ammonkey 3.3.2

A package for organizing, processing and performing markerless 3D tracking on primate behavior videos using DeepLabCut, Anipose, with a user-friendly GUI.

Ambiguous Monkey

A fully rewritten deeplabcut+anipose data managing and processing package for better efficiency and data clarity.

Mel

2025.6

Overview

Currently it runs as separate sub-modules and uses command lines. Previous GUI is kept and still works itself but the file structure is outdated for dlc and anipose steps.

Dependency

Considering that deeplabcut requires conda environment, it is recommended to install this package with deeplabcut in the same conda environment.

See amm-env.yaml. Some syntaxes require Python >= 3.10.

Installation

To install ammonkey, you will need to create a conda env per amm-env.yaml. The new env is named amm by default.

conda env create -f "whatever-path/amm-env.yaml"

Due to the package resolving and installation, this might take a while to finish.

Then the GUI can be started with:

ammonkey

Basic setups can be configured in site-packages/ammonkey/cfgs/amm-config.yaml.

If there is already an env with deeplabcut installed, ammonkey can also be installed via: [this hasn't been tested for compatibility yet]

pip install ammonkey

Workflow

Code examples are in ammonkey_example.py. Here explains the ideas how it works

Some terms

• Experiment note is the source according to which the package processes data.

  • Object is ExpNote()

• DAET

  • refers to a combination of Date-Animal-Experiment-Task;
  • Created by and according to ExpNote
  • Used as a unique identifier of each experiment entry (one row in ExpNote).

• Raw path is the path where raw data is dumped.

  • Must contain string 'DATA_RAW'. This behavior can be changed by modifying .core.expNote.getDataPath().
  • Must follow DATA_RAW/(animal_name)/yyyy/mm/dd/yyyymmdd. This is parsed in ExpNote._parsePathInfo()
  • Data in the path should be kept intact and not allowed to modify for integrity.

• Data path is the path where all processed data will appear.

  • Converted from raw path by just replacing DATA_RAW -> DATA.

• dlc == deeplabcut

• ani == anipose

Steps

1. Data setup
   • Create data folders for storage
2. Vid sync;
3. DLC;
4. Anipose;
5. Collect final CSVs.

Details are covered in sections below and in example code ammonkey_example.py.

Batch process

The modular architecture of this package allows for date-wide processing which can be very useful if one looks back and realize multiple datasets are not processed yet.

• Batch sync: ammonkey/scripts/batch_sync.py;
• Batch DLC: ammonkey/scripts/batch_dlc.py.

Tech Details

Below are module-specific docs generated by chatGPT.

ExpNote / DAET

1. Purpose

ExpNote and DAET together provide a lightweight, hash-friendly way to label, load, and query experimental video-note spreadsheets. DAET is an immutable identifier (Date-Animal-Experiment-Task), while ExpNote is a high-level interface that turns an Excel sheet plus a folder tree into easily filterable Python objects.:contentReference[oaicite:0]{index=0} :contentReference[oaicite:1]{index=1}

---

2. Key Concepts

Concept	What it means
DAET	A frozen dataclass stringified as YYYYMMDD-Animal-Experiment-Task; unique & hashable. Two DAET objects w/ same four fields are exactly the same. Thus, it is not recommended to write same experiment-task fields when taking notes. Upon creation, if there are entries with same four fields, they will be auto-renamed, so will the output folders.
Task	Enum describing experiment categories (TS, BBT, BRKM, PULL, CALIB, ALL).
ExpNote	Loader/manager that converts an Excel notes file into DAET-centric records and video paths.

---

3. Quick-start Example

from pathlib import Path
from ammonkey.expNote import ExpNote, Task

root = Path(r'P:\projects\monkeys\DATA_RAW\Pici\2025\04\20250403')
notes = ExpNote(root)          # auto-parses animal/date and loads Excel
print(notes)                   # ➜ ExperimentNotes(Pici 20250403 with N entries)

# Access all DAETs
for daet in notes.daets:
    print(daet)

# Filter for a specific task
ts_df = notes.filterByTask(Task.TS)
print(ts_df[['daet', 'Experiment']])

# Get video existence map for the first valid entry
first_valid = notes.getValidDaets()[0]
print(notes.checkVideoExistence(first_valid))

---

4. Public API Summary

4.1 class DAET(date:str, animal:str, experiment:str, task:str)

Member	Description
.d	Alias for str(daet)
.year	Four-digit int year
.isCalib	True if "calib" in experiment
fromString(s)	Parse a YYYYMMDD-Animal-Exp-Task string
fromRow(row, date, animal)	Build from a Pandas row inside ExpNote

4.2 class ExpNote(path, *, header_key='Experiment', …)

