itproject-code/pipeline/feature_extractors.py

from abc import ABC
import json
import logging
import os
import random
import subprocess
from ast import literal_eval
from pipeline.utils import SourceMedia, Source, Feature, Interval

# for loudness detection
import soundfile
import pyloudnorm

# for word detection
from faster_whisper import WhisperModel, BatchedInferencePipeline

logger = logging.getLogger(__name__)

class FeatureExtractor(ABC):
    """Feature extractor interface."""
    def _run_get_output(self, cmd: list, cwd:str=".") -> str:
        """Run a command and return the output as a string

        Defined to be mocked out in tests via unittest.mock.patch
        """
        return subprocess.run(cmd, stdout=subprocess.PIPE, cwd=cwd).stdout.decode("utf-8")

    def setup(self):
        """Setup the feature extractor -- validate input files & config"""

    def run(self):
        """Run the feature extractor -- extract features"""

    def teardown(self):
        """Teardown the feature extractor -- clean up any temporary files created during setup"""

class LaughterFeatureExtractor(FeatureExtractor):
    """Feature extractor for laughter detection.

    This class is responsible for extracting features corresponding to laughter in media files. Uses jrgillick's laughter-detection library.

    Here:

     setup() (not needed for laughter-detection, as it can work with AV files directly)

     run() is used to extract features from the audio using jrgillick's laughter-detection

     teardown() (not needed)

    @see: https://github.com/jrgillick/laughter-detection for the laughter-detection library
    """
    _PREPEND_TIME = 7.0 # seconds before the laugh to capture whatever was funny
    _APPEND_TIME = 3.0 # seconds after the laugh to capture the reaction
    _CONFIG_LAUGH_DETECTOR_DIR = "/home/robert/mounts/980data/code/laughter-detection/"


    def __init__(self, input_files=None, config=None):
        """It is expected that input_files is a SourceMedia object"""
        self.input_files = input_files
        self.config = config
        self.features = []

    def _laughdetect(self, audio_file, laugh_detector_dir=_CONFIG_LAUGH_DETECTOR_DIR) -> list:
        """Run laughter detection on the audio file

        Returns a list of 2-tuples, each representing a laugh instance in the audio file
        in the format: (start, end) in seconds
        """
        laugh_detector_script = "segment_laughter.py"
        # fake output for testing
        # laugh_detector_path = "tests/fake_segment_laughter.py"
        laugh_detector_cmd = ["python", f"{laugh_detector_dir}{laugh_detector_script}",
                              f"--input_audio_file={audio_file}"]

        # run command, capture output, ignore exit status
        # use self._run_get_output to allow mocking in tests
        laugh_output = self._run_get_output(laugh_detector_cmd, laugh_detector_dir)

        # ↑ have to include cwd to keep laughter-detection imports happy
        # also, it isn't happy if no output dir is specified but we get laughs so it's grand

        # laughs are lines in stdout that start with "instance:", followed by a space and a 2-tuple of floats
        # so jump to the 10th character and evaluate the rest of the line
        return [literal_eval(instance[10:])
                for instance in laugh_output.splitlines()
                if instance.startswith("instance: ")]

    def _adjust_features(self) -> None:
        """Adjust features according to config

        Generically, this ensures features conform to config - min/max feature length, etc.

        In the context of LaughterFeatureExtractor, there is some secret sauce: things that
        cause a laugh generally /precede/ the laugh, so we want more team before the detected start
        than at the end. For example, for a minimum feature length of 15s, we might prepend 10 seconds,
        and append 5 seconds (for example), or 12s and 3s. We may wish to do this pre/post adjustment
        for all laughter features found, regardless of length.

        """
        for feature in self.features:
            # do the pre & post adjustment
            feature.interval.move_start(-self._PREPEND_TIME, relative=True)
            feature.interval.move_end(self._APPEND_TIME, relative=True)

    def setup(self):
        """Setup the laughter feature extractor -- not needed.

        jrgillick's laughter-detection library can work with AV files directly!
        """
        if not self.input_files or len(self.input_files) == 0:
            raise ValueError("No input files provided!")


    def run(self):
        """Extract laughter features for each input file.

