Fix zero gradient for subtensor assignment.

[bartvm]

@alexbw Partially fixes #126 and adds a unit test. It gives the correct answer in direct mode now, but still crashes in optimized mode.

[alexbw]

Will merge when Assignment test passes on CI

On Fri, May 27, 2016 at 5:38 PM Bart van Merriënboer <
notifications@github.com> wrote:

@alexbw https://github.com/alexbw Partially fixes #126
#126 and adds a unit
test. It gives the correct answer in direct mode now, but still crashes in

optimized mode.

You can view, comment on, or merge this pull request online at:

#127
Commit Summary

• Fix zero gradient for subtensor assignment.

File Changes

• M src/gradfuns.lua
  https://github.com/twitter/torch-autograd/pull/127/files#diff-0 (7)
• M test/test.lua
  https://github.com/twitter/torch-autograd/pull/127/files#diff-1 (5)

Patch Links:

• https://github.com/twitter/torch-autograd/pull/127.patch
• https://github.com/twitter/torch-autograd/pull/127.diff

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#127, or mute the thread
https://github.com/notifications/unsubscribe/AAJ4jzSbpB_VT7x5bzVhKIpXaZRAPPBNks5qF2RegaJpZM4Io4la
.

[bartvm]

The test that I added fails on the current master as well, it's a bug somewhere in optimized mode. I'd have to get a bit more familiar with the details of optimized mode before I can fix that.

[bartvm]

Maybe someone with better knowledge of the optimized mode internals can understand what is going on. The problem seems to be this line in Node.lua which specifically deals with Value.__internal_set:

mutationFlow:alias(self.inputs[1], valueAlias)

Unlike valueAlias = Value.from(outputs[1], Source.computed(self, 1)), self.inputs[1] here is not transformed into a Value instance. This line in runtime/codegen/backend/lua/init.lua takes issue with that:

aliasOp = graph.mutationFlow.history[i]
addNodeTargets(aliasOp.from.source.node, hazardNodes)

aliasOp.from.source doesn't have a node, it's simply a table that looks like this:

from    {
  type : "tensor"
  source :
    {
      type : "table"
      parent :
        {
          type : "param"
          name : 1
        }
      key : "x"
    }
  raw : DoubleTensor - size: 10x10
}

[bartvm]

Because I need this for something else as well now, another stab at it. I'm using torch.clone to sidestep the issue of params.x not being a node (as was done in the other tests).

However, now the test for optimized mode fails for another reason: The gradient gets computed correctly the first time grad(f) is called, but when it is called a second time it overwrites the gradient in place and sets it to zero. This surfaced accidentally because gradcheck happens to call grad(f) twice to check for determinism. I made the check more explicit, but I haven't figured out how to fix it yet.

[bartvm]

@luketwitter @alexbw Although I can't figure out the current bug in optimized mode triggered by the new unit test I added, can I propose at least merging my changes to src/gradfuns.lua? That way the gradients for index assignment in direct mode aren't silently set to zero anymore, which is a really nasty bug because it can make models completely fail to train without giving a single error.

I made a new PR for this: #139

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}