Recherche avancée

Sur d’autres sites (8595)

• Revision c82de3bede : Extending ext_tx expt to include dst variants Extends the ext-tx experiment to

  15 novembre 2014, par Deb Mukherjee

  Changed Paths :
   Modify /vp9/common/vp9_blockd.h

  
   Modify /vp9/common/vp9_entropymode.c

  
   Modify /vp9/common/vp9_entropymode.h

  
   Modify /vp9/common/vp9_enums.h

  
   Modify /vp9/common/vp9_idct.c

  
   Modify /vp9/decoder/vp9_decodeframe.c

  
   Modify /vp9/decoder/vp9_decodemv.c

  
   Modify /vp9/encoder/vp9_bitstream.c

  
   Modify /vp9/encoder/vp9_encodeframe.c

  
   Modify /vp9/encoder/vp9_encodemb.c

  
   Modify /vp9/encoder/vp9_encoder.h

  
   Modify /vp9/encoder/vp9_rd.c

  
   Modify /vp9/encoder/vp9_rdopt.c

  
  
  Extending ext_tx expt to include dst variants

  Extends the ext-tx experiment to include regular and flipped
  DST variants. A total of 9 transforms are thus possible for
  each inter block with transform size <= 16x16.

  In this patch currently only the four ADST_ADST variants
  (flipped or non-flipped in both dimensions) are enabled
  for inter blocks.

  The gain with the ext-tx experiment grows to +1.12 on derflr.
  Further experiments are underway.

  Change-Id : Ia2ed19a334face6135b064748f727fdc9db278ec

• lavu : fix memory leaks by using a mutex instead of atomics

  14 novembre 2014, par wm4

  lavu : fix memory leaks by using a mutex instead of atomics
  The buffer pool has to atomically add and remove entries from the linked
  
  list of available buffers. This was done by removing the entire list
  
  with a CAS operation, working on it, and then setting it back again
  
  (using a retry-loop in case another thread was doing the same thing).
  This could effectively cause memory leaks : while a thread was working on
  
  the buffer list, other threads would allocate new buffers, increasing
  
  the pool’s total size. There was no real leak, but since these extra
  
  buffers were not needed, but not free’d either (except when the buffer
  
  pool was destroyed), this had the same effects as a real leak. For some
  
  reason, growth was exponential, and could easily kill the process due
  
  to OOM in real-world uses.
  Fix this by using a mutex to protect the list operations. The fancy
  
  way atomics remove the whole list to work on it is not needed anymore,
  
  which also avoids the situation which was causing the leak.
  Signed-off-by : Anton Khirnov <anton@khirnov.net>
  

  • [DBH] libavutil/buffer.c
  • [DBH] libavutil/buffer_internal.h

• Parallelize Youtube video frame download using yt-dlp and cv2

  4 mars 2023, par zulle99

  My task is to download multiple sequences of successive low resolution frames of Youtube videos.

  &#xA;

  I summarize the main parts of the process :

  &#xA;

  &#xA;
  • Each bag of shots have a dimension of half a second (depending on the current fps)

  &#xA;

  • In order to grab useful frames I've decided to remove the initial and final 10% of each video since it is common to have an intro and outro. Moreover

  &#xA;

  • I've made an array of pair of initial and final frame to distribute the load on multiple processes using ProcessPoolExecutor(max_workers=multiprocessing.cpu_count())

  &#xA;

  • In case of failure/exception I completly remove the relative directory

  &#xA;

  &#xA;

  The point is that it do not scale up, since while running I noticesd that all CPUs had always a load lower that the 20% more or less. In addition since with these shots I have to run multiple CNNs, to prevent overfitting it is suggested to have a big dataset and not a bounch of shots.

  &#xA;

  Here it is the code :

  &#xA;

  import yt_dlp
import os
from tqdm import tqdm
import cv2
import shutil
import time
import random
from concurrent.futures import ProcessPoolExecutor
import multiprocessing
import pandas as pd
import numpy as np
from pathlib import Path
import zipfile


# PARAMETERS
percentage_train_test = 50
percentage_bag_shots = 20
percentage_to_ignore = 10

zip_f_name = f'VideoClassificationDataset_{percentage_train_test}_{percentage_bag_shots}_{percentage_to_ignore}'
dataset_path = Path('/content/VideoClassificationDataset')

# DOWNOAD ZIP FILES
!wget --no-verbose https://github.com/gtoderici/sports-1m-dataset/archive/refs/heads/master.zip

# EXTRACT AND DELETE THEM
!unzip -qq -o '/content/master.zip' 
!rm '/content/master.zip'

DATA = {'train_partition.txt': {},
        'test_partition.txt': {}}

LABELS = []

train_dict = {}
test_dict = {}

path = '/content/sports-1m-dataset-master/original'

for f in os.listdir(path):
  with open(path + '/' + f) as f_txt:
    lines = f_txt.readlines()
    for line in lines:
      splitted_line = line.split(' ')
      label_indices = splitted_line[1].rstrip('\n').split(',') 
      DATA[f][splitted_line[0]] = list(map(int, label_indices))

with open('/content/sports-1m-dataset-master/labels.txt') as f_labels:
  LABELS = f_labels.read().splitlines()


TRAIN = DATA['train_partition.txt']
TEST = DATA['test_partition.txt']
print('Original Train Test length: ', len(TRAIN), len(TEST))

# sample a subset percentage_train_test
TRAIN = dict(random.sample(TRAIN.items(), (len(TRAIN)*percentage_train_test)//100))
TEST = dict(random.sample(TEST.items(), (len(TEST)*percentage_train_test)//100))

print(f'Sampled {percentage_train_test} Percentage  Train Test length: ', len(TRAIN), len(TEST))


if not os.path.exists(dataset_path): os.makedirs(dataset_path)
if not os.path.exists(f'{dataset_path}/train'): os.makedirs(f'{dataset_path}/train')
if not os.path.exists(f'{dataset_path}/test'): os.makedirs(f'{dataset_path}/test')


  &#xA;

  Function to extract a sequence of continuous frames :

  &#xA;

  def extract_frames(directory, url, idx_bag, start_frame, end_frame):
  capture = cv2.VideoCapture(url)
  count = start_frame

  capture.set(cv2.CAP_PROP_POS_FRAMES, count)
  os.makedirs(f'{directory}/bag_of_shots{str(idx_bag)}')

  while count < end_frame:

    ret, frame = capture.read()

    if not ret: 
      shutil.rmtree(f'{directory}/bag_of_shots{str(idx_bag)}')
      return False

    filename = f'{directory}/bag_of_shots{str(idx_bag)}/shot{str(count - start_frame)}.png'

    cv2.imwrite(filename, frame)
    count += 1

  capture.release()
  return True


  &#xA;

  Function to spread the load along multiple processors :

  &#xA;

  def video_to_frames(video_url, labels_list, directory, dic, percentage_of_bags):
  url_id = video_url.split('=')[1]
  path_until_url_id = f'{dataset_path}/{directory}/{url_id}'
  try:   

    ydl_opts = {
        'ignoreerrors': True,
        'quiet': True,
        'nowarnings': True,
        'simulate': True,
        'ignorenoformatserror': True,
        'verbose':False,
        'cookies': '/content/all_cookies.txt',
        #https://stackoverflow.com/questions/63329412/how-can-i-solve-this-youtube-dl-429
    }
    ydl = yt_dlp.YoutubeDL(ydl_opts)
    info_dict = ydl.extract_info(video_url, download=False)

    if(info_dict is not None and  info_dict['fps'] >= 20):
      # I must have a least 20 frames per seconds since I take half of second bag of shots for every video

      formats = info_dict.get('formats', None)

