27a1beff70
purged the visualizer
2019-12-15 22:17:46 -08:00
05480606b0
small refactoring and kicked the visualizer again
2019-12-15 09:51:22 -08:00
79e8b9874b
for some reason I dockerized it and it works
2019-12-13 17:35:06 -08:00
4a9e328884
visualizer kinda working, more work needs done
2019-12-13 11:05:38 -08:00
1338df9db2
very nice scalability curve up to 12 procs
...
parameters: wiki, BPE 250, BS 32, TARGET 8.3, the rest is commited
also it might be that more pipelines would work if there were
MANY more pipelines, but that remains to be seen until i buy
myself a Ryzen 3700x and destroy the world with it i guess i
should go to sleep now
2019-12-12 22:26:50 -08:00
2bbb1d243c
stop refactoring and get some stuff huggin' done
2019-12-12 21:10:41 -08:00
966bbc904c
mainly faster startup thanks to 'lazy' loading
2019-12-12 20:02:03 -08:00
d954a9832b
somehow i managed to improve the build system
...
to become even more awesome
2019-12-12 10:17:17 -08:00
0372f0ee9c
using adam seems like a sane idea
2019-12-11 22:44:18 -08:00
598e59ca2e
parallelizing doesn't bring anything
2019-12-11 22:23:11 -08:00
2e5042f0e3
termination seems to be implemented okayish
2019-12-11 15:29:59 -08:00
136e5938c3
side-quest successful, go eat mom's spaghetti
...
the side-quest was to flatten the pipeline so that the only
branching is allowed after batcher
2019-12-11 11:19:56 -08:00
32742059c7
commit working termination code for two nodes
...
before embarking on yet another side-quest
2019-12-11 11:12:32 -08:00
7043b65532
this is the baseline for evaluation
2019-12-11 10:31:16 -08:00
5abe7bb413
a lot of small stuff
2019-12-07 19:56:23 -08:00
997d3a7618
the pipeline kinda works, you can even demo it
2019-12-01 22:04:57 -08:00
b2a4681fb2
implemented some kind of dispatching
2019-12-01 17:49:09 -08:00
bc6d34e253
now it kinda learns again and code is kinda clean
2019-12-01 15:44:34 -08:00
a2ef842ef8
take a more windowed approach
2019-12-01 15:04:59 -08:00
4bf66bf85e
network trains and tests but slow? -- investigate
2019-12-01 11:26:54 -08:00
b115077c3b
now slave correctly gets data
2019-12-01 10:21:58 -08:00
569bbf7397
tokenize FASTER
2019-12-01 09:44:40 -08:00
7409eca38b
added batcher and even made it twice as fast
...
as it was a minute ago
2019-11-30 22:40:23 -08:00
101248965c
tokenizer and batcher are sane and work
...
i wonder how much pain it's gonna cause me to change
something in the code
2019-11-30 22:16:42 -08:00
04c35ed9b6
after literally hours of pain, keras learns mnist
2019-11-29 22:10:25 -08:00
c4a21b6be8
reached some mildly amusing result
...
of slaves having "independent" data sources and master still
training through them
2019-11-27 17:37:40 -08:00
5af90383f0
commiting halfways because i'm feeling stuck
2019-11-26 22:11:54 -08:00
9727499f4d
set up synchronous comm for many slaves
...
and also abusing the c preprocessor
2019-11-26 21:20:40 -08:00
725130e51f
implemented a very smooth frankenstein function
2019-11-25 22:37:34 -08:00
76f8d7dcb6
a great work has been done here
...
now there is no mentioning python in c code (except for the
boilerplate at the beginning, but the rest is nice clean c).
all the bridging is being done in cython (where it belongs).
and no memory leaks so there's that!
2019-11-25 22:01:39 -08:00
253c833b5b
commit small improvements before going craZy
2019-11-25 20:54:27 -08:00
04e0b9829c
well done sending a network over mpi, pat pat
...
now do a lot of nets in parallel and then we'll talk
2019-11-25 20:23:33 -08:00
a80a3de4fa
it started working so commit it
2019-11-24 22:14:26 -08:00
