The DVS128 Gesture and SSC reference configurations are registered in scripts/train_reference_models.py and the code path is complete and tested, but no event dataset can currently produce an honest number. All three blockers are outside this repository.
| Dataset |
Status |
Detail |
| DVS128 Gesture |
cannot download |
figshare returns HTTP 202 with an empty body; tonic then reports "File not found or corrupted" |
| CIFAR10-DVS |
cannot download |
same figshare behaviour |
| N-MNIST |
cannot download |
tonic's pinned Mendeley URL returns HTTP 404 |
| SHD / SSC |
downloads, decode broken upstream |
tonic scales float16 timestamps by 1e6; the scale factor does not fit in float16, so every timestamp becomes INT64_MIN |
SSC is worked around locally by spikeforge/events/hsd_reader.py, which reads the HDF5 directly and scales in float64. That makes SSC trainable, but the accuracy the stock defaults reach is weak enough not to be worth a catalog slot — fc_legacy β=0.5 reaches 11.88%, β=0.9 reaches 10.22%, and recurrent_net reaches 9.16%, all against a 2.86% chance baseline on 35 classes with loss pinned near ln(35). Three configurations in the same band means it is the defaults, not the configuration.
To revisit, in order:
- Re-test figshare from another network:
curl -sI -r 0-1023 -o /dev/null -w "%{http_code}\n" https://figshare.com/ndownloader/files/38022171. A 206/200 means DVS128 Gesture is unblocked; the run is minutes of CPU (~50 ms/sample at 128×128×2, ~715 KiB checkpoint).
- Watch the upstream tonic issues (drafts in
plans/upstream/) for the float16 fix and the dead N-MNIST URL.
- If DVS128 lands, update the README's "What it scores" framing, which is still implicitly all-MNIST-family.
EventTimestampError refuses a corrupted stream rather than binning every event into one time step, so this cannot regress into a silently wrong published number.
The DVS128 Gesture and SSC reference configurations are registered in
scripts/train_reference_models.pyand the code path is complete and tested, but no event dataset can currently produce an honest number. All three blockers are outside this repository.HTTP 202with an empty body; tonic then reports "File not found or corrupted"HTTP 404float16timestamps by1e6; the scale factor does not fit infloat16, so every timestamp becomesINT64_MINSSC is worked around locally by
spikeforge/events/hsd_reader.py, which reads the HDF5 directly and scales infloat64. That makes SSC trainable, but the accuracy the stock defaults reach is weak enough not to be worth a catalog slot —fc_legacyβ=0.5 reaches 11.88%, β=0.9 reaches 10.22%, andrecurrent_netreaches 9.16%, all against a 2.86% chance baseline on 35 classes with loss pinned nearln(35). Three configurations in the same band means it is the defaults, not the configuration.To revisit, in order:
curl -sI -r 0-1023 -o /dev/null -w "%{http_code}\n" https://figshare.com/ndownloader/files/38022171. A206/200means DVS128 Gesture is unblocked; the run is minutes of CPU (~50 ms/sample at 128×128×2, ~715 KiB checkpoint).plans/upstream/) for thefloat16fix and the dead N-MNIST URL.EventTimestampErrorrefuses a corrupted stream rather than binning every event into one time step, so this cannot regress into a silently wrong published number.