def process_visual_genome():
setup_django()
import os, shutil, gzip, json
from django.core.files.uploadedfile import SimpleUploadedFile
from dvaapp.view_shared import handle_uploaded_file
from dvaapp import models
from dvaapp.models import TEvent
from dvaapp.tasks import perform_dataset_extraction, export_video
from collections import defaultdict
os.system('aws s3api get-object --request-payer "requester" --bucket visualdatanetwork --key visual_genome_objects.txt.gz /root/DVA/visual_genome_objects.txt.gz')
data = defaultdict(list)
with gzip.open('/root/DVA/visual_genome_objects.txt.gz') as metadata:
for line in metadata:
entries = line.strip().split('\t')
data[entries[1]].append({
'x': int(entries[2]),
'y': int(entries[3]),
'w': int(entries[4]),
'h': int(entries[5]),
'object_id': entries[0],
'object_name': entries[6],
'text': ' '.join(entries[6:]), })
name = "visual_genome"
fname = "visual_genome.zip"
f = SimpleUploadedFile(fname, "", content_type="application/zip")
v = handle_uploaded_file(f, name)
outpath = "/root/DVA/dva/media/{}/video/{}.zip".format(v.pk, v.pk)
os.system('rm {}'.format(outpath))
os.system(
'aws s3api get-object --request-payer "requester" --bucket visualdatanetwork --key visual_genome.zip {}'.format(
outpath))
perform_dataset_extraction(TEvent.objects.create(video=v).pk)
video = v
models.Region.objects.all().filter(video=video).delete()
buffer = []
batch_count = 0
for frame in models.Frame.objects.all().filter(video=video):
frame_id = str(int(frame.name.split('/')[-1].split('.')[0]))
for o in data[frame_id]:
annotation = models.Region()
annotation.region_type = models.Region.ANNOTATION
annotation.video = v
annotation.frame = frame
annotation.x = o['x']
annotation.y = o['y']
annotation.h = o['h']
annotation.w = o['w']
annotation.object_name = o['object_name']
annotation.metadata = o
annotation.text = o['text']
buffer.append(annotation)
if len(buffer) == 1000:
try:
models.Region.objects.bulk_create(buffer)
batch_count += 1
print "saved {}".format(batch_count)
except:
print "encountered an error doing one by one"
for k in buffer:
try:
k.save()
except:
print "skipping"
print k.object_name
buffer = []
try:
models.Region.objects.bulk_create(buffer)
print "saved {}".format(batch_count)
except:
print "encountered an error doing one by one"
for k in buffer:
try:
k.save()
except:
print "skipping"
print k.object_name
print "exporting"
export_video(TEvent.objects.create(video=v).pk)
def ci():
"""
Used in conjunction with travis for Continuous Integration testing
:return:
"""
import django
sys.path.append(os.path.dirname(__file__))
os.environ.setdefault("DJANGO_SETTINGS_MODULE", "dva.settings")
django.setup()
import base64
from django.core.files.uploadedfile import SimpleUploadedFile
from dvaapp.views import handle_uploaded_file, handle_youtube_video, pull_vdn_list\
,import_vdn_dataset_url
from dvaapp.models import Video, Clusters,IndexEntries,TEvent,VDNServer, DVAPQL
from django.conf import settings
from dvaapp.operations.processing import DVAPQLProcess
from dvaapp.tasks import extract_frames, perform_indexing, export_video, import_video_by_id,\
perform_clustering, perform_analysis, perform_detection,\
segment_video, crop_regions_by_id
for fname in glob.glob('tests/ci/*.mp4'):
name = fname.split('/')[-1].split('.')[0]
f = SimpleUploadedFile(fname, file(fname).read(), content_type="video/mp4")
handle_uploaded_file(f, name, False)
if sys.platform != 'darwin':
for fname in glob.glob('tests/*.mp4'):
name = fname.split('/')[-1].split('.')[0]
f = SimpleUploadedFile(fname, file(fname).read(), content_type="video/mp4")
handle_uploaded_file(f, name, False)
for fname in glob.glob('tests/*.zip'):
name = fname.split('/')[-1].split('.')[0]
f = SimpleUploadedFile(fname, file(fname).read(), content_type="application/zip")
handle_uploaded_file(f, name)
# handle_youtube_video('world is not enough', 'https://www.youtube.com/watch?v=P-oNz3Nf50Q') # Temporarily disabled due error in travis
for i,v in enumerate(Video.objects.all()):
if v.dataset:
arguments = {'sync':True}
