File size: 11,090 Bytes
2e36228 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 |
import os, subprocess, glob, pandas, tqdm, cv2, numpy
from scipy.io import wavfile
import random
import torch
import numpy as np
def init_args(args):
# The details for the following folders/files can be found in the annotation of the function 'preprocess_AVA' below
args.modelSavePath = os.path.join(args.WORKSPACE, 'model')
args.scoreSavePath = os.path.join(args.WORKSPACE, 'score.txt')
args.trialPathAVA = os.path.join(args.DATA.dataPathAVA, 'csv')
args.audioOrigPathAVA = os.path.join(args.DATA.dataPathAVA, 'orig_audios')
args.visualOrigPathAVA = os.path.join(args.DATA.dataPathAVA, 'orig_videos')
args.audioPathAVA = os.path.join(args.DATA.dataPathAVA, 'clips_audios')
args.visualPathAVA = os.path.join(args.DATA.dataPathAVA, 'clips_videos')
args.trainTrialAVA = os.path.join(args.trialPathAVA, 'train_loader.csv')
if args.evalDataType == 'val':
args.evalTrialAVA = os.path.join(args.trialPathAVA, 'val_loader.csv')
args.evalOrig = os.path.join(args.trialPathAVA, 'val_orig.csv')
args.evalCsvSave = os.path.join(args.WORKSPACE, 'val_res.csv')
else:
args.evalTrialAVA = os.path.join(args.trialPathAVA, 'test_loader.csv')
args.evalOrig = os.path.join(args.trialPathAVA, 'test_orig.csv')
args.evalCsvSave = os.path.join(args.WORKSPACE, 'test_res.csv')
os.makedirs(args.modelSavePath, exist_ok=True)
os.makedirs(args.DATA.dataPathAVA, exist_ok=True)
return args
def make_deterministic(seed, strict=False):
#https://github.com/pytorch/pytorch/issues/7068#issuecomment-487907668
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed) # if you are using multi-GPU.
# torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
# torch.set_deterministic(True)
if strict:
#https://github.com/pytorch/pytorch/issues/7068#issuecomment-515728600
torch.backends.cudnn.enabled = False
print(
"strict reproducability required! cudnn disabled. make sure to set num_workers=0 too!")
def download_pretrain_model_AVA():
if os.path.isfile('pretrain_AVA.model') == False:
Link = "1NVIkksrD3zbxbDuDbPc_846bLfPSZcZm"
cmd = "gdown --id %s -O %s" % (Link, 'pretrain_AVA.model')
subprocess.call(cmd, shell=True, stdout=None)
def preprocess_AVA(args):
# This preprocesstion is modified based on this [repository](https://github.com/fuankarion/active-speakers-context).
# The required space is 302 G.
# If you do not have enough space, you can delate `orig_videos`(167G) when you get `clips_videos(85G)`.
# also you can delate `orig_audios`(44G) when you get `clips_audios`(6.4G).
# So the final space is less than 100G.
# The AVA dataset will be saved in 'AVApath' folder like the following format:
# ```
# ├── clips_audios (The audio clips cut from the original movies)
# │ ├── test
# │ ├── train
# │ └── val
# ├── clips_videos (The face clips cut from the original movies, be save in the image format, frame-by-frame)
# │ ├── test
# │ ├── train
# │ └── val
# ├── csv
# │ ├── test_file_list.txt (name of the test videos)
# │ ├── test_loader.csv (The csv file we generated to load data for testing)
# │ ├── test_orig.csv (The combination of the given test csv files)
# │ ├── train_loader.csv (The csv file we generated to load data for training)
# │ ├── train_orig.csv (The combination of the given training csv files)
# │ ├── trainval_file_list.txt (name of the train/val videos)
# │ ├── val_loader.csv (The csv file we generated to load data for validation)
# │ └── val_orig.csv (The combination of the given validation csv files)
# ├── orig_audios (The original audios from the movies)
# │ ├── test
# │ └── trainval
# └── orig_videos (The original movies)
# ├── test
# └── trainval
# ```
download_csv(args) # Take 1 minute
download_videos(args) # Take 6 hours
extract_audio(args) # Take 1 hour
extract_audio_clips(args) # Take 3 minutes
extract_video_clips(args) # Take about 2 days
def download_csv(args):
# Take 1 minute to download the required csv files
Link = "1C1cGxPHaJAl1NQ2i7IhRgWmdvsPhBCUy"
cmd = "gdown --id %s -O %s" % (Link, args.dataPathAVA + '/csv.tar.gz')
subprocess.call(cmd, shell=True, stdout=None)
cmd = "tar -xzvf %s -C %s" % (args.dataPathAVA + '/csv.tar.gz', args.dataPathAVA)
subprocess.call(cmd, shell=True, stdout=None)
os.remove(args.dataPathAVA + '/csv.tar.gz')
def download_videos(args):
# Take 6 hours to download the original movies, follow this repository: https://github.com/cvdfoundation/ava-dataset
for dataType in ['trainval', 'test']:
fileList = open('%s/%s_file_list.txt' % (args.trialPathAVA, dataType)).read().splitlines()
