Allow inference on multiple individual videos via sleap-track

[emdavis02]

Discussed in #1439

Problem Description

^{Originally posted by roomrys August 4, 2023}
Currently, to run inference on multiple videos via the sleap-track command, users need to call this command many times either manually or in a script. It might be a nice feature to allow multiple videos/inputs.

Feature Proposal

• Add an option for a folder of files in the cli instead of only an individual file path
• Later expand this to also include a csv of file paths

Implementation Details

Currently, sleap_track takes in an argument data_path from the command line that is the file path to a .slp file, this means that the user must run this function once for every video they wish to run an inference on either manually or through a script. We would like to allow the argument data_path to also take in a path to a folder of .slp files and run an inference on each of these files. This will all be acomplished in sleap/nn/inference.py

1. Create an optional flag to the cli to specify if data_path is a folder

• Add this component to the _make_cli_parser() function
• Also add a description to help docs
  

  sleap/sleap/nn/inference.py

  Lines 5051 to 5068 in 43a4f13

  	def _make_cli_parser() -> argparse.ArgumentParser:
  	"""Create argument parser for CLI.
  	Returns:
  	The `argparse.ArgumentParser` that defines the CLI options.
  	"""
  	parser = argparse.ArgumentParser()
  	parser.add_argument(
  	"data_path",
  	type=str,
  	nargs="?",
  	default="",
  	help=(
  	"Path to data to predict on. This can be a labels (.slp) file or any "
  	"supported video format."
  	),
  	)

2. Make data_path a list to enable iteration

• Check if the flag was called in args to communicate that args.data_path is a folder
• If it is, enter the folder and add all compatible files to data_path
• Else, data_path = [args.data_path]

Note: The "labels" cli argument has been deprecated and will not need to be edited to accomodate this new function.

sleap/sleap/nn/inference.py

Lines 5292 to 5296 in 43a4f13

	labels_path = getattr(args, "labels", None)
	if labels_path is not None:
	data_path = labels_path
	else:
	data_path = args.data_path

3. Add a loop to file loading lines

• Iterate through data_path. The loop will encompass the entire code section shown below.
• Change provider to a list to store a value for each item in data_path
  

  sleap/sleap/nn/inference.py

  Lines 5304 to 5328 in 43a4f13

  	if data_path.endswith(".slp"):
  	labels = sleap.load_file(data_path)
  	if args.only_labeled_frames:
  	provider = LabelsReader.from_user_labeled_frames(labels)
  	elif args.only_suggested_frames:
  	provider = LabelsReader.from_unlabeled_suggestions(labels)
  	elif getattr(args, "video.index") != "":
  	provider = VideoReader(
  	video=labels.videos[int(getattr(args, "video.index"))],
  	example_indices=frame_list(args.frames),
  	)
  	else:
  	provider = LabelsReader(labels)
  	else:
  	print(f"Video: {data_path}")
  	# TODO: Clean this up.
  	video_kwargs = dict(
  	dataset=vars(args).get("video.dataset"),
  	input_format=vars(args).get("video.input_format"),
  	)
  	provider = VideoReader.from_filepath(
  	filename=data_path, example_indices=frame_list(args.frames), **video_kwargs
  	)

4. Add a loop to main() for running inference and tracking

• Iterate through data_path in the section of main shown below. The loop will start at line 5476, before we run the inference but after the predictor.tracker is set.
  

  sleap/sleap/nn/inference.py

  Lines 5473 to 5485 in 43a4f13

  	if args.models is not None:
  	# Setup models.
  	predictor = _make_predictor_from_cli(args)
  	predictor.tracker = tracker
  	# Run inference!
  	labels_pr = predictor.predict(provider)
  	if output_path is None:
  	output_path = data_path + ".predictions.slp"
  	labels_pr.provenance["model_paths"] = predictor.model_paths
  	labels_pr.provenance["predictor"] = type(predictor).__name__

