ale/main.py
2022-10-21 20:47:39 -04:00

211 lines
8.3 KiB
Python

import argparse
import structlog
from functools import partial
from pathlib import Path
import sys
import hashlib
import random
import multiprocessing
import concurrent.futures
import time
import numpy as np
from src.mediautils.audio import extract_audio_from_video
from src.mediautils.video import render_moments
from src.editors.amplitude.editor import AmplitudeEditor
from src.editors.sentiment.editor import SentimentEditor
from src.math.cost import quadratic_loss
from src.math.distribution import create_distribution
log = structlog.get_logger()
EDITORS = {
'amplitude': AmplitudeEditor,
'sentiment': SentimentEditor
}
ERROR_FUNCS = {
'quadratic': quadratic_loss
}
def main(args):
# Check video existance
input_file = args.file
in_vid_path = Path(input_file)
if not in_vid_path.is_file():
log.error("the specified input path does not exist", path=input_file)
sys.exit(-1)
log.info("preparing video", input_video=input_file)
# Hash the video, we use this to see if we have processed this video before
# and as a simple way to generate temp file names
sha1 = hashlib.sha1()
BUF_SIZE = 1655360
with open(in_vid_path, 'rb') as f:
while True:
data = f.read(BUF_SIZE)
if not data:
break
sha1.update(data)
temp_file_name = sha1.hexdigest()
log.info("hash computed", hash=temp_file_name)
temp_file_name = f"ale-{temp_file_name}"
# Prepare the input video
audio_path, audio_cached = extract_audio_from_video(str(in_vid_path.resolve()), temp_file_name)
if audio_cached:
log.info("using cached audio file", cache_path=audio_path)
else:
log.info("extracted audio", cache_path=audio_path)
# Initalize Editor
log.info("initializing editor", editor=args.editor)
editor = EDITORS[args.editor](str(in_vid_path.resolve()), audio_path, vars(args))
log.info("initialized editor", editor=args.editor)
costfunc = ERROR_FUNCS[args.cost]
desired = args.duration
# Generate center of large window and small window size
large_window_center = random.uniform(30, 50)
small_window_center = random.uniform(5, 15)
# The spread multiplier, or epsilon, slowly decays as we approach the center of the gradient
spread_multiplier = random.uniform(0.15, 0.18)
# The decay rate, or how quickly our spread multiplier decreases as we approach the center of the gradient
spread_decay = random.uniform(0.0001, 0.001)
parallelism = args.parallelism
# The main loop of the program starts here
# we first create distributions
# use workers to simultanously create many possible edits
# find the best edit of the lot -> this is determined by lowest "cost"
# if the best fits within our desitred time range, output, otherwise
# reset the distributions using the best as the new center, then repeat
# Create distribution of large and small
complete = False
iterations = 0
while not complete:
log.info("creating distributions", large_start=large_window_center, small_start=small_window_center, spread=spread_multiplier, decay=spread_decay)
large_distribution = create_distribution(large_window_center, spread_multiplier, parallelism)
np.random.shuffle(large_distribution)
small_distribution = create_distribution(small_window_center, spread_multiplier, parallelism)
np.random.shuffle(small_distribution)
# Fire off workers to generate edits
moment_results = []
with concurrent.futures.ThreadPoolExecutor() as executor:
futures = []
pairs = list(zip(large_distribution, small_distribution))
for pair in pairs:
futures.append(
executor.submit(
editor.edit,
pair[0],
pair[1],
vars(args)
)
)
for future in concurrent.futures.as_completed(futures):
try:
moment_results.append(future.result())
except Exception:
log.exception("error during editing")
sys.exit(-2)
moment_results
costs = []
durations = []
for result in moment_results:
total_duration = 0
for moment in result[0]:
total_duration = total_duration + moment.get_duration()
costs.append(costfunc(desired, total_duration))
durations.append(total_duration)
index_min = min(range(len(costs)), key=costs.__getitem__)
large_window_center = moment_results[index_min][1]
small_window_center = moment_results[index_min][2]
log.info("batch complete", best_large=large_window_center, best_small=small_window_center, duration=durations[index_min])
if durations[index_min] > desired * 0.95 and desired * 1.05 > durations[index_min]:
log.info("found edit within target duration", target=desired, duration=durations[index_min])
out_path = Path(args.destination)
log.info("rendering...")
start = time.time()
render_moments(moment_results[index_min][0], str(in_vid_path.resolve()), str(out_path.resolve()))
log.info("render complete", duration=time.time() - start, output=str(out_path.resolve()))
sys.exit(0)
iterations = iterations + parallelism
if iterations > 50000:
log.error("could not find a viable edit in the target duration, try other params", target=desired)
sys.exit(-4)
spread_multiplier = spread_multiplier - spread_decay
if spread_multiplier < 0:
log.warn("spread reached 0, resetting")
large_window_center = random.uniform(30, 50)
small_window_center = random.uniform(5, 15)
spread_multiplier = random.uniform(0.15, 0.18)
spread_decay = random.uniform(0.0001, 0.001)
if __name__ == "__main__":
parser = argparse.ArgumentParser(
prog="ALE", description="ALE: Automatic Linear Editor.",
formatter_class=partial(argparse.HelpFormatter, width=100)
)
parser.add_argument('file', help='Path to the video file to edit')
parser.add_argument('duration', help='Target length of the edit, in seconds', type=int)
parser.add_argument('destination', help='Edited video save location')
subparsers = parser.add_subparsers(dest='editor', help='The editing algorithm to use')
parser_audio_amp = subparsers.add_parser('amplitude', help='The amplitude editor uses audio amplitude moving averages to find swings from relatively quiet moments to loud moments. This is useful in videos where long moments of quiet are interspersed with loud action filled moments.')
parser_audio_amp.add_argument(
"--factor",
default=16000,
help="Subsampling factor",
dest="factor",
type=int,
)
parser_sentiment = subparsers.add_parser('sentiment', help='The sentiment editor transcribes the speech in a video and runs sentiment analysis on the resulting text. Using moving averages, large swings in sentiment can be correlated to controversial or exciting moments. A GPU with CUDA is recommended for fast results.')
parser_sentiment.add_argument(
"--model",
default="base",
help="The size of the sentiment analysis model being used. Larger models increase computation time.",
dest="model_size",
choices=["base", "tiny", "small", "medium", "large"],
)
parser.add_argument("-p", "--parallelism", dest="parallelism", type=int, default=multiprocessing.cpu_count() - 2, help="The number of cores to use, defaults to N - 2 cores.")
parser.add_argument("--cost-function", dest="cost", choices=ERROR_FUNCS.keys(), default='quadratic')
parser.add_argument(
"--minduration",
default=8,
help="Minimum clip duration",
dest="mindur",
type=int,
)
parser.add_argument(
"--maxduration",
default=15,
help="Maximum clip duration",
dest="maxdur",
type=int,
)
args = parser.parse_args()
try:
main(args)
except Exception:
log.exception("uncaught error!")
sys.exit(-2)
sys.exit(0)