lazynwb 0.2.59

An attempt to speed-up access to large NWB (Neurodata Without Borders) files stored in the cloud.

lazynwb

Purpose

1. Make NWB table access faster and/or consume less memory by reading only the data required, when it's needed

As of 2025 and pynwb==3.0, there are a couple of ways to access data stored in an NWB file as a DynamicTable (e.g. trials, units):

• get the pandas dataframe for the table and access the desired column
• or access specific columns as arrays from disk

The schema for the units table includes columns for list or nested-list type data, including spike_times, and waveform_mean and waveform_sd which can be large for Neuropixels probes and often not needed for analysis. Reading the entire table into memory may be unnecessary and, especially when reading from NWBs stored in the cloud, can be slow.

Accessing individual columns as arrays, on the other hand, means we no longer have the convenience of a DataFrame.

Ideally, we would filter our table based on metrics in some columns, then access the larger columns for the filtered subset of rows, seemlessly with a single command.

To this end, lazynwb.scan_nwb() provides a polars.LazyFrame() interface to NWB tables, which supports both predicate pushdown and projection of columns.

It also supports reading multiple NWB files in one operation, producing a concatenated table:

import lazynwb
import polars as pl

(
  lazynwb.scan_nwb(
    [nwb_path_0, nwb_path_1, ...],  # single path or iterable
    table_path='/units',             # or '/intervals/trials' etc
  )
  .filter(
    pl.col('activity_drift') <= 0.2,
    pl.col('amplitude_cutoff') <= 0.1,
    pl.col('presence_ratio') >= 0.7,
    pl.col('isi_violations_ratio') <= 0.5,
    pl.col('decoder_label') != 'noise',
  )
  .select('unit_id', 'location', 'spike_times', '_nwb_path', '_table_row_index')
  # _nwb_path and _table_row_index are not columns in the NWB table: they're added to identify source of each row in a table that spans multiple NWBs
)
>>> shape: (101, 4)
┌─────────┬─────────────────────────────────┬─────────────────────────────────┬──────────────┐
│ unit_id ┆ spike_times                     ┆ _nwb_path                       ┆ _table_index │
│ ---     ┆ ---                             ┆ ---                             ┆ ---          │
│ i64     ┆ list[f64]                       ┆ str                             ┆ u32          │
╞═════════╪═════════════════════════════════╪═════════════════════════════════╪══════════════╡
│ 193     ┆ [2722.628735, 2723.620493, … 4… ┆ /data/ecephys_702960_2024-03-1… ┆ 5            │
│ 23      ┆ [1784.801304, 1784.804037, … 3… ┆ /data/ecephys_725805_2024-07-1… ┆ 4            │
│ 0       ┆ [9.2712e6, 9.2712e6, … 9.2731e… ┆ /data/ecephys_737812_2024-08-0… ┆ 0            │
│ 300     ┆ [9.2713e6, 9.2714e6, … 9.2731e… ┆ /data/ecephys_737812_2024-08-0… ┆ 6            │
│ 19      ┆ [6115.424355, 6116.428649, … 7… ┆ /data/ecephys_702960_2024-03-1… ┆ 5            │
│ …       ┆ …                               ┆ …                               ┆ …            │
│ 437     ┆ [581.476385, 598.829113, … 331… ┆ /data/ecephys_666859_2023-06-1… ┆ 40           │
│ 439     ┆ [929.656482, 1134.993272, … 33… ┆ /data/ecephys_666859_2023-06-1… ┆ 41           │
│ 446     ┆ [626.940861, 661.785209, … 331… ┆ /data/ecephys_666859_2023-06-1… ┆ 42           │
│ 449     ┆ [618.939192, 618.991564, … 331… ┆ /data/ecephys_666859_2023-06-1… ┆ 43           │
│ 609     ┆ [594.415999, 646.51812, … 3312… ┆ /data/ecephys_666859_2023-06-1… ┆ 44           │
└─────────┴─────────────────────────────────┴─────────────────────────────────┴──────────────┘