        Heavy lifting is performed in _laughdetect()

        Tuples from _laughdetect are used to create Feature objects, which are appended to self.features by convention

        @see: utils.py:Feature, Interval
        """
        if self.input_files:
            for file in self.input_files:
                laughs = self._laughdetect(file.path)
                for laugh in laughs:
                    start, end = laugh
                    self.features.append(Feature(interval=Interval(start=start, end=end),
                                                 source=file, feature_extractor="laughter"))

            # adjust features
            self._adjust_features()


    def teardown(self):
        """No cleanup needed!"""


class RandomFeatureExtractor(FeatureExtractor):
    """Feature extractor for random feature generation.

    This class is responsible for generating random features for testing purposes.

    Here:

     setup() is not needed

     run() is used to generate random features

     teardown() is not needed
    """
    NUM_FEATURES = 30
    MAX_DURATION = 15.0
    MIN_DURATION = 5.0

    def __init__(self, input_files=None, config=None):
        """It is expected that input_files is a SourceMedia object"""
        self.input_files = input_files
        self.config = config
        self.features = []

    def setup(self):
        """Setup the random feature extractor -- validate input files & config"""
        logger.debug("RandomFeatureExtractor setup")

        # Validate input files
        if not self.input_files:
            raise ValueError("No input files provided")

    def run(self):
        """Generate random features for each input file"""
        # check self.input_files is of type SourceMedia
        if not self.input_files or not isinstance(self.input_files, SourceMedia):
            raise ValueError("No input files provided")

        for file in self.input_files:
            for _ in range(self.NUM_FEATURES):
                # determine duration between MIN and MAX, round to 3 decimal places
                duration = round(random.uniform(self.MIN_DURATION, self.MAX_DURATION), 3)

                start = random.random() * file.duration() - duration
                self.features.append(Feature(interval=Interval(start=start, duration=duration),
                                             source=file, feature_extractor="random"))

    def teardown(self):
        pass


class LoudAudioFeatureExtractor(FeatureExtractor):
    """Feature extractor for loud audio detection.

    This class is responsible for extracting features corresponding to loud audio in media files.

    Here:

     setup() is used to validate input files & config, and extracting audio

     run() uses pyloudnorm to detect loud audio

     teardown() is used to clean up temporary files created during setup (if specified by config)
    """
    _CONFIG_DEFAULT_NUM_FEATURES = 15 # keep the top 5 loudnesses
    _CONFIG_DEFAULT_MIN_DURATION = 5.00 # seconds
    def __init__(self, input_files=None, config=None,
                 num_features=_CONFIG_DEFAULT_NUM_FEATURES,
                 min_duration=_CONFIG_DEFAULT_MIN_DURATION):
        if not input_files:
            raise ValueError("No input files provided!")
        self.input_files = input_files
        self.config = config
        self.features = []
        self._num_features = num_features
        self._min_duration = min_duration

    def _audio_file_from_path(self, path: str) -> str:
        """Return the audio file path given a video file path

        Example:
          - in = "/path/to/video.mp4"
          - out = "/tmp/video.mp4.wav"
        """
        OUTPUT_DIR = "/tmp"
        return f"{OUTPUT_DIR}/{os.path.basename(path)}.wav"

    def _get_loudnesses(self, data, meter, rate, window_size, stride_size):
        """Extract loudnesses from the audio data using pyloudnorm

        return a list of 2-tuples, each representing a timecode and loudness value
        """
        loudnesses = []

        for w in range(0, len(data)-window_size, stride_size):
            window = data[w:w+window_size, 0:2] # extract window
            loudnesses.append( (w/rate, meter.integrated_loudness(window)) )

        return loudnesses

    def _loudnorm(self, audio_file):
        """Run pyloudnorm on the audio file"""
        data, rate = soundfile.read(audio_file) # load audio (with shape (samples, channels))
        meter = pyloudnorm.Meter(rate=rate,block_size=0.3) # create BS.1770 meter

        loudness_features = []
        window_size = int(rate * 0.5) # 500ms
        stride_size = int(rate * 0.5) # 500ms -- no overlap