Member	Use
.daets / getDaets()	List of all DAET objects in the sheet
dupWithBlackList(daets:list[DAET])	Creates new ExpNote object that excludes any entry in the input list.
dupWithWhiteList(daets:list[DAET])	Creates new ExpNote object that keeps only entries in the input list.
getAllTaskTypes() -> list[Task]	Returns all task types involved in this note. e.g. returns [Task.CALIB, Task.TS, Task.PULL]
applyTaskFilter(tasks:list[Task])	Creates new ExpNote object that keeps only entries that match tasks in the list.
checkSanity() -> bool	Returns whether this note looks okay.
is_daet_void(daet) -> bool	Returns whether this DAET entry is marked as void ("do not process" mark)
getRow(daet)	Return Pandas row for an entry (DAET)
getVidSetIdx(daet)	Returns a list[int] of all the videos' index of given DAET.
getVidPath(daet, cam_idx)	Absolute Path to a single video file
checkVideoExistence(daet)	{cam_idx: bool} map of found files. e.g. {1:True, 2:False, 3:True, 4:True}
getValidDaets(min_videos=2, skip_void=True)	DAETs that are ready for processing
getSummary()	Dict with counts for quick reporting

---

5. Typical Workflow

1. Instantiate ExpNote with the raw-data folder that contains an Animal_Date.xlsx.
2. Inspect notes.getSummary() to verify entry counts.
3. Filter/Select DAETs via getValidDaets() or filterByTask().
4. Locate Videos by calling getVidPath() or getVidSetPaths().
5. Process only when checkVideoExistence() confirms required files.

---

6. Error Handling & Logging

• Missing Excel columns raise ValueError; missing files log warnings but do not crash.
• Duplicate DAETs or malformed rows are reported through the module’s logger (__name__).
• Incorrect date formats in DAET raise ValueError at construction time.

---

7. Best Practices

• Keep Excel headers exactly matched; customize header_key only if the sheet layout differs.
• Use the Task enum everywhere instead of raw strings to avoid typos.
• Because DAET is frozen and hashable, it can be used as a dictionary key or set element safely across a session.
• For reproducibility across runs, store the string form of a DAET rather than relying on Python hash() values, which vary per interpreter session.

---

8. Version Compatibility

Both classes rely only on the Python ≥ 3.10 standard library plus pandas ≥ 1.5. If your workflow touches on video paths, ensure the folder layout remains:

root/
│  Animal_Date.xlsx
├─ cam1/
├─ cam2/
├─ cam3/
└─ cam4/

Any deviation requires overriding cam_headers or extending getVidPath.

Video Synchronization

1. Purpose

sync.py implements a complete pipeline to detect and correct timestamp offsets across multi-camera video sets using LED flashes and audio cues. It produces per-DAET JSON configs and optionally writes synchronized output videos. :contentReference[oaicite:0]{index=0}

---

2. Key Components

Class / Function	Role
SyncConfig	Holds thresholds, durations, file extensions and flags for detection & override.
SyncResult	Immutable record of LED starts, audio starts, corrected frames, status & message.
VidSynchronizer	Orchestrates detection (audio + LED), cross-validation, config writing, and final video sync.
syncVideos(...)	Convenience function: prepares folders, invokes VidSynchronizer, returns results.

---

3. Quick-start Example

from ammonkey.expNote import ExpNote, Task
from ammonkey.sync import syncVideos, SyncConfig, CamConfig

notes = ExpNote(r'P:\projects\monkeys\DATA_RAW\Pici\2025\04\20250403')
cam_cfg = CamConfig()                  # default camera ROIs & LED colors
sync_cfg = SyncConfig(override_existing=True)

# Run full sync for Task.BBT only
results = syncVideos(notes,
                     cam_cfg=cam_cfg,
                     sync_cfg=sync_cfg,
                     task=Task.BBT)

# Inspect outcomes
for res in results:
    print(f"{res.daet}: {res.status} — {res.message}")

---

4. Public API Summary

4.1 SyncConfig(...)

• audio_test_duration: seconds for detection window
• led_threshold, cross_validation_threshold: pixel & frame-offset tolerances
• video_extension, output_size: format & resolution for output
• override_existing: reprocess existing detections/syncs

4.2 SyncResult

• .daet: associated DAET
• .led_starts, .audio_starts, .corrected_starts: per-cam frame lists
• .status: 'success', 'warning', or 'failed'
• .message: diagnostic text
• .config_path: JSON file location if created

4.3 VidSynchronizer

• .setROI(daet=None, frame=500): launch ROI selector for LED zones.
• .syncAll(task, skip_existing=True): run detection + video sync for all valid DAETs.
• ._runAudioSync(daets), ._runLEDSync(daet, audio_starts): low-level detectors.
• ._crossValidate(led_starts, audio_starts): align LED vs audio, interpolate or flag errors.
• ._createSyncConfig(...): write per-DAET JSON for SyncLED.process_videos.