2. Quickly provide a summary of the metadata for all NWB files in a project

lazynwb.get_metadata_df(nwb_paths, as_polars=True)
>>>
```Getting metadata: 100%|█████████████████████| 252/252 [00:17<00:00, 14.51file/s]
shape: (252, 28)
┌────────────┬────────────┬───────────┬───────────┬───┬────────┬───────────┬───────────┬───────────┐
│ identifier ┆ session_st ┆ session_i ┆ session_d ┆ … ┆ weight ┆ strain    ┆ date_of_b ┆ _nwb_path │
│ ---        ┆ art_time   ┆ d         ┆ escriptio ┆   ┆ ---    ┆ ---       ┆ irth      ┆ ---       │
│ str        ┆ ---        ┆ ---       ┆ n         ┆   ┆ null   ┆ str       ┆ ---       ┆ str       │
│            ┆ datetime[μ ┆ str       ┆ ---       ┆   ┆        ┆           ┆ datetime[ ┆           │
│            ┆ s, UTC]    ┆           ┆ str       ┆   ┆        ┆           ┆ μs, UTC]  ┆           │
╞════════════╪════════════╪═══════════╪═══════════╪═══╪════════╪═══════════╪═══════════╪═══════════╡
│ 0514cf12-2 ┆ 2024-08-07 ┆ 713655_20 ┆ ecephys   ┆ … ┆ null   ┆ Sst-IRES- ┆ 2023-11-2 ┆ /data/dyn │
│ 41f-4ab2-a ┆ 19:03:44   ┆ 24-08-07  ┆ session   ┆   ┆        ┆ Cre;Ai32  ┆ 3         ┆ amicrouti │
│ ce9-1c2619 ┆ UTC        ┆           ┆ (day 3)   ┆   ┆        ┆           ┆ 08:00:00  ┆ ng_datacu │
│ …          ┆            ┆           ┆ with b…   ┆   ┆        ┆           ┆ UTC       ┆ be_…      │
│ 5c032dff-e ┆ 2024-12-06 ┆ 743199_20 ┆ ecephys   ┆ … ┆ null   ┆ VGAT-ChR2 ┆ 2024-05-1 ┆ /data/dyn │
│ 04f-4884-9 ┆ 19:06:17   ┆ 24-12-06  ┆ session   ┆   ┆        ┆ -YFP      ┆ 8         ┆ amicrouti │
│ 85d-055ac7 ┆ UTC        ┆           ┆ (day 4)   ┆   ┆        ┆           ┆ 07:00:00  ┆ ng_datacu │
│ …          ┆            ┆           ┆ with b…   ┆   ┆        ┆           ┆ UTC       ┆ be_…      │
│ 4a7e9fdb-4 ┆ 2022-09-27 ┆ 636397_20 ┆ ecephys   ┆ … ┆ null   ┆ C57BL6J(N ┆ 2022-06-0 ┆ /data/dyn │
│ fab-4052-a ┆ 18:36:50   ┆ 22-09-27  ┆ session   ┆   ┆        ┆ P)        ┆ 2         ┆ amicrouti │
│ 7fc-f2d109 ┆ UTC        ┆           ┆ (day 2)   ┆   ┆        ┆           ┆ 07:00:00  ┆ ng_datacu │
│ …          ┆            ┆           ┆ with b…   ┆   ┆        ┆           ┆ UTC       ┆ be_…      │
│ 9b4aab77-5 ┆ 2025-01-16 ┆ 744279_20 ┆ ecephys   ┆ … ┆ null   ┆ Sst-IRES- ┆ 2024-05-2 ┆ /data/dyn │
│ 021-43f3-9 ┆ 22:01:37   ┆ 25-01-16  ┆ session   ┆   ┆        ┆ Cre;Ai32  ┆ 5         ┆ amicrouti │
│ f18-b13291 ┆ UTC        ┆           ┆ (day 4)   ┆   ┆        ┆           ┆ 07:00:00  ┆ ng_datacu │
│ …          ┆            ┆           ┆ with b…   ┆   ┆        ┆           ┆ UTC       ┆ be_…      │
│ b0ba34cb-4 ┆ 2024-04-22 ┆ 706401_20 ┆ ecephys   ┆ … ┆ null   ┆ Sst-IRES- ┆ 2023-10-0 ┆ /data/dyn │
...
│ 971-495d-b ┆ 19:18:59   ┆ 25-03-18  ┆ session   ┆   ┆        ┆ -YFP      ┆ 6         ┆ amicrouti │
│ 6ed-dc7b08 ┆ UTC        ┆           ┆ (day 1)   ┆   ┆        ┆           ┆ 07:00:00  ┆ ng_datacu │
│ …          ┆            ┆           ┆ withou…   ┆   ┆        ┆           ┆ UTC       ┆ be_…      │
└────────────┴────────────┴───────────┴───────────┴───┴────────┴───────────┴───────────┴───────────┘

