Conv2DGrad & MaxPoolGradHelper #12665
Conversation
|
Can one of the admins verify this patch? |
|
@bpiel, thanks for your PR! By analyzing the history of the files in this pull request, we identified @tensorflower-gardener, @DeNeutoy and @vrv to be potential reviewers. |
| } | ||
| REGISTER_GRADIENT_OP("Conv2D", Conv2DGrad); | ||
|
|
||
| Status MaxPoolGradHelper(const Scope& scope, const Operation& op, |
There was a problem hiding this comment.
I am not sure what the status of MaxPool and MaxPoolV2 is, but my guess is that MaxPoolV2 is the one more commonly used now. You may want to also add that gradient
tensorflow/tensorflow/python/ops/nn_grad.py
Line 545 in 86d9171
If you feel like it you can add MaxPoolGradGrad and MaxPoolGradGradV2, but I doubt these are blocking your models.
There was a problem hiding this comment.
Ah, didn't know about MaxPoolV2. Added it.
|
should this get merged? Just worried about the branch becoming in conflict with master |
|
Jenkins, test this please. |
|
GPU build rerun at http://ci.tensorflow.org/job/tensorflow-pull-requests-gpu/6635 |
|
What's the difference between MaxPool and MaxPoolV2? Should it be used one over the other? |
|
We (almost) never change ops, whenever we have to, we instead introduce a new V2 (or V3, or V4) op. That makes sure that old serialized graphdefs keep working. So in your code, use V2, it's the newest version. |
Gradients for the nn ops, Conv2D and MaxPool. Ported from python:
tensorflow/tensorflow/python/ops/nn_grad.py
Line 457 in 86d9171
These are simple enough that I grouped them into a single PR. I can split into two, if necessary.
cc @suharshs @dguerra