You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
SpikeGadgetsRawIO: detect timestamp gaps and expose acquisition timestamps (impl of #1773) #1856
I have been working with SpikeGadgets and I realized that this format also has timestamps in each packet's header (4-byte uint32, see the packet diagram in spikegadgetsrawio.py), but SpikeGadgetsRawIO does not currently read it. We should use it to run an integrity check on parse (like the one added to IntanRawIO in #1484) and to expose the per-packet hardware timestamps via a get_spikegadgets_timestamps() method (parallel to #1652 for Intan). This is the SpikeGadgets-specific implementation of the general gap-handling API in #1773.
AI-assisted reproduction of #1856, checked locally on d50b9b9 (Neo 0.15.0.dev0, Python 3.12.14, NumPy 2.5.4, quantities 0.16.4).
I generated five 340-byte .rec files with one channel at 30 kHz; no external recordings were needed. All five parse without a warning into one segment, and SpikeGadgetsIO produces the same AnalogSignal.times: [0, 33.333, 66.667, 100] µs relative to the first sample.
Stored ticks
Synthetic case
Timing result
[100,101,102,103]
Continuous
Correct relative times
[100,101,103,104]
One missing tick
Last two samples placed 33.333 µs early
[2**32-2,2**32-1,0,1]
Normal uint32 wrap
Correct relative times
[2**32-2,0,1,2]
Missing tick at wrap
Last three samples placed 33.333 µs early
[100,101,0,1]
Counter reset
Silently treated as continuous; elapsed time is unknown
I can help with a focused implementation and synthetic tests. Two scope questions before proposing a change:
Would you prefer (A) native tick access plus default continuity errors and an explicit integrity-check bypass first, or (B) full gap_tolerance_ms segmentation in the same PR? A would change today's read behavior: files with gaps would need an explicit opt-out.
Should the accessor be private, as suggested in the recent General API for handling sample gaps on rawio #1773 discussion, or use the publicget_spikegadgets_timestamps name originally proposed here? Either way, I suggest preserving the original uint32 ticks, with wrap/gap/reset tests, rather than silently unwrapping or interpolating them.
Thanks for asking. The use case I am exploring is timing QC before aligning invasive neural recordings with behavioral labels for decoding: being able to inspect the acquisition ticks, including a nonzero initial tick or a discontinuity, before relying on a nominally regular time axis. This is prospective validation work; I do not currently have a verified real recording showing the missing-tick/reset problem. The examples in my earlier comment are synthetic packet fixtures.
I have now checked the public Neo/Gin fixtures as well. Three complete, structurally valid excerpts contain 8,000, 2 and 6,228 packets. All 14,227 adjacent transitions are +1, and the proposed getter agrees with an independent struct.unpack_from("<I") decode of every stored tick. I did not find an acquisition gap, reset or rollover in these excerpts. Together they contain less than half a second of data, so this is limited negative evidence. Two other downloaded stubs did not pass complete-packet/framing checks and were excluded from the acquisition-timing conclusion.
That is why #1949 is kept as a draft accessor-only step: it exposes the stored ticks without changing segmentation or nominal times, and does not claim to solve real-world gap handling. I would be glad to validate against an affected example if one is available before proposing continuity-check behavior.
Activity
DanceoftheAurora commentedon Nov 6, 2026
AI-assisted reproduction of #1856, checked locally on
d50b9b9(Neo 0.15.0.dev0, Python 3.12.14, NumPy 2.5.4, quantities 0.16.4).I generated five 340-byte
.recfiles with one channel at 30 kHz; no external recordings were needed. All five parse without a warning into one segment, andSpikeGadgetsIOproduces the sameAnalogSignal.times:[0, 33.333, 66.667, 100]µs relative to the first sample.[100,101,102,103][100,101,103,104][2**32-2,2**32-1,0,1][2**32-2,0,1,2][100,101,0,1]For either missing-tick example,
t_stopis 133.333 µs instead of the known synthetic 166.667 µs. The sample values themselves are intact. Minimal synthetic reproducer and measured results.I can help with a focused implementation and synthetic tests. Two scope questions before proposing a change:
gap_tolerance_mssegmentation in the same PR? A would change today's read behavior: files with gaps would need an explicit opt-out.get_spikegadgets_timestampsname originally proposed here? Either way, I suggest preserving the original uint32 ticks, with wrap/gap/reset tests, rather than silently unwrapping or interpolating them.feat: expose native SpikeGadgets timestamps privately
h-mayorquin commentedon Nov 6, 2026
Can I ask you what is your use case? do you have real files that have this problem?
DanceoftheAurora commentedon Nov 7, 2026
Thanks for asking. The use case I am exploring is timing QC before aligning invasive neural recordings with behavioral labels for decoding: being able to inspect the acquisition ticks, including a nonzero initial tick or a discontinuity, before relying on a nominally regular time axis. This is prospective validation work; I do not currently have a verified real recording showing the missing-tick/reset problem. The examples in my earlier comment are synthetic packet fixtures.
I have now checked the public Neo/Gin fixtures as well. Three complete, structurally valid excerpts contain 8,000, 2 and 6,228 packets. All 14,227 adjacent transitions are +1, and the proposed getter agrees with an independent
struct.unpack_from("<I")decode of every stored tick. I did not find an acquisition gap, reset or rollover in these excerpts. Together they contain less than half a second of data, so this is limited negative evidence. Two other downloaded stubs did not pass complete-packet/framing checks and were excluded from the acquisition-timing conclusion.Script, results, source links and file hashes are available for reproduction. No recording bytes were modified or fault-injected in that check.
That is why #1949 is kept as a draft accessor-only step: it exposes the stored ticks without changing segmentation or nominal times, and does not claim to solve real-world gap handling. I would be glad to validate against an affected example if one is available before proposing continuity-check behavior.