GDriveData

[pauladkisson]

Notes for the new data shared over Google Drive

See #20 for description from Datta Lab.

[pauladkisson]

High Level Organization

• We have a list of folders labeled _aggregate_results_arhmm_***, whose mapping to experiment_type can be found in the analysis_configuration.TEMPLATE file in the code repo.
  • experiment folders are NOT split by animal (i.e. multiple mice are contained in each)
• Each folder has 3 example session folders labeled w/ a date ex. session_20210215162554-455929

Kai's Description:

Each session directory contains:

• proc, directory – Output of moseq-extract.
• depth.avi – Raw, lossless-compressed depth frames.
• depth_ts.txt – Timestamps for depth frames.
• feedback_ts.txt – Timestamps at which the target syllable was detected, and a boolean denoting whether or not reinforcement was given.
• A number of other intermediate processing files.

If the experiment included photometry, the session folder will contain:

• ir.avi – Infrared video containing LED flashes used to align video stream with photometry data.
• ir_ts.avi – Timestamps for frames in ir.avi.
• tdt_data_*.dat – Raw recording from the TDT.
• tdt_data_*.json - Metadata from the TDT recording.

[pauladkisson]

Moseq Output in proc directory

• bground.tiff = looks like an average of the video --> background
• first_frame.tiff = first frame of depth video
• roi_00.tiff = mask of the operant space in the video
• results_00.mp4 = false color depth video of mouse with auto extracted/oriented mouse in the top left corner
• results_00.h5 = Moseq results from the session including:
  • kinematic variables (centroid_x_mm, centroid_y_mm, etc.) under scalars
  • session metadata (uuid, SessionName, etc.) under metadata/acquisition
  • MoSeq parameters under metadata/extraction/parameters
  • frames = extracted video frames of isolated mouse corresponding to upper left corner of results_00.mp4 (Tx80x80)
  • timestamps = same as depth_ts.txt = semi-regular timestamps for each frame

Conclusion: The .h5 file may effectively replace the kinematic data from the .parquet files since it is non-redundant and also comes w/ relevant MoSeq parameters. Might want to develop a MoSeqInterface in neuroconv like DeepLabCutInterface (depending on usage outside Datta Lab).

[pauladkisson]

Video Data

• depth.avi, results_00.mp4, ir.avi, and frames from results_00.h5 all show different views of the same session
• depth.avi is very low contrast --> the moseq output probably combines the depth video with the regular video

Conclusion: Can use VideoInterface from neuroconv (potentially w/ some modifications) for original depth.avi and ir.avi files. Can store .mp4 version as a separate file w/ a disclaimer that it should only be used for qualitative inspection.

[pauladkisson]

Metadata

Several metdata files throughout the directory

• metadata.json = metadata related to depth video (DepthDataType, etc.) and session (SessionName, etc.)
• status.yaml = various metadata from the session relating to triggering, MoSeq params, syllables
• proc/results.yaml = MoSeq Metadata = redundant w/ .h5 file

Conclusion: Everything work keeping should be in the moseq .h5 file (I think).

[pauladkisson]

TDT Data

• can load in raw data with provided function from datta lab
• tstep = timestamps
• clock = original integer clock
• sync = IR LED synchronization pulses
• pmt00 = has peak in spectral power at 181Hz --> raw signal?
• pmt00 = has peak in spectral power at 381Hz --> raw reference?
• pmt00_x = much more regular sinusoid w/ peak at 181Hz --> commanded voltage for blue LEDs?
• pmt01_x = much more regular sinusoid w/ peak at 381Hz --> commanded voltage for UV LEDs?
• pmt00_y and pmt01_y = peak around 30Hz --> ???

Conclusion: Add raw photometry to FPInterface. Double check interpretation w/ Datta Lab.

[pauladkisson]

Parquet tables

In the photometry sessions,

• photometry_df.parquet and photometry_df_v2.parquet have aligned photometry data
• alignment_df.parquet have alignment info including IR indices from tdt and video

Conclusion: Could be useful for verifying output.

[pauladkisson]

TXT Files

• depth_ts.txt, ir_ts.txt = semi-regular (with some skips) timestamps
• predictions.txt = syllable prediction from MoSeq (offline or online?)
• pc_scores.txt = 10 PC values of depth used to train MoSeq --> ignore bc derived

Conclusion: timestamp info should be in .h5 file, predictions need to be included in the MoSeqInterface, PCs can be ignored.