Allow sharing of tensors whose size is unknown to some parties #184
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Summary: When a party wants to secret-share a tensor, the CrypTen API currently assumes that the other parties know what the size of that tensor will be. This may be problematic, for example, if a party wants to share a data tensor whose size is unknown a priori. To make matters worse, if a party makes an incorrect assumption about the size of the tensor they will receive, the PRZS initialization does not work correctly and the parties may end up with secret shares of a tensor that do not match in size (which, somewhat surprisingly, does not always appear to be caught by Gloo during computations). The diff adds an `broadcast_size` option in the construction of secret-shared tensors. If set to `True`, the `src` party will broadcast the size of the tensor to all other parties so that they can execute the PRZS initialization correctly. We could also add an additional check on the size of the secret shares for all initializations, but I did not do that because it would add a communication round (somewhat defeating the purpose of doing PRZS initialization). This means that the failure case described above still exists; I added documentation describing it. Differential Revision: D24381193 fbshipit-source-id: 3baf2373d40fc880f7d5741178e182c0661b46b2
facebook-github-bot
added
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D24381193 |
gmuraru
reviewed
Oct 17, 2020
| @@ -264,7 +264,7 @@ def send_obj(self, obj, dst, group=None): | |||
|
|
|||
| buf = pickle.dumps(obj) | |||
| size = torch.tensor(len(buf), dtype=torch.int32) | |||
| arr = torch.from_numpy(numpy.frombuffer(buf, dtype=numpy.int8)) | |||
| arr = torch.from_numpy(numpy.copy(numpy.frombuffer(buf, dtype=numpy.int8))) | |||
There was a problem hiding this comment.
From what I have read from here[1], the arr should share the same memory with the numpy array in case numpy.copy is not used.
Q: It is needed to create a copy of the data? (by using this I think we are playing it safe, but if buf becomes bigger it might introduce an overhead because of the copy)
[1] https://pytorch.org/docs/stable/generated/torch.from_numpy.html
There was a problem hiding this comment.
Without this copy, I am seeing this warning. I tried altering the writeable flag of the numpy array directly but that does not appear to be allowed here. The copy makes the warning go away.
There was a problem hiding this comment.
Yep, you are right!
I have read here that pytorch does not support for the moment read-only tensors!
gmuraru
approved these changes
Oct 18, 2020
facebook-github-bot
pushed a commit
that referenced
this pull request
Oct 21, 2020
Summary: Pull Request resolved: #184 When a party wants to secret-share a tensor, the CrypTen API currently assumes that the other parties know what the size of that tensor will be. This may be problematic, for example, if a party wants to share a data tensor whose size is unknown a priori. To make matters worse, if a party makes an incorrect assumption about the size of the tensor they will receive, the PRZS initialization does not work correctly and the parties may end up with secret shares of a tensor that do not match in size (which, somewhat surprisingly, does not always appear to be caught by Gloo during computations). The diff adds an `broadcast_size` option in the construction of secret-shared tensors. If set to `True`, the `src` party will broadcast the size of the tensor to all other parties so that they can execute the PRZS initialization correctly. We could also add an additional check on the size of the secret shares for all initializations, but I did not do that because it would add a communication round (somewhat defeating the purpose of doing PRZS initialization). This means that the failure case described above still exists; I added documentation describing it. Reviewed By: knottb Differential Revision: D24381193 fbshipit-source-id: 0b2ae19df973dc777f546944adc1e7676d3ef621
knottb
pushed a commit
to gmuraru/CrypTen
that referenced
this pull request
Dec 18, 2020
…h#184) Summary: Pull Request resolved: fairinternal/CrypTen#184 This is a first working version of the Tensorflow model conversion to CrypTen. This currently works on simple three layer fully connected network. Future diffs will ensure that additional modules are supported. Reviewed By: lvdmaaten Differential Revision: D20661986 fbshipit-source-id: 60bffed2ecd4363a7afe0658453f8856460e60ca
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
When a party wants to secret-share a tensor, the CrypTen API currently assumes that the other parties know what the size of that tensor will be. This may be problematic, for example, if a party wants to share a data tensor whose size is unknown a priori. To make matters worse, if a party makes an incorrect assumption about the size of the tensor they will receive, the PRZS initialization does not work correctly and the parties may end up with secret shares of a tensor that do not match in size (which, somewhat surprisingly, does not always appear to be caught by Gloo during computations).
The diff adds an
broadcast_sizeoption in the construction of secret-shared tensors. If set toTrue, thesrcparty will broadcast the size of the tensor to all other parties so that they can execute the PRZS initialization correctly.We could also add an additional check on the size of the secret shares for all initializations, but I did not do that because it would add a communication round (somewhat defeating the purpose of doing PRZS initialization). This means that the failure case described above still exists; I added documentation describing it.
Differential Revision: D24381193