outFolder = '%s/%s' % (args.visualOrigPathAVA, dataType)
for fileName in fileList:
cmd = "wget -P %s https://s3.amazonaws.com/ava-dataset/%s/%s" % (outFolder, dataType,
fileName)
subprocess.call(cmd, shell=True, stdout=None)
def extract_audio(args):
# Take 1 hour to extract the audio from movies
for dataType in ['trainval', 'test']:
inpFolder = '%s/%s' % (args.visualOrigPathAVA, dataType)
outFolder = '%s/%s' % (args.audioOrigPathAVA, dataType)
os.makedirs(outFolder, exist_ok=True)
videos = glob.glob("%s/*" % (inpFolder))
for videoPath in tqdm.tqdm(videos):
audioPath = '%s/%s' % (outFolder, videoPath.split('/')[-1].split('.')[0] + '.wav')
cmd = (
"ffmpeg -y -i %s -async 1 -ac 1 -vn -acodec pcm_s16le -ar 16000 -threads 8 %s -loglevel panic"
% (videoPath, audioPath))
subprocess.call(cmd, shell=True, stdout=None)
def extract_audio_clips(args):
# Take 3 minutes to extract the audio clips
dic = {'train': 'trainval', 'val': 'trainval', 'test': 'test'}
for dataType in ['train', 'val', 'test']:
df = pandas.read_csv(os.path.join(args.trialPathAVA, '%s_orig.csv' % (dataType)),
engine='python')
dfNeg = pandas.concat([df[df['label_id'] == 0], df[df['label_id'] == 2]])
dfPos = df[df['label_id'] == 1]
insNeg = dfNeg['instance_id'].unique().tolist()
insPos = dfPos['instance_id'].unique().tolist()
df = pandas.concat([dfPos, dfNeg]).reset_index(drop=True)
df = df.sort_values(['entity_id', 'frame_timestamp']).reset_index(drop=True)
entityList = df['entity_id'].unique().tolist()
df = df.groupby('entity_id')
audioFeatures = {}
outDir = os.path.join(args.audioPathAVA, dataType)
audioDir = os.path.join(args.audioOrigPathAVA, dic[dataType])
for l in df['video_id'].unique().tolist():
d = os.path.join(outDir, l[0])
if not os.path.isdir(d):
os.makedirs(d)
for entity in tqdm.tqdm(entityList, total=len(entityList)):
insData = df.get_group(entity)
videoKey = insData.iloc[0]['video_id']
start = insData.iloc[0]['frame_timestamp']
end = insData.iloc[-1]['frame_timestamp']
entityID = insData.iloc[0]['entity_id']
insPath = os.path.join(outDir, videoKey, entityID + '.wav')
if videoKey not in audioFeatures.keys():
audioFile = os.path.join(audioDir, videoKey + '.wav')
sr, audio = wavfile.read(audioFile)
audioFeatures[videoKey] = audio
audioStart = int(float(start) * sr)
audioEnd = int(float(end) * sr)
audioData = audioFeatures[videoKey][audioStart:audioEnd]
wavfile.write(insPath, sr, audioData)
def extract_video_clips(args):
# Take about 2 days to crop the face clips.
# You can optimize this code to save time, while this process is one-time.
# If you do not need the data for the test set, you can only deal with the train and val part. That will take 1 day.
# This procession may have many warning info, you can just ignore it.
dic = {'train': 'trainval', 'val': 'trainval', 'test': 'test'}
for dataType in ['train', 'val', 'test']:
df = pandas.read_csv(os.path.join(args.trialPathAVA, '%s_orig.csv' % (dataType)))
dfNeg = pandas.concat([df[df['label_id'] == 0], df[df['label_id'] == 2]])
dfPos = df[df['label_id'] == 1]
insNeg = dfNeg['instance_id'].unique().tolist()
insPos = dfPos['instance_id'].unique().tolist()
df = pandas.concat([dfPos, dfNeg]).reset_index(drop=True)
df = df.sort_values(['entity_id', 'frame_timestamp']).reset_index(drop=True)
entityList = df['entity_id'].unique().tolist()
df = df.groupby('entity_id')
outDir = os.path.join(args.visualPathAVA, dataType)
audioDir = os.path.join(args.visualOrigPathAVA, dic[dataType])
for l in df['video_id'].unique().tolist():
d = os.path.join(outDir, l[0])
if not os.path.isdir(d):
os.makedirs(d)
for entity in tqdm.tqdm(entityList, total=len(entityList)):
insData = df.get_group(entity)
videoKey = insData.iloc[0]['video_id']
entityID = insData.iloc[0]['entity_id']
videoDir = os.path.join(args.visualOrigPathAVA, dic[dataType])
videoFile = glob.glob(os.path.join(videoDir, '{}.*'.format(videoKey)))[0]
V = cv2.VideoCapture(videoFile)
insDir = os.path.join(os.path.join(outDir, videoKey, entityID))
if not os.path.isdir(insDir):
os.makedirs(insDir)
j = 0
for _, row in insData.iterrows():
imageFilename = os.path.join(insDir, str("%.2f" % row['frame_timestamp']) + '.jpg')
V.set(cv2.CAP_PROP_POS_MSEC, row['frame_timestamp'] * 1e3)
_, frame = V.read()
h = numpy.size(frame, 0)
w = numpy.size(frame, 1)
x1 = int(row['entity_box_x1'] * w)
y1 = int(row['entity_box_y1'] * h)
x2 = int(row['entity_box_x2'] * w)
y2 = int(row['entity_box_y2'] * h)
face = frame[y1:y2, x1:x2, :]
j = j + 1
cv2.imwrite(imageFilename, face)
|