        # for w in range(data.shape[0]//100):
        #     loudnesses.append(meter.integrated_loudness(data[w:w+int(0.3*rate),0:2]))
        loudnesses = self._get_loudnesses(data, meter, rate, window_size, stride_size)

        for timecode, loudval in sorted([l for l in loudnesses if float(l[1]) != float("-inf")], key=lambda x: x[1], reverse=True):
            # print(f"Timecode: {timecode}, Loudness: {loudval}")
            loudness_features.append((timecode, round(loudval, 3))) # round to 3 DP

        return loudness_features

    def _keep_num(self, features, num=_CONFIG_DEFAULT_NUM_FEATURES, trim_overlap=False) -> list:
        """Keep the top n features (default: 5)

        Approach:
         - for range in 0-n
           + expand the nth top feature to min duration
            (move start back by 0.5*min_duration, end forward by 0.5*min_duration)
           + drop any features that are now in that feature's range (optional)
         - return the top n features

        Each feature is a Feature object, with an Interval object
        """
        keep_features = []
        # ensure features are sorted by score
        features = sorted(features, key=lambda x: x.score, reverse=True)
        for i in range(num):
            current_feature = features.pop(0)
            # expand the feature to min_duration - try and keep centered at current start
            if self._min_duration > current_feature.interval.duration:
                current_feature.interval.move_start(-0.5*self._min_duration, relative=True)
            if current_feature.interval.duration < self._min_duration:
                current_feature.interval.update_duration(self._min_duration)
            keep_features.append(current_feature)
            # drop any features that are now in that feature's range (plus margin)
            # features = [f for f in features if
            #             (f.interval.start < current_feature.interval.start-margin and
            #             f.interval.end > current_feature.interval.start-margin) or
            #              (f.interval.end > current_feature.interval.end+margin and
            #               f.interval.start < current_feature.interval.end+margin)]
            if trim_overlap:
                features = [f for f in features if f.interval.overlaps(current_feature.interval)]

        return keep_features

    def setup(self):
        """extract audio from video files to be processed by pyloudnorm

        TODO: config -- hardcoded for now
        """
        # pyloudnorm expects WAV files
        for file in self.input_files:
            audio_file = self._audio_file_from_path(file.path)
            # ffmpeg -i input.mp4 -vn -acodec pcm_s16le output.wav
            subprocess.run(["ffmpeg", "-y", "-i", file.path, "-vn", "-acodec", "pcm_s16le", audio_file],
                           stdout=subprocess.PIPE, stderr=subprocess.PIPE)

    def run(self):
        """Use pyloudnorm to detect loud audio"""
        for file in self.input_files:
            audio_file = self._audio_file_from_path(file.path)
            loudnesses = self._loudnorm(audio_file)

            features = []
            for time, loudness in loudnesses:
                features.append(Feature(interval=Interval(start=time, duration=0.500),
                                             source=file, feature_extractor="loudness",
                                             score=loudness))

            # prune features list to keep self.num_features
            if len(features) > self._num_features:
                self.features = self._keep_num(features, self._num_features)
            else:
                self.features = features


class VideoActivityFeatureExtractor(FeatureExtractor):
    """Feature extractor for video activity detection.

    This class is responsible for extracting features corresponding to high activity in video files.

    Uses ffmpeg's scdet filter with threshold of zero.

    Here:

     setup() is used to validate input files & config

     run() is used to extract features from the video using OpenCV

     teardown() is used to clean up any temporary files created during setup according to the config