• transplate the following lines of code into the above loop. This will need to be run for each item in data_path
  

  sleap/sleap/nn/inference.py

  Lines 5510 to 5541 in 43a4f13

  	if args.no_empty_frames:
  	# Clear empty frames if specified.
  	labels_pr.remove_empty_frames()
  	finish_timestamp = str(datetime.now())
  	total_elapsed = time() - t0
  	print("Finished inference at:", finish_timestamp)
  	print(f"Total runtime: {total_elapsed} secs")
  	print(f"Predicted frames: {len(labels_pr)}/{len(provider)}")
  	# Add provenance metadata to predictions.
  	labels_pr.provenance["sleap_version"] = sleap.__version__
  	labels_pr.provenance["platform"] = platform.platform()
  	labels_pr.provenance["command"] = " ".join(sys.argv)
  	labels_pr.provenance["data_path"] = data_path
  	labels_pr.provenance["output_path"] = output_path
  	labels_pr.provenance["total_elapsed"] = total_elapsed
  	labels_pr.provenance["start_timestamp"] = start_timestamp
  	labels_pr.provenance["finish_timestamp"] = finish_timestamp
  	print("Provenance:")
  	pprint(labels_pr.provenance)
  	print()
  	labels_pr.provenance["args"] = vars(args)
  	# Save results.
  	labels_pr.save(output_path)
  	print("Saved output:", output_path)
  	if args.open_in_gui:
  	subprocess.call(["sleap-label", output_path])

5. Add an aditional loop to main() for just running tracking

• Iterate through data_path for the following code. The loop will start after the elif and contain the rest of the attatched lines.
  

  sleap/sleap/nn/inference.py

  Lines 5487 to 5500 in 43a4f13

  	elif getattr(args, "tracking.tracker") is not None:
  	# Load predictions
  	print("Loading predictions...")
  	labels_pr = sleap.load_file(args.data_path)
  	frames = sorted(labels_pr.labeled_frames, key=lambda lf: lf.frame_idx)
  	print("Starting tracker...")
  	frames = run_tracker(frames=frames, tracker=tracker)
  	tracker.final_pass(frames)
  	labels_pr = Labels(labeled_frames=frames)
  	if output_path is None:
  	output_path = f"{data_path}.{tracker.get_name()}.slp"

• Again, we will have to transplant the following lines of code into the loop.
  

  sleap/sleap/nn/inference.py

  Lines 5510 to 5541 in 43a4f13

  	if args.no_empty_frames:
  	# Clear empty frames if specified.
  	labels_pr.remove_empty_frames()
  	finish_timestamp = str(datetime.now())
  	total_elapsed = time() - t0
  	print("Finished inference at:", finish_timestamp)
  	print(f"Total runtime: {total_elapsed} secs")
  	print(f"Predicted frames: {len(labels_pr)}/{len(provider)}")
  	# Add provenance metadata to predictions.
  	labels_pr.provenance["sleap_version"] = sleap.__version__
  	labels_pr.provenance["platform"] = platform.platform()
  	labels_pr.provenance["command"] = " ".join(sys.argv)
  	labels_pr.provenance["data_path"] = data_path
  	labels_pr.provenance["output_path"] = output_path
  	labels_pr.provenance["total_elapsed"] = total_elapsed
  	labels_pr.provenance["start_timestamp"] = start_timestamp
  	labels_pr.provenance["finish_timestamp"] = finish_timestamp
  	print("Provenance:")
  	pprint(labels_pr.provenance)
  	print()
  	labels_pr.provenance["args"] = vars(args)
  	# Save results.
  	labels_pr.save(output_path)
  	print("Saved output:", output_path)
  	if args.open_in_gui:
  	subprocess.call(["sleap-label", output_path])

Documentation Changes

Changes will be made to the sleap-track section of the documentation

positional arguments:
data_path Path to data to predict on. This can be one of the following:

• A .slp file containing labeled data.
• A folder containing multiple video files in supported formats.
• An individual video file in a supported format.

optional arguments:
...
-o OUTPUT, --output OUTPUT The output filename or directory path to use for the predicted data. If not provided, defaults to '[data_path].predictions.slp'.

[eberrigan]

Great job @emdavis02!

• We could try it with and without the additional argument to determine if the input is a directory. It might be sufficient to use Path.isdir() and Path.isfile() documented here.
• There is either inference alone, or inference with tracking.
• Make sure our current implementation of the CLI works so that these changes are backwards-compatible.
• Please add examples of intended use cases and test that the new implementation behaves as expected. These examples will be added to the CLI documentation.
• Add necessary tests for inference.py to make sure changes are covered.

[talmo]

• We could try it with and without the additional argument to determine if the input is a directory. It might be sufficient to use os.path.isdir() and os.path.isfile()` documented here.

Just jumping in to say: please use pathlib instead of os.path APIs!

[eberrigan]

Thanks! Here is the correspondence between os.path and Path: correspondence to tools in the os module