-
Notifications
You must be signed in to change notification settings - Fork 2.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
BadOptimization when using gpuarray backend in DebugMode #5216
Comments
Can you give the full error message? @abergeron, do you have time to look at this? On Fri, Nov 11, 2016 at 4:54 PM, bbudescu notifications@github.com wrote:
|
Hi, sorry for the delays (both the one until answering now, and the one until I'll be able to come up with a minimal scenario in which the error is raised). I'll get back on it as soon as I have the time. This will be another few days, I guess. The model I'm currently trying out is based on this: https://github.com/jocicmarko/ultrasound-nerve-segmentation. As to #5206, as far as I understand, that one happens all the time. Mine only surfaces in DebugMode. If it's of any help, I'm currently experimenting on a ubuntu machine with cuda 7.5 and cudnn 5005 (ubuntu 16.04). |
Do you still have the full error message? Without this, we can't do something about that. |
Hi, as I anticipated, I'd like to apologize for the delay with which I'm writing. Compiling the net in debug mode and with verbose output takes a lot of time, and I didn't get to doing that lately. I didn't take the time again to debug-step to the location I pointed out before, but I assume it's the same, as the stderr output looks the same. Note that I also have saved the whole stdout / stderr output and theano logging verbose output, but the two files have about 14MB, so I just copied the exception's message and a few frames above. If you also need the verbose output or my theano config flags, please let me know. This is the error I get (ignore the "timestamp - WARNING:" and the "(std streams)") : New Graph: Hint: relax the tolerance by setting tensor.cmp_sloppy=1 |
Would any other output help for identifying the problem (e.g. optimizer verbose log / optimizer profiling etc.)? |
There are two different problems here:
This should prevent DebugMode from applying to that internal function: #5387 |
I merged the PR, so updating Theano to the dev version should fix it. If you still see the problem, tell us. |
@abergeron do you have a fix for the second part? That GraphToGPU do a replacement when nothing is changed? |
No I didn't touch that. |
When using the 'cuda' gpu backend in DebugMode function compilation fails here https://github.com/Theano/Theano/blob/c0548210e11603141d70250871de8191e0a37037/theano/gpuarray/dnn.py#L216, because both expected and obtained outputs are PyCapsule objects containing NULL pointers (which, I assume, are supposed to be cudnnHandle_t's), which compare as different. BadOptimization is raised.
The text was updated successfully, but these errors were encountered: