Conversation
|
|
So when I re-enable the |
hunse
force-pushed
the
cifar-notebook
branch
2 times, most recently
from
55e9410 to
65f7b27
Compare
|
Ok, this should be ready to go. The failing test was just because we weren't upping the snip spikes per step (maybe that default should be higher? Also, we've got quite a few mostly useless warnings now that get spewed for most tests, so we might want to clean those up to make real warnings like this one more obvious). It's ready for review, and I've rebased it so hopefully everything passes. As for squashing it, the first three commits can be squashed no problem. The "Added ImageShifter" commit could go with the notebook commit. The last commit could probably go with the first ones, since it's just enabling tests that relate to that functionality. |
|
The "docs" test is failing because it can't run the new notebook, because it needs the GPU TensorFlow. The CPU TensorFlow doesn't support having the channels first in conv, which is what I use here. We might be able to write our TensorFlow code to get around this, basically by explicitly transposing things if we realize we're on the CPU. Thoughts? |
tbekolay
force-pushed
the
cifar-notebook
branch
3 times, most recently
from
e22cfc6 to
b81f32b
Compare
tbekolay
left a comment
tbekolay
left a comment
There was a problem hiding this comment.
Added some fixups and separate commits for the C file. Not sure if the commits to the C file should be separate or squashed into one 🤷♂️ But with those changes, this looks good to me.
|
Note: merging this is blocked by nengo/nengo-examples#19 as we link to the example added in that PR here. |
|
All @tbekolay's additions/changes look good to me. Just two minor comments. |
This commit also makes spike packing more efficient, and removes the 'xfail' mark from tests that require population spikes.
Instead of having to comment / uncomment print statements.
This value should not change, but it makes the code that uses it more readable.
No camel case for our own code (unavoidable for NxSDK things), and one variable definition per line.
tbekolay
force-pushed
the
cifar-notebook
branch
from
a2426fa to
d0bfed1
Compare
3 participants
This has CIFAR-10 running on Loihi. It replaces #117.
The main feature addition here is allowing population spikes on inputs. This allows us to re-use axons in the three-channel CIFAR-10 input images, making it much easier to get things on the chip.
A lot of these commits can probably be squashed together. The script in the sandbox should also be moved to the sandbox repository.