4.4 syncVideos(...)

• Prepares folders via dataSetup
• Instantiates VidSynchronizer and runs ROI setup
• Returns list of SyncResult objects

---

5. Typical Workflow

1. Prepare raw folder with Excel notes (ExpNote) and camera subfolders.
2. Import and configure: instantiate SyncConfig (tuning thresholds) and optional CamConfig.
3. Run syncVideos(notes, cam_cfg, sync_cfg, task) to detect & sync.
4. Review console/log for SyncResult statuses.
5. Inspect generated JSON configs in each DAET’s SynchronizedVideos folder.
6. Locate final synced videos under SynchronizedVideos/{group}/ subfolders.

---

6. Error Handling & Logging

• Uses Wood logger to write per-phase logs under notes.data_path.
• Falls back to audio-only if LED fails; marks as warning or failed based on deviation counts.
• Missing videos or ROI configs emit warnings but continue processing other cams.
• Exceptions during detection/sync produce a SyncResult with status='failed' and error message.

---

7. Best Practices

• ROI Configuration: call .setROI() once interactively before batch runs.
• Threshold Tuning: adjust led_threshold and cross_validation_threshold on a small sample to minimize false detections.
• Data Organization: maintain consistent camera folders (cam1–cam4) and file naming to match ExpNote conventions.
• Idempotency: use override_existing=False in SyncConfig to skip already processed entries.

---

8. Version Compatibility

• Python ≥ 3.10, pandas ≥ 1.5, plus dependencies for SyncLED, SyncAud, and ROIConfig.
• Relies on the same folder layout as ExpNote.
• Tested on Windows & Linux file paths but uses pathlib for cross-platform support.

DLC Analysis

1. Purpose

dlc.py defines DLCModel and DLCProcessor to run DeepLabCut pose estimation on synchronized video folders. dlcCollector.py adapts and merges per-model outputs into a single folder for downstream analysis.

---

2. Key Components

Component	Role
initDlc()	Lazy-imports DeepLabCut, sets ready flag.
DLCModel	Frozen dataclass wrapping model config; methods: runOnce(), pee(), path helpers.
DLCProcessor	Orchestrates batch processing: loads sync roots, skips calibration, merges outputs.
mergeDlcOutput(*folders)	Combines separate model outputs into a merged folder, writes combined inherit.json.
copyH5(src, dst)	Copies filtered H5 files for merging.
getDLCMergedFolderName(...)	Generates merged folder names from per-model final names.

---

3. Quick-start Example

from ammonkey.expNote import ExpNote
from ammonkey.sync import syncVideos, Task
from ammonkey.dlc import initDlc, createProcessor_BBT

# 1. Initialize DLC
initDlc()

# 2. Load experiment notes and synchronize videos
notes = ExpNote(r'P:\projects\monkeys\DATA_RAW\Pici\2025\04\20250403')
syncVideos(notes, task=Task.BBT)

# 3. Create processor for BBT and run batch DLC
processor = createProcessor_BBT(notes)
results = processor.batchProcess()  # dict of DAET → success flag

---

4. Public API Summary

initDlc() → int

Imports DeepLabCut; returns 1 on success, 0 on failure.

class DLCModel

• .runOnce(vid_path, override_exist=True) → bool — run DLC on one folder.
• .md5_short, .final_folder_name — unique identifiers.
• .information() → list[str] — diagnostic info.
• .pee(vid_path) — isolates output files, writes logs and inherit.json.

class DLCProcessor

• .analyzeSingleDaet(daet) → bool — process one DAET.
• .batchProcess(daets=None, min_videos=2) → dict[DAET, bool] — loop over DAETs.

mergeDlcOutput(*folders) → int

Validates input folders, copies H5 files, writes merged inherit.json.

---

5. Typical Workflow

1. Call initDlc() at startup.
2. Instantiate ExpNote and synchronize videos with syncVideos().
3. Create a DLCProcessor via a factory (createProcessor_TS/Pull/BBT/Brkm).
4. Run processor.batchProcess() to perform DLC analysis.
5. Find merged outputs under {DAET}/DLC/{merged_folder}.

---

6. Error Handling & Logging

• Initialization errors and missing models log via Python logging.
• runOnce() logs and returns False on import errors, missing paths, or analysis failures.
• mergeDlcOutput() raises on nonexistent folders or invalid argument counts.
• All per-DAET logs and merged JSON live under note.data_path / 'SynchronizedVideos'.