    #TODO: minimum duration -- consider whether to do here, or expand duration post-consolidation
    """
    _CONFIG_DEFAULT_NUM_FEATURES = 15 # keep the top 5 activity moments
    _CONFIG_DEFAULT_MIN_DURATION = 5.00 # seconds
    def __init__(self, input_files=None, config=None,
                 num_features=_CONFIG_DEFAULT_NUM_FEATURES,
                 min_duration=_CONFIG_DEFAULT_MIN_DURATION):
        if not input_files:
            raise ValueError("No input files provided!")
        self.input_files = input_files
        self.config = config
        self.features = []
        self._num_features = num_features
        self._min_duration = min_duration

    def _scdet(self, video_file):
        """Run scdet filter on the video file"""
        ffmpeg_cmd = ["ffmpeg", "-i", video_file, "-vf", "scdet=threshold=0", "-f", "null", "-"]
        # output is of the form:
        #   [scdet @ 0x7f0798003d00] lavfi.scd.score: 0.031, lavfi.scd.time: 23.65
        #   [scdet @ 0x7f0798003d00] lavfi.scd.score: 0.006, lavfi.scd.time: 23.70
        # capture output, extract time & score
        scdet_output = subprocess.run(ffmpeg_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE).stderr.decode("utf-8")
        # extract time & score
        scores = []
        for line in scdet_output.splitlines():
            if "lavfi.scd.score" in line:
                scores.append( (float(line.split(",")[1].split(":")[1]),
                                float(line.split(",")[0].split(":")[1]))
                                )
        return scores

    def _nonoverlap_mean(self, scores, window_size=0.500) -> list:
        """Take the mean of non-overlapping windows of scores

        Input: list of tuples in the format (time, score)
        Output: list of tuples in the format (time, mean_score) (reduced set)
        """
        means = []
        current_window = []
        current_window_start = 0.0

        for time, score in scores:
            if time - current_window_start > window_size:
                # calculate mean of current window
                mean_score = sum([s for _, s in current_window]) / len(current_window)
                means.append((current_window_start, round(mean_score, 3)))
                # reset window
                current_window = []
                current_window_start = time
            current_window.append((time, score))

        return means

    def _drop_lowest(self, scores, percent=33):
        """Drop the lowest n% scores from the list"""
        scores = sorted(scores, key=lambda x: x[1], reverse=True)
        return scores[:int(len(scores) * (percent / 100))]

    def _keep_num(self, features, num=_CONFIG_DEFAULT_NUM_FEATURES, trim_overlap=False) -> list:
        """Keep the top n features (default: 5)

        Approach:
         - for range in 0-n
           + expand the nth top feature to min duration
            (move start back by 0.5*min_duration, end forward by 0.5*min_duration)
           + drop any features that are now in that feature's range (optional)
         - return the top n features

        Each feature is a Feature object, with an Interval object
        """
        keep_features = []
        # ensure features are sorted by score
        features = sorted(features, key=lambda x: x.score, reverse=True)
        for i in range(num):
            current_feature = features.pop(0)
            # expand the feature to min_duration - try and keep centered at current start
            if self._min_duration > current_feature.interval.duration:
                current_feature.interval.move_start(-0.5*self._min_duration, relative=True)
            if current_feature.interval.duration < self._min_duration:
                current_feature.interval.update_duration(self._min_duration)
            keep_features.append(current_feature)
            # drop any features that are now in that feature's range (plus margin)
            # features = [f for f in features if
            #             (f.interval.start < current_feature.interval.start-margin and
            #             f.interval.end > current_feature.interval.start-margin) or
            #              (f.interval.end > current_feature.interval.end+margin and
            #               f.interval.start < current_feature.interval.end+margin)]
            if trim_overlap:
                features = [f for f in features if f.interval.overlaps(current_feature.interval)]

        return keep_features

    def setup(self):
        pass

    def run(self):
        for file in self.input_files:
            scores = self._scdet(file.path)
            means = sorted(self._nonoverlap_mean(scores), key=lambda x: x[1], reverse=True)

            features = []
            for time, score in self._drop_lowest(means, 66):
                features.append(Feature(interval=Interval(start=time, duration=0.500),
                                             source=file, feature_extractor="videoactivity",
                                             score=score))

            # prune features list to keep self.num_features
            if len(self.features) > self._num_features:
                self.features = self._keep_num(features, self._num_features)
            else:
                self.features = features

    def teardown(self):
        pass


class JSONFeatureExtractor(FeatureExtractor):
    """(Re-)create features from a JSON file