extract_frames(TEvent.objects.create(video=v,arguments=arguments).pk)
else:
arguments = {'sync':True}
segment_video(TEvent.objects.create(video=v,arguments=arguments).pk)
arguments = {'index': 'inception'}
perform_indexing(TEvent.objects.create(video=v,arguments=arguments).pk)
if i ==0: # save travis time by just running detection on first video
# face_mtcnn
arguments = {'detector': 'face'}
dt = TEvent.objects.create(video=v,arguments=arguments)
perform_detection(dt.pk)
arguments = {'filters':{'event_id':dt.pk},}
crop_regions_by_id(TEvent.objects.create(video=v,arguments=arguments).pk)
# coco_mobilenet
arguments = {'detector': 'coco'}
dt = TEvent.objects.create(video=v, arguments=arguments)
perform_detection(dt.pk)
arguments = {'filters':{'event_id':dt.pk},}
crop_regions_by_id(TEvent.objects.create(video=v,arguments=arguments).pk)
# inception on crops from detector
arguments = {'index':'inception','target': 'regions','filters': {'event_id': dt.pk, 'w__gte': 50, 'h__gte': 50}}
perform_indexing(TEvent.objects.create(video=v,arguments=arguments).pk)
# assign_open_images_text_tags_by_id(TEvent.objects.create(video=v).pk)
fname = export_video(TEvent.objects.create(video=v).pk)
f = SimpleUploadedFile(fname, file("{}/exports/{}".format(settings.MEDIA_ROOT,fname)).read(), content_type="application/zip")
vimported = handle_uploaded_file(f, fname)
import_video_by_id(TEvent.objects.create(video=vimported).pk)
dc = Clusters()
dc.indexer_algorithm = 'inception'
dc.included_index_entries_pk = [k.pk for k in IndexEntries.objects.all().filter(algorithm=dc.indexer_algorithm)]
dc.components = 32
dc.save()
clustering_task = TEvent()
clustering_task.arguments = {'clusters_id':dc.pk}
clustering_task.operation = 'perform_clustering'
clustering_task.save()
perform_clustering(clustering_task.pk)
query_dict = {
'process_type': DVAPQL.QUERY,
'image_data_b64':base64.encodestring(file('tests/query.png').read()),
'indexer_queries':[
{
'algorithm':'inception',
'count':10,
'approximate':False
}
]
}
qp = DVAPQLProcess()
qp.create_from_json(query_dict)
# execute_index_subquery(qp.indexer_queries[0].pk)
query_dict = {
'process_type': DVAPQL.QUERY,
'image_data_b64':base64.encodestring(file('tests/query.png').read()),
'indexer_queries':[
{
'algorithm':'inception',
'count':10,
'approximate':True
}
]
}
qp = DVAPQLProcess()
qp.create_from_json(query_dict)
# execute_index_subquery(qp.indexer_queries[0].pk)
server, datasets, detectors = pull_vdn_list(1)
for k in datasets:
if k['name'] == 'MSCOCO_Sample_500':
print 'FOUND MSCOCO SAMPLE'
import_vdn_dataset_url(VDNServer.objects.get(pk=1),k['url'],None)
test_backup()
annotation.object_name = 'caption'
buf.append(annotation)
if len(buf) > 1000:
try:
models.Region.objects.bulk_create(buf)
batch_count += 1
print "saved {}".format(batch_count)
except:
print "encountered an error doing one by one"
for k in buf:
try:
k.save()
except:
print "skipping"
print k.object_name
buf = []
try:
models.Region.objects.bulk_create(buf)
batch_count += 1
print "saved {}".format(batch_count)
except:
print "encountered an error doing one by one"
for k in buf:
try:
k.save()
except:
print "skipping"
print k.object_name
buf = []
export_video(TEvent.objects.create(video=v).pk)
def create_yolo_test_data():
import shutil
import numpy as np
import os
from PIL import Image
setup_django()
from dvaapp.shared import handle_uploaded_file
from django.core.files.uploadedfile import SimpleUploadedFile
from dvaapp.models import Region,TEvent,Frame,Label,RegionLabel
from dvaapp.tasks import extract_frames,export_video
try:
shutil.rmtree('/Users/aub3/tests/yolo_test')
except:
pass
try:
os.mkdir('/Users/aub3/tests/yolo_test')
except:
pass
data = np.load('shared/underwater_data.npz')
json_test = {}
json_test['anchors'] = [(0.57273, 0.677385), (1.87446, 2.06253), (3.33843, 5.47434), (7.88282, 3.52778), (9.77052, 9.16828)]
id_2_boxes = {}
class_names = {
0:"red_buoy",
1:"green_buoy",
2:"yellow_buoy",
3:"path_marker",
4:"start_gate",
5:"channel"
}
labels = {k: Label.objects.create(name=v, set="test") for k, v in class_names}
for i,image in enumerate(data['images'][:500]):
path = "/Users/aub3/tests/yolo_test/{}.jpg".format(i)
Image.fromarray(image).save(path)
id_2_boxes[path.split('/')[-1]] = data['boxes'][i].tolist()
local('zip /Users/aub3/tests/yolo_test.zip -r /Users/aub3/tests/yolo_test/* ')
fname = "/Users/aub3/tests/yolo_test.zip"
name = "yolo_test"
f = SimpleUploadedFile(fname, file(fname).read(), content_type="application/zip")
dv = handle_uploaded_file(f, name)
extract_frames(TEvent.objects.create(video=dv).pk)
for df in Frame.objects.filter(video=dv):
for box in id_2_boxes[df.name]:
r = Region()
r.video = dv
r.frame = df
c , top_x, top_y, bottom_x, bottom_y = box
r.object_name = class_names[c]
r.region_type = Region.ANNOTATION
r.x = top_x
r.y = top_y
r.w = bottom_x - top_x
r.h = bottom_y - top_y
r.save()
l = RegionLabel()
l.frame = df
l.video = dv
l.label = labels[c]
l.region = r
l.save()
export_video(TEvent.objects.create(video=dv).pk)
try:
shutil.rmtree('/Users/aub3/tests/yolo_test')
except:
pass
def generate_vdn(fast=False):
kill()
setup_django()
from django.core.files.uploadedfile import SimpleUploadedFile
from dvaapp.views import handle_uploaded_file, handle_youtube_video
from dvaapp import models
from dvaapp.models import TEvent
from dvaapp.tasks import extract_frames, perform_face_detection_indexing_by_id, inception_index_by_id, \
perform_ssd_detection_by_id, perform_yolo_detection_by_id, inception_index_regions_by_id, \
export_video
dirname = get_coco_dirname()
local(
'wget https://www.dropbox.com/s/2dq085iu34y0hdv/coco_input.zip?dl=1 -O coco.zip'
)
local('unzip coco.zip')
with lcd(dirname):
local("zip coco_input.zip -r *.jpg")
fname = '{}/coco_input.zip'.format(dirname)
with open('{}/coco_sample_metadata.json'.format(dirname)) as datafile:
data = json.load(datafile)
f = SimpleUploadedFile("coco_input.zip",
file(fname).read(),
content_type="application/zip")
v = handle_uploaded_file(f, 'mscoco_sample_500')
extract_frames(TEvent.objects.create(video=v).pk)
video = v
models.Region.objects.all().filter(video=video).delete()
for frame in models.Frame.objects.all().filter(video=video):
frame_id = str(int(frame.name.split('_')[-1].split('.')[0]))
annotation = models.Region()
annotation.region_type = models.Region.ANNOTATION
annotation.video = v
annotation.frame = frame
annotation.full_frame = True
annotation.metadata = data[frame_id]['image']
annotation.object_name = 'metadata'
annotation.save()
for frame in models.Frame.objects.all().filter(video=video):
frame_id = str(int(frame.name.split('_')[-1].split('.')[0]))
for a in data[frame_id][u'annotations']:
annotation = models.Region()
annotation.region_type = models.Region.ANNOTATION
annotation.video = v
annotation.frame = frame
annotation.metadata = a
annotation.full_frame = False
annotation.x = a['bbox'][0]
annotation.y = a['bbox'][1]
annotation.w = a['bbox'][2]
annotation.h = a['bbox'][3]
annotation.object_name = 'coco_instance/{}/{}'.format(
a[u'category'][u'supercategory'], a[u'category'][u'name'])
annotation.save()
for a in data[frame_id][u'keypoints']:
annotation = models.Region()
annotation.region_type = models.Region.ANNOTATION
annotation.video = v
annotation.frame = frame
annotation.metadata = a
annotation.x = a['bbox'][0]
annotation.y = a['bbox'][1]
annotation.w = a['bbox'][2]
annotation.h = a['bbox'][3]
annotation.object_name = 'coco_keypoints/{}/{}'.format(
a[u'category'][u'supercategory'], a[u'category'][u'name'])
annotation.save()
for caption in data[frame_id][u'captions']:
annotation = models.Region()
annotation.region_type = models.Region.ANNOTATION
annotation.video = v
annotation.frame = frame
annotation.text = caption['caption']
annotation.full_frame = True
annotation.object_name = 'caption'
annotation.save()
if not fast:
inception_index_by_id(TEvent.objects.create(video=v).pk)
perform_ssd_detection_by_id(TEvent.objects.create(video=v).pk)
perform_face_detection_indexing_by_id(
TEvent.objects.create(video=v).pk)
inception_index_regions_by_id(TEvent.objects.create(video=v).pk)
export_video(TEvent.objects.create(video=v).pk)