72ca047

facebook-github-bot reviewed Sep 6, 2018

View reviewed changes

facebook-github-bot closed this in 83a1ab2 Sep 6, 2018

ssnl deleted the sp_prt branch September 6, 2018 02:54

ezyang added open source merged labels Jun 24, 2019

Conversation

ssnl commented Aug 2, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

ssnl commented Aug 3, 2018

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

weiyangfb left a comment

Choose a reason for hiding this comment

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

gchanan left a comment

Choose a reason for hiding this comment

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

fmassa commented Aug 28, 2018

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

ezyang commented Sep 4, 2018

Uh oh!

ssnl commented Sep 4, 2018

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

ssnl commented Sep 4, 2018

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

ssnl commented Sep 5, 2018

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

ssnl commented Aug 2, 2018 •

edited

Loading