Add gpu support for LRN #211
Comments
|
Thanks for this feature request! In the meantime, if you want to test your model without explicitly specifying the device for each op, you can do: sess = tf.Session(tf.ConfigProto(allow_soft_placement=True))...when constructing your session. This allows you to request a GPU device for any of your ops, and it will fall back to running on a CPU if there is no GPU kernel available. |
|
Just in case anyone is using this code, theres a typo. It should be: |
|
just ran into this, anyone working on it? |
|
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/stream_executor/dnn.h#L822 does exist, if someone wants to plumb that call through, I think we'd have LRN support for GPU :) |
|
Any updates on this? not having this on GPU really slows things down. |
|
Right, cudnn has an LRN implementation, so the point is to have it use that one. |
|
FYI to anyone looking at contributing a fix for this -- there is some movement internally. |
|
Thanks, removing the contributions welcome tag, since this is being worked on. |
|
@vrv: Does the person working on it have a Github username? |
|
I don't know, @rryan do you know? |
|
Once Stream Executor support is finished I was planning to work on it -- so you can assign me if you'd like :). |
|
@rryan: Assigned, thanks! |
|
Is there any news on this bug? Should I better compute it with different operations? |
|
Sorry for the delay, I carved out some time to work on it this weekend and will be sending it out for review as soon as the GPU tests pass. |
|
Thanks @rryan , looking forward! |
|
Added in 35df3ed. |
|
Yay! |
When we removed a pattern, we removed it from worklist but not from worklistMap. Then, when we tried to add a new pattern on the same Operation again, the pattern wasn't added since it already existed in the worklistMap (but not in the worklist). Closes #211 PiperOrigin-RevId: 277319669 Change-Id: I4d919ea19eb5ef229b1ee001ddd26faba26e6879
When we removed a pattern, we removed it from worklist but not from worklistMap. Then, when we tried to add a new pattern on the same Operation again, the pattern wasn't added since it already existed in the worklistMap (but not in the worklist). Closes #211 PiperOrigin-RevId: 277319669
…pstream-rccl-distribute add rccl to tf.contrib.distribute
When I launch a network that uses local response normalization, it works perfectly on a CPU, but it appears to not have a gpu implementation and results in the following error when I switch to a gpu device:
tensorflow.python.framework.errors.InvalidArgumentError: Cannot assign a device to node 'LRN': Could not satisfy explicit device specification '/gpu:0' [[Node: LRN = LRN[alpha=0.0005, beta=0.75, bias=2, depth_radius=5, _device="/gpu:0"](conv1/conv1)]]I could probably explicitly deploy this operation on the cpu, but it since this is a sliding window algorithm I'm surprised it doesn't have a gpu implementation.
The text was updated successfully, but these errors were encountered: