Closed resonatevision closed 7 years ago
@resonatevision, thanks for the pull request! Have you tested the changes?
Yes. Tested on Tensorflow 1.0.1. Couldn't get the training to run without the changes. Also, there were some WARNINGS about the deprecating apis this month so I changed those too.
Actually I am having trouble with chainer now. Don't know if it is related to the changes I made to global counter. It gives me I/O error in self._invoke_prefetch() when evaluate_pcp is called. Could be completely different issue. But you can hold off on merging if you have never seen this before. Update: I reverted back the global variable changes. Still seeing the evaluate_pcp problem. It's something else. Any thoughts?
This is not related, I think. As far as I remember your problem is related to chainer multi-threading.
Yes. Exactly. Any thoughts on how you solved it?
I fixed it by replacing the MultiprocessIterator with SerialIterator. Obviously will impact training speed but that's the best I could do for now.
Sorry, I'm quiet busy now with other project, but I will check your pull request and your problem in 2 weeks :)
I upgraded to tensorflow r1.0 using script provided with tensorflow.
Remove deprecated function calls.