    The JSON file can have one of two formats:
      - the format produced by the pipleline (@see: video_producers.py:JSONProducer)
      - a simplified format which is easier for manual creation
    """

    def __init__(self, input_files=None, config=None):
        if not input_files:
            raise ValueError("No input files provided!")
        self.input_files = input_files
        self.config = config
        self.features = []

    def setup(self):
        pass

    def _interval_from_dict(self, d):
        return Interval(start=d["start"], duration=d["duration"])

    def _source_from_dict(self, d):
        return Source(d["source"], d["path"], d["provider"])

    def _read_json_from_file(self, file):
        """Read a JSON file and return the contents

        Method exists to allow for mocking in tests
        """
        with open(file, "r") as f:
            return json.load(f)

    def run(self):

        # only pipeline JSON format for now
        # TODO: add support for simplified format
        for file in self.input_files:
            features_from_json = self._read_json_from_file(file.path)

            for feature in features_from_json:
                self.features.append(Feature(interval=self._interval_from_dict(feature["interval"]),
                                             source=self._source_from_dict(feature["source"]),
                                             feature_extractor=feature["feature_extractor"],
                                             score=feature["score"]))

    def teardown(self):
        pass


class WordFeatureExtractor(FeatureExtractor):
    """Feature extractor for specific word detection (uses Whisper)"""
    # set defaults for whisper settings
    DEFAULT_MODEL_SIZE = "medium"
    DEFAULT_DEVICE = "cpu"
    DEFAULT_COMPUTE_TYPE = "int8"
    DEFAULT_BEAM_SIZE = 5
    DEFAULT_BATCH_SIZE = 16
    DEFAULT_PIPELINE_TYPE = "batched" # or "stream"

    words = []

    def _transcribe(self, model, file, **kwargs):
        """Defined here to allow for mocking in tests"""
        return model.transcribe(file, **kwargs)

    def _whispermodel(self, model_size=DEFAULT_MODEL_SIZE,
                      device=DEFAULT_DEVICE, compute_type=DEFAULT_COMPUTE_TYPE):
        """Defined here to allow for mocking out in tests"""
        return WhisperModel(model_size, device=device, compute_type=compute_type)

    def _batched_inference_pipeline(self, model):
        """Defined here to allow for mocking out in tests"""
        return BatchedInferencePipeline(model=model)

    def __init__(self, input_files=None, config=None):
        if not input_files:
            raise ValueError("No input files provided!")
        self.input_files = input_files
        self.config = config
        self.features = []

    def setup(self, words=[]):
        """Setup the word feature extractor -- validate input files & config

        Whisper expects a list of words to search for in the audio
        """
        logger.debug("WordFeatureExtractor setup")

        # Validate words - raise a notice if none provided
        if len(words) == 0:
            logger.warning("No words provided for detection")
        self.words = words
        # TODO: consider stripping punctuation since Whisper produces words+punctuation
        # and we might want to strip the punctuation there too

    def run(self):
        """Extract features corresponding to supplied target words (defined in setup) for each input file

        Use Whisper to detect words in the audio, then match these to target words and create features

        Note: if no words are supplied we can exit early
        """
        if len(self.words) == 0: return

        if self.DEFAULT_PIPELINE_TYPE == "batched":
            batched = True
        else:
            batched = False

        # no early exit
        # TODO: consider maybe loglevel notice of estimated time! consider also: max execution time config?
        # TODO: config options for model size, device, compute type
        model = self._whispermodel() # NB uses defaults, TODO: add config options

        # NOTE: batched not available on pypi yet at time of writing
        if batched:
            batched_model = self._batched_inference_pipeline(model)

        for file in self.input_files:
            # transcribe the audio file
            if batched:
                segments, _ = self._transcribe(batched_model, file.path, batch_size=self.DEFAULT_BATCH_SIZE)
            else:
                segments, _ = self._transcribe(model, file.path, beam_size=self.DEFAULT_BEAM_SIZE)

            # process the segments
            # segment has: start, end, text
            for segment in segments:
                # check if any of the words are in the segment
                for word in segment.text.split():
                    if word in self.words:
                        self.features.append(Feature(interval=Interval(start=segment.start, end=segment.end),
                                                     source=file, feature_extractor="word",
                                                     score=1.0))