Multidimensional learning #22
Conversation
|
You'd have to ask @celiasmith about that for loop (since according to I'll try to review this today. |
There was a problem hiding this comment.
I've finished my online review. I started trying to rebase this, but I ran into a conflict that I don't know how to resolve. Commit 24ed25f consolidated all error information into one message, and I'm not sure how this will work with multidim errors (I'm not familiar enough with that part of the code). So I'd like to leave this for someone else to rebase. Once it's rebased, I'm happy to help make some of the changes I've suggested.
| @@ -460,6 +460,7 @@ def build_connection(model, conn): | |||
| if not conn.solver.weights: | |||
| needs_interneurons = True | |||
| else: | |||
| # TODO: Does this code ever run? What case is this for? | |||
There was a problem hiding this comment.
This is for if pre_obj is Neurons. I'll add a note.
There was a problem hiding this comment.
This is actually fixed in master.
| self.loihi.nengo_io_h2c.write(1, [0]) | ||
| self.loihi.nengo_io_h2c.write(1, [0]) | ||
| self.loihi.nengo_io_h2c.write(1, [0]) | ||
| # TODO: fix this when are the close writes? |
There was a problem hiding this comment.
This phrase does not make sense to me. Could you elaborate?
| # TODO: fix this when are the close writes? | ||
| # This needs to be the number of dimensions+2 | ||
| # I've just set it big so we can use lots of dims | ||
| self.loihi.nengo_io_h2c.write(25, [0]*25) |
There was a problem hiding this comment.
Could we do this in loihi.wait_for_completion(), and there maybe we'd be able to figure out the number of dimensions?
| probe = receiver.target | ||
| conn = self.model.probe_conns[probe] | ||
| dec_cx = self.model.objs[conn]['decoded'] | ||
| # TODO: Assumes 1 chip |
There was a problem hiding this comment.
Can we have an assert for this? Then it will break immediately if that assumption changes.
| @@ -25,15 +25,17 @@ void nengo_io(runState *s) { | |||
| } | |||
|
|
|||
| //Communicate with learning snip | |||
| int error[2]; | |||
| //assumes only 1 core is learning | |||
There was a problem hiding this comment.
Should this be asserted somewhere? Maybe it is?
| @@ -3,63 +3,69 @@ | |||
| #include "nengo_learn.h" | |||
|
|
|||
| void nengo_learn(runState *s) { | |||
| int core = s->userData[0]; | |||
| int error = (signed char) s->userData[1]; | |||
| int n_errors = s->userData[0]; | |||
There was a problem hiding this comment.
We use the terms error and n_errors a lot of places (not new to this PR, I think), and it's not clear to me what it means. I'm wondering if we could have something more specific. Are these error traces? Could we say n_error_traces or something?
| print('saving...') | ||
| plt.savefig('pes_learning_snips2.png') | ||
| plt.savefig('pes_learning_snips{}d.png'.format(d)) |
There was a problem hiding this comment.
Should the end goal be to turn this file into a test? That seems like something that I could work on.
|
Closing this in favor of #89. |
This is, I believe, the branch that is closest to working for multidimensional learning. The last time it was tested there were other issues preventing it from working, but now it should be possible to test this out through the
learn_nd.pyfile.I also had an unanswered question about the for loop in
simulator.pystarting L424: