Closed davidparks21 closed 8 years ago
Can you try to run the provided MNIST example? I tried to run it locally it could save and load snapshot without issue and I could not reproduce your error. I am not sure what that error mean, it might be that you are using a different version of JLD that does not haskey
?
I'm testing it further, I wasn't able to reproduce it on linux, I'm going to re-try again on windows.
I haven't been able to reproduce this and I've tested it on windows and linux, so I probably had some other error that I wasn't correlating correctly. closing it.
I'm trying to load a snapshot and continue training from that point, I ran a training of 600 iterations successfully, then tried running the same network setup again. The logs indicate that it found the correct snapshot, but I get an error in loading on
haskey
.I've been looking into it for the past hour, but I'm not quite clear if the problem is with JLD or with something Mocha's code is doing before that point.
The warning text that is precursor to the error,
not present in workspace; reconstructing
is found injld_types.jl
, and seems to suggest that the datatype in the JLD file "isn't supported", though I'm not sure what that means.I wondered if the error makes more immediate sense to you?
Update (31May): This worked fine on my linux cluster, and I noticed it's using memory mapped IO in JLD. I recall there being a warning in the docs saying that memory map IO wasn't working on windows. So that's my best guess right now. I'll test further on windows and update this later, but at least it doesn't look like a systemic bug.