itproject-code/pipeline/producers.py

"""Classes for producing videos"""

from abc import ABC
import json
import logging
import os
import subprocess
import tempfile

# for visualisations:
import matplotlib.pyplot as plt

# for encoding as JSON
from pipeline.utils import Feature


class Producer(ABC):
    """Generic producer interface."""
    def __init__(self, features):
        """All producers should take a list of features as input"""
    def produce(self):
        """All Producers should produce something!"""

class VideoProducer(Producer):
    """Video producer interface."""


class FfmpegVideoProducer(VideoProducer):
    """Produce videos using ffmpeg"""
    # TODO: consider output filename options

    def _run_no_output(self, cmd: list, cwd:str=".") -> None:
        """Run a command and return the output as a string

        Defined to be mocked out in tests via unittest.mock.patch
        """
        subprocess.run(cmd, stdout=None, stderr=None, cwd=cwd)

    def __init__(self, features):
        if not features:
            raise ValueError("No features provided")
        # TODO: consider if we want to permit empty features (producing no video)
        self.features = features

    def _ffmpeg_feature_to_clip(self, feature=None, output_filepath=None):
        """use ffmpeg to produve a video clip from a feature"""
        OVERWRITE = True # TODO: consider making this a config option
        if not feature or not feature.interval:
            raise ValueError("No feature provided")

        if not output_filepath:
            raise ValueError("No output filepath provided")

        ffmpeg_prefix = ["ffmpeg", "-y"] if OVERWRITE else ["ffmpeg"]
        ffmpeg_suffix = ["-r", "60", "-c:v", "libx264", "-crf", "26", "-c:a", "aac", "-preset", "ultrafast"]
        # TODO: match framerate of input video
        # TODO: adjustable encoding options
        seek = ["-ss", str(feature.interval.start)]
        duration = ["-t", str(feature.interval.duration)]
        ffmpeg_args = ffmpeg_prefix + seek + ["-i"] + [feature.source.path] +\
            duration + ffmpeg_suffix + [output_filepath]
        logging.info(f"ffmpeg_args: {ffmpeg_args}")
        self._run_no_output(ffmpeg_args)

    def _ffmpeg_concat_clips(self, clips=None, output_filepath=None):
        """use ffmpeg to concatenate clips into a single video"""
        OVERWRITE = True
        ffmpeg_prefix = ["ffmpeg"]
        ffmpeg_prefix += ["-y"] if OVERWRITE else []
        ffmpeg_prefix += ["-f", "concat", "-safe", "0", "-i"]

        # there is a method to do this via process substitution, but it's not portable
        # so we'll use the input file list method

        if not clips:
            raise ValueError("No clips provided")

        if not output_filepath:
            raise ValueError("No output filepath provided")

        # generate a temporary file with the list of clips
        join_file = tempfile.NamedTemporaryFile(mode="w")
        for clip in clips:
            join_file.write(f"file '{clip}'\n")
        join_file.flush()

        ffmpeg_args = ffmpeg_prefix + [join_file.name] + ["-c", "copy", output_filepath]
        logging.info(f"ffmpeg_args: {ffmpeg_args}")
        self._run_no_output(ffmpeg_args)
        join_file.close()

    def produce(self):
        OUTPUT_DIR = "/tmp/" # TODO: make this a config option

        clips = []
        for num, feature in enumerate(self.features):
            output_filepath = f"{OUTPUT_DIR}/highlight_{num}.mp4"
            self._ffmpeg_feature_to_clip(feature, output_filepath)
            clips.append(output_filepath)

        # concatenate the clips
        output_filepath = f"{OUTPUT_DIR}/highlights.mp4"
        self._ffmpeg_concat_clips(clips, output_filepath)
        logging.info(f"Produced video: {output_filepath}")

class VisualisationProducer(Producer):
    """Visualisation producer -- illustrate the features we have extracted"""
    def __init__(self, features):
        if not features:
            raise ValueError("No features provided")
        self.features = features

    def produce(self):
        """Produce visualisation"""
        # basic idea: use matplotlib to plot:
        #  - a wide line segment representing the source video[s]
        #  - shorter line segments representing the features extracted where:
        #    + width represents duration
        #    + colour represents feature type
        #    + position represents time
        #  - save as image
        plotted_source_videos = []
        bar_labels = []

        fig, ax = plt.subplots()
        for feature in self.features:
            # plot source video line if not done already
            if feature.source not in plotted_source_videos:
                # use video duration as width
                # ax.plot([0, feature.source.duration()], [0, 0], color='black', linewidth=10)
                ax.broken_barh([(0, feature.source.duration())], (0, 5), facecolors='grey')
                plotted_source_videos.append(feature.source)
                bar_labels.append(os.path.basename(feature.source.path))
                # annotate the source video
                ax.text(0.25, 0.25, os.path.basename(feature.source.path), ha='left', va='bottom',
                        fontsize=16)

            # plot feature line
            # ax.plot([feature.interval.start, feature.interval.end], [1, 1], color='red', linewidth=5)
            ax.broken_barh([(feature.interval.start, feature.interval.duration)], (10, 5), facecolors='red')
            if feature.feature_extractor not in bar_labels:
                bar_labels.append(feature.feature_extractor)
                # label bar with feature extractor
                ax.text(0, 8, feature.feature_extractor, ha='left', va='bottom',
                        fontsize=16)

        # label the plot's axes
        ax.set_xlabel('Time')
        # ax.set_yticks([], labels=bar_labels)
        ax.set_yticks([])
        # ax.tick_params(axis='y', labelrotation=90, ha='right')
        # save the plot
        plt.savefig("/tmp/visualisation.png")
        plt.close()

class PipelineJSONEncoder(json.JSONEncoder):
    def default(self, obj):
        if hasattr(obj, 'to_json'):
            return obj.to_json()
        else:
            return json.JSONEncoder.default(self, obj)

class JSONProducer(Producer):
    """Produce JSON output"""
    def __init__(self, features):
        if not features:
            raise ValueError("No features provided")
        self.features = features

    def produce(self):
        # FIXME: config option for output path
        with open("/tmp/features.json", "w") as jsonfile:
            jsonfile.write(json.dumps(self.features, cls=PipelineJSONEncoder, indent=4))