Closed rcoreilly closed 1 year ago
false alarm -- nmda variation again it seems. the real culprit was dwt submean.
really need this test -- it looks like Lvis (which massively uses pooled inhibition) is behaving very differently on GPU in layer V2 which depends critically on pooled inhib.
fix for divergence in v2: 9448f4d94da2326b2c3c4a791bbf662f28cfcf3e
Now Lvis CPU vs. GPU are nearly identical, except for the the critical output layer when increasing NData higher. crazy.
did thorough test in pool_test with large layers and ndata = 8 and GABAB, NMDA set to 0 (to avoid contamination) -- Gi does not diverge, but other things are a bit diff still (INet) -- this is just an unavoidable issue apparently: #237
found the issue affecting Lvis -- this should be all good now and test is robust with large scale test.
shows up in the bench_objrec model. None of our tests use pool 4D layers -- will write one for this.