3. Quickly provide a summary of the contents of a single NWB file

lazynwb.get_internal_paths(nwb_paths[0])
>>> {
  '/acquisition/frametimes_eye_camera/timestamps': <HDF5 dataset "timestamps": shape (267399,), type "<f8">,
  '/acquisition/frametimes_front_camera/timestamps': <HDF5 dataset "timestamps": shape (267204,), type "<f8">,
  '/acquisition/frametimes_side_camera/timestamps': <HDF5 dataset "timestamps": shape (267374,), type "<f8">,
  '/acquisition/lick_sensor_events/data': <HDF5 dataset "data": shape (2734,), type "<f8">,
  '/acquisition/lick_sensor_events/timestamps': <HDF5 dataset "timestamps": shape (2734,), type "<f8">,
  '/intervals/aud_rf_mapping_trials': <HDF5 group "/intervals/aud_rf_mapping_trials" (10 members)>,
  '/intervals/epochs': <HDF5 group "/intervals/epochs" (9 members)>,
  '/intervals/performance': <HDF5 group "/intervals/performance" (21 members)>,
  '/intervals/trials': <HDF5 group "/intervals/trials" (48 members)>,
  '/intervals/vis_rf_mapping_trials': <HDF5 group "/intervals/vis_rf_mapping_trials" (12 members)>,
  '/processing/behavior/dlc_eye_camera': <HDF5 group "/processing/behavior/dlc_eye_camera" (110 members)>,
  '/processing/behavior/eye_tracking': <HDF5 group "/processing/behavior/eye_tracking" (26 members)>,
  '/processing/behavior/facemap_front_camera/data': <HDF5 dataset "data": shape (267204, 500), type "<f4">,
  '/processing/behavior/facemap_front_camera/timestamps': <HDF5 dataset "timestamps": shape (267204,), type "<f8">,
  '/processing/behavior/facemap_side_camera/data': <HDF5 dataset "data": shape (267374, 500), type "<f4">,
  '/processing/behavior/facemap_side_camera/timestamps': <HDF5 dataset "timestamps": shape (267374,), type "<f8">,
  '/processing/behavior/licks/data': <HDF5 dataset "data": shape (2707,), type "<f8">,
  '/processing/behavior/licks/timestamps': <HDF5 dataset "timestamps": shape (2707,), type "<f8">,
  '/processing/behavior/lp_front_camera': <HDF5 group "/processing/behavior/lp_front_camera" (57 members)>,
  '/processing/behavior/lp_side_camera': <HDF5 group "/processing/behavior/lp_side_camera" (57 members)>,
  '/processing/behavior/quiescent_interval_violations/timestamps': <HDF5 dataset "timestamps": shape (131,), type "<f8">,
  '/processing/behavior/rewards/timestamps': <HDF5 dataset "timestamps": shape (130,), type "<f8">,
  '/processing/behavior/running_speed/data': <HDF5 dataset "data": shape (251998,), type "<f8">,
  '/processing/behavior/running_speed/timestamps': <HDF5 dataset "timestamps": shape (251998,), type "<f8">
 }

---

Development

See instructions in https://github.com/bjhardcastle/lazynwb/CONTRIBUTING.md and the original template: https://github.com/bjhardcastle/copier-pdm-npc/blob/main/README.md

notes

• hdf5 access seems to have a mutex lock that threads spend a long time waiting to acquire (with remfile)
• seems to slow down over time in single-threaded loop
  • on laptop, first 5 are fast (2-3 s per iteration) - successive iterations are much slower (>60 s)