Issue #1380: Add SP test checking results with no learning and 0 boos… #1381
Conversation
There was a problem hiding this comment.
I am also investigating this issue, so it's great to see some other confirmation.
I see it after some time (~10 mins) of running a large (2-4000 cols) SP&TM on quite noisy data.
For me the issue manifests as "the network stops acting" which is caused by weight decay in SP, which I presume is caused by the "random" mistakes in SP output.
| const UInt nColumns = 20; | ||
| SpatialPooler sp; | ||
| sp.initialize({inputSize}, {nColumns}); | ||
| sp.setBoostStrength(0); |
There was a problem hiding this comment.
I'd suggest some explanatory comment here:
"This test verifies possible issue where SP with 0-boosting and learning disabled produces different (random) output, compared to the python SP. "
and add
bool learning = False;
| vector<UInt> out1(nColumns, 0); | ||
| vector<UInt> out2(nColumns, 0); | ||
| sp.compute(input.data(), false, out1.data()); | ||
| sp.compute(input.data(), false, out2.data()); |
There was a problem hiding this comment.
- we'll definitely need more then one iteration, the issue seems to manifest "after some time" (-> an overflow?); I'm suggesting to test on a sequence of
random with a fixed seed - I'm not sure what @scottpurdy was thinking, I'd either compare two c++ SP instances, or better a c++ and py SP.
There was a problem hiding this comment.
I take it back. Neither of that is needed, I can replicate the issue now and will post the test-case. Seems to be related to initial params (and initialization via constructor vs. sp.setXXX())
|
Moving to #1382 |
…ting
Tests verifying the issue.
Fixes: #1380