Its saved weights load fine but genuinely reconstruct worse than the
mean baseline (likely an under-trained or abandoned snapshot) -- a data
problem, not a code bug, and retraining is out of scope for now. Add a
KNOWN_ISSUES map so cases like this report SKIP with a reason instead of
FAIL, and stop counting toward the suite's failure exit code, while
still being visible in the output. norb_small_16h_v2/many (missing
weight files entirely) remain plain FAILs since that wasn't in scope
for this change.
9 passed, 1 known issue, 2 failed (12 total).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016K8Gu7Qejd11JbdiHZqYAs
Single-layer Deep stack (28x28->256), has saved weights and a larger
training batch than the other mnist variants (test run took noticeably
longer). Passes: reconstruction error 0.005 vs. mean-baseline 0.067.
9/12 pass.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016K8Gu7Qejd11JbdiHZqYAs
Single-layer Deep stack (28x28->32), weights load fine but genuinely
fail the reconstruction check: error 0.231 vs. mean-baseline 0.067,
worse than the 1.5x tolerance. Unlike norb_small_16h_v2/many (which fail
because no current-format weights exist at all), this one has weights
that load successfully but just don't reconstruct well -- training
params look ordinary (RaoBlackwell, 1000 epochs) and the training data
looks like normal-scale MNIST pixels, so this looks like a genuinely
under-trained or abandoned saved snapshot rather than a test bug.
7/11 pass.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016K8Gu7Qejd11JbdiHZqYAs
101-layer Deep stack (64->64 x100). Fails the same way as
norb_small_16h_v2: only legacy *.weights.dat files exist (the pre-split
combined-weights format current code never reads), no .w/.bh/.bv.dat for
any layer -- a pre-existing data gap, not a bug. 6/8 pass.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016K8Gu7Qejd11JbdiHZqYAs
2-layer DeepStack (9x9->24, 24->4), has saved weights and training data.
Passes: reconstruction error 0.0007 vs. mean-baseline 0.14 -- exercises
upDownPass through more than one layer for the first time in the suite.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016K8Gu7Qejd11JbdiHZqYAs
The old main.cpp (TEST target) called APIs that no longer exist
(DeepStack::load()/loadWeights()/loadTrainingBatch() with no args,
Layer::upDownPass()) -- it hadn't compiled against the current AStack/
Layer API in a long time and was a manual smoke-test script, not an
actual test suite.
Replace it with a small dependency-free runner matching this repo's
existing no-framework style: for each named project, load it, load
weights and training batch, run a full up-pass/down-pass reconstruction
via DeepStack::upDownPass, and check the reconstruction error is finite
and beats the trivial per-feature-mean baseline (a project-agnostic sanity
bound, rather than a hardcoded threshold). Prints PASS/FAIL per project.
Starting coverage: 1-hot, prims, norb_small_16h_v2, mnist_2. 3/4 pass;
norb_small_16h_v2 fails because it has no saved weight files at all
(only .prj/.training.dat/.test.dat) -- a pre-existing gap in that
fixture's data, not a bug introduced here.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016K8Gu7Qejd11JbdiHZqYAs
- load / store training batch with context
- on load: add context part to legacy training batches
- removed Rbm::setBatch()
git-svn-id: http://moon:8086/svn/software/trunk/projects/Rbm@790 b431acfa-c32f-4a4a-93f1-934dc6c82436
jensandClaude Sonnet 5