      # excluding the initial and final 10% of each video to avoid noise
      video_length = info_dict['duration'] * info_dict['fps']

      shots = info_dict['fps'] // 2

      to_ignore = (video_length * percentage_to_ignore) // 100
      new_len = video_length - (to_ignore * 2)
      tot_stored_bags = ((new_len // shots) * percentage_of_bags) // 100   # ((total_possbile_bags // shots) * percentage_of_bags) // 100
      if tot_stored_bags == 0: tot_stored_bags = 1 # minimum 1 bag of shots

      skip_rate_between_bags = (new_len - (tot_stored_bags * shots)) // (tot_stored_bags-1) if tot_stored_bags > 1 else 0

      chunks = [[to_ignore+(bag*(skip_rate_between_bags+shots)), to_ignore+(bag*(skip_rate_between_bags+shots))+shots] for bag in range(tot_stored_bags)]
      # sequence of [[start_frame, end_frame], [start_frame, end_frame], [start_frame, end_frame], ...]


      # ----------- For the moment I download only shots form video that has 144p resolution -----------

      res = {
          '160': '144p',
          '133': '240p',
          '134': '360p',
          '135': '360p',
          '136': '720p'
      }

      format_id = {}
      for f in formats: format_id[f['format_id']] = f
      #for res in resolution_id:
      if list(res.keys())[0] in list(format_id.keys()):
          video = format_id[list(res.keys())[0]]
          url = video.get('url', None)
          if(video.get('url', None) != video.get('manifest_url', None)):

            if not os.path.exists(path_until_url_id): os.makedirs(path_until_url_id)

            with ProcessPoolExecutor(max_workers=multiprocessing.cpu_count()) as executor:
              for idx_bag, f in enumerate(chunks): 
                res = executor.submit(
                  extract_frames, directory = path_until_url_id, url = url, idx_bag = idx_bag, start_frame = f[0], end_frame = f[1])
                
                if res.result() is True: 
                  l = np.zeros(len(LABELS), dtype=int) 
                  for label in labels_list: l[label] = 1
                  l = np.append(l, [shots]) # appending the number of shots taken in the list before adding it on the dictionary

                  dic[f'{directory}/{url_id}/bag_of_shots{str(idx_bag)}'] = l.tolist()


  except Exception as e:
    shutil.rmtree(path_until_url_id)
    pass


  &#xA;

  Download of TRAIN bag of shots :

  &#xA;

  start_time = time.time()
pbar = tqdm(enumerate(TRAIN.items()), total = len(TRAIN.items()), leave=False)

for _, (url, labels_list) in pbar: video_to_frames(
  video_url = url, labels_list = labels_list, directory = 'train', dic = train_dict, percentage_of_bags = percentage_bag_shots)

print("--- %s seconds ---" % (time.time() - start_time))


  &#xA;

  Download of TEST bag of shots :

  &#xA;

  start_time = time.time()
pbar = tqdm(enumerate(TEST.items()), total = len(TEST.items()), leave=False)

for _, (url, labels_list) in pbar: video_to_frames(
  video_url = url, labels_list = labels_list, directory = 'test', dic = test_dict, percentage_of_bags = percentage_bag_shots)

print("--- %s seconds ---" % (time.time() - start_time))


  &#xA;

  Save the .csv files

  &#xA;

  train_df = pd.DataFrame.from_dict(train_dict, orient='index', dtype=int).reset_index(level=0)
train_df = train_df.rename(columns={train_df.columns[-1]: 'shots'})
train_df.to_csv('/content/VideoClassificationDataset/train.csv', index=True)

test_df = pd.DataFrame.from_dict(test_dict, orient='index', dtype=int).reset_index(level=0)
test_df = test_df.rename(columns={test_df.columns[-1]: 'shots'})
test_df.to_csv('/content/VideoClassificationDataset/test.csv', index=True)


  &#xA;