---

7. Best Practices

• Use processor factories (createProcessor_*) to ensure consistent model presets.
• Skip calibration DAETs automatically; no extra filtering needed.
• Inspect JSON logs (inherit.json) for reproducibility rather than relying on in-memory hashes.

---

8. Version Compatibility

• Python ≥ 3.10
• DeepLabCut installed in the same environment
• pandas ≥ 1.5 (for ExpNote)
• Compatible with Windows & Linux via pathlib

Anipose Processing & Finalization

1. Purpose

ani.py automates camera-calibration, video preparation, and 3D triangulation via the Anipose CLI. finalize.py gathers all generated 3D-pose CSVs into a clean folder structure and logs the collection.

---

2. Key Components

Component / Function	Role
getH5Rename(file_name, stem_only=False)	Strip DLC postfix from H5 filenames
insertModelToH5Name(...)	(stub) Intended to inject model tag into H5 filename
CalibLib	Indexes historical calibration TOML files and finds closest match by date
AniposeProcessor	Manages per-DAET folder setup, calibration copy, CLI calls (calibrate, triangulate), and log writing
violentCollect(ani_path, clean_path)	Copy all *.csv from pose-3d subfolders into a single "clean" directory
writeProcedureSummaryCsv(note, ani_path, clean_path)	(stub) Intended to summarize DAET procedures into a single CSV

---

3. Quick-start Example

from pathlib import Path
from monkeylab.expNote import ExpNote
from monkeylab.ani import AniposeProcessor, CalibLib
from monkeylab.finalize import violentCollect

# 1. Load notes and configure Anipose
notes = ExpNote(r'P:\projects\monkeys\DATA_RAW\Pici\2025\04\20250403')
processor = AniposeProcessor(note=notes, model_set_name='BBT-LR')

# 2. Prepare calibration files and root config
processor.setupCalibs()
processor.setupRoot()

# 3. Batch-setup 2D outputs and run triangulation
processor.batchSetup(use_filtered=True)
processor.calibrate()       # runs `anipose calibrate`
processor.triangulate()     # runs `anipose triangulate`

# 4. Collect all final CSVs
clean_dir = Path(r'P:\projects\monkeys\CLEAN_CSV')
violentCollect(processor.ani_root_path, clean_dir)

---

4. Public API Summary

getH5Rename(file_name, stem_only=False) → str

Remove auto-generated DLC postfix (_DLC_resnet…_shuffle…) from H5 filenames.

class CalibLib(lib_path: Path)

• .updateLibIndex() — scan *.toml calibration files into a date→list index.
• .lookUp(date: int) → Path | None — return exact or closest‐prior TOML; logs fallback.

class AniposeProcessor(note, model_set_name, ...)

• .setupCalibs() — copy each DAET’s raw videos into per-DAET calibration folders.
• .setupRoot() — copy config.toml and ensure calibration.toml exist in root.
• .setupSingleDaet(daet, use_filtered, copy_videos) — prepare H5, logs, and calibs for one DAET.
• .batchSetup(use_filtered, copy_videos) — run setupSingleDaet for all DAETs.
• .calibrate() / .triangulate() — invoke Anipose CLI via subprocess.
• .pee(daet_root) — append triangulation record to scent.log.

violentCollect(ani_path: Path, clean_path: Path) → None

Copy all CSVs under ani_path/pose-3d into clean_path/{ani_name} and write a timestamped log.

writeProcedureSummaryCsv(...)

(to be implemented) Summarize per-DAET processing steps into a master CSV.

---

5. Typical Workflow

1. Instantiate ExpNote and create AniposeProcessor.
2. Populate calibration library with setupCalibs().
3. Initialize root with setupRoot().
4. Copy filtered or unfiltered H5 and calibration files via batchSetup().
5. Run Anipose CLI commands: .calibrate() then .triangulate().
6. Collect final CSVs using violentCollect() (and later implement writeProcedureSummaryCsv()).

---

6. Error Handling & Logging

• Missing calibration folder raises FileNotFoundError.
• Fallback calibration selection logs a WARNING.
• Copy errors log via logger.error and continue.
• violentCollect skips existing CSVs; raises on invalid paths.

---

7. Best Practices

• Pre-populate calib history with well-named TOML files (calibration-YYYYMMDD…).
• Run setupCalibs() once per model set before any triangulation.
• Verify that config.toml and calibration.toml exist under ani_root_path.
• Use use_filtered=True to process only filtered 2D poses.
• Implement writeProcedureSummaryCsv to automate batch reporting.

---

8. Version Compatibility

• Python ≥ 3.10 (relies on dataclasses, pathlib, subprocess).
• Anipose CLI installed in specified conda_env.
• pandas ≥ 1.5 (for ExpNote integration).
• Compatible with Windows & POSIX paths via pathlib.