def mapper(videohash, metadata):
print('videohash[%s]' % videohash)
print('hdfs_path[%s]' % metadata['hdfs_path'])
print 'mapper', videohash
filename = 'hardcodedvideo.' + metadata['extension']
#print filename, metadata.keys()
try:
picarus.io._record_to_file(metadata, filename)
except IOError:
hadoopy.counter('INPUT_ERROR', 'REMOTE READ FAILED')
return
min_interval = float(os.environ['MIN_INTERVAL'])
resolution = float(os.environ['RESOLUTION'])
try:
iter1 = lambda : viderator.frame_iter(filename, frame_skip=5, frozen=True)
iter2 = lambda : viderator.convert_video_ffmpeg(filename, frame_skip=5, frozen=True)
kf = keyframe.Histogram(min_interval)
# Do this instead of 'return' in order to keep the tempfile around
try:
for k, v in keyframes(iter1, iter2, metadata, kf, resolution):
#print 'yield', k
yield k, v
except:
hadoopy.counter('INPUT_ERROR', 'VIDEO_READ_ERROR')
return
finally:
os.remove(filename)
def map(self, key, value):
"""
Args:
key: Image name
value: Image as jpeg byte data
Yields:
A tuple in the form of (key, value)
key: Imagename-face-x0<x_tl_val>-y0<y_tl_val>-x1<x_br_val>-y1<y_br_val>
value: Cropped face binary data
"""
try:
image_pil, image_cv = self._load_cv_image(value)
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
return
try:
faces = self._detect_faces(image_cv)
except:
hadoopy.counter('DATA_ERRORS2', 'ImageLoadError')
return
if not faces:
return
if self._output_boxes:
yield key, (value, faces)
else:
for x0, y0, x1, y1 in faces:
image_pil_crop = image_pil.crop((x0, y0, x1, y1))
out_fp = StringIO.StringIO()
image_pil_crop.save(out_fp, 'JPEG')
out_fp.seek(0)
yield '%s-face-x0%d-y0%d-x1%d-y1%d' % (key, x0, y0, x1, y1), out_fp.read()
def collect(self,key,value):
if len(self.data) == 0:
self.first_key = key
if self.ncols == None:
self.ncols = len(value)
print >>sys.stderr, "Matrix size: %i columns"%(self.ncols)
else:
# TODO should we warn and truncate here?
# No. that seems like something that will introduce
# bugs. Maybe we could add a "liberal" flag
# for that.
assert(len(value) == self.ncols)
self.data.append(value)
self.nrows += 1
if len(self.data)>self.blocksize*self.ncols:
hadoopy.counter('Program','QR Compressions',1)
# compress the data
self.compress()
# write status updates so Hadoop doesn't complain
if self.nrows%50000 == 0:
hadoopy.counter('Program','rows processed',50000)
def map(self, event_filename, video_data):
hadoopy.counter('CombinedFeatures', 'DontHave')
sys.stderr.write('%s\n' % str(event_filename))
for event_filename, features in self.r.map(event_filename, video_data):
sys.stderr.write('%s\n' % str(event_filename))
for x in self.b.map(event_filename, features):
yield x
def _map(self, row, image_binary):
try:
image = Image.open(StringIO.StringIO(image_binary))
if not hasattr(image, "_getexif"):
yield row, json.dumps({})
else:
image_tags = image._getexif()
if image_tags is None:
yield row, json.dumps({})
else:
yield row, json.dumps(
dict(
(name, base64.b64encode(image_tags[id]))
if isinstance(image_tags[id], str)
else image_tags[id]
for id, name in TAGS.items()
if id in image_tags
)
)
except:
sys.stdout.flush()
hadoopy.counter("STATUS", "badRows")
else:
sys.stdout.flush()
hadoopy.counter("STATUS", "goodRows")
def map(self, image_hash, image_data):
"""
Args:
image_hash: Unique image string
image_data: Binary image data
Yields:
A tuple in the form of (classifier_name, label_value)
classifier_name: String representing the classifier
label_value: (label, feature) where label is an int
"""
try:
image = Image.open(StringIO.StringIO(image_data))
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
return
bgen = imfeat.BlockGenerator(image, imfeat.CoordGeneratorRectRotate,
output_size=(self._image_height, self._image_width),
step_delta=(self._image_height / 2, self._image_width / 2), angle_steps=1)
for num, (image_out, sim) in enumerate(bgen):
feature = np.asfarray(imfeat.compute(self._feat, image_out)[0])
pred = dict((classifier_name, classifier.predict(feature))
for classifier_name, classifier in self._classifiers)
if any(x for x in pred.values() if x[0][0] * x[0][1] > 0): # At least 1 class needs to be > 0
image_out_fp = StringIO.StringIO()
imfeat.convert_image(image_out, ['RGB']).save(image_out_fp, 'JPEG')
image_out_fp.seek(0)
yield (image_hash, sim), (pred, image_out_fp.read())
def map(self, key, value):
"""
Args:
key: Image name
value: Image as jpeg byte data
Yields:
A tuple in the form of (key, value)
key: Constant dummy value
value: (l2sqr_dist, value)
"""
try:
image = imfeat.resize_image(imfeat.image_fromstring(value, {'type': 'numpy', 'mode': 'bgr', 'dtype': 'uint8'}),
100, 100)
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
return
# Distance metric
diff = image - self.target_image
dist = np.sum(diff * diff)
# Output
if dist < self.min_dist:
self.min_dist = dist
self.min_key = key
self.min_value = value
def reduce(self, image_hash, values):
"""
Args:
image_hash: (see mapper)
values: Iterator of values (see mapper)
Yields:
A tuple in the form of (image_hash, value)
image_hash: Image hash
value: The provided value (not the prediction)
"""
predictions = None
out_val = None
for value in values:
if isinstance(value, dict):
predictions = value
else:
out_val = value
if predictions is None or out_val is None:
hadoopy.counter('DATA_ERR', 'MISSING_PREDICTIONS_OR_DATA')
return
label, conf = predictions[self._class_name][0]
if (self._class_thresh <= label * conf) == (self._output_class == 1): # Both true or both false
yield image_hash, out_val
def convert_matrix(self, matrix):
sparsity = len(matrix.nonzero()[0]) / float(matrix.size)
print('Sparsity[%f]' % sparsity)
if sparsity < self.min_sparsity:
hadoopy.counter('SPARSITY', 'SPARSE')
return sp.sparse.csr_matrix(matrix)
hadoopy.counter('SPARSITY', 'DENSE')
return matrix
def _map(self, row, input_binary):
try:
yield row, self.job.process_binary(input_binary)
except:
sys.stdout.flush()
hadoopy.counter('ERROR', 'BadRow')
else:
hadoopy.counter('STATUS', 'GoodRow')
def _map(self, row, image_binary):
try:
image = imfeat.image_fromstring(image_binary)
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
out = self.sp(image, 25)
hulls = self.sp.label_image_to_contours(out, 1.)
yield row, json.dumps(hulls, separators=(',', ':'))
def map(self, image_id, image_binary):
image = resize(imfeat.image_fromstring(image_binary))
print(image.shape)
st = time.time()
box_num = -1
for box_num, (box, f) in enumerate(feature.image_patch_features_dense(image, normalize_box=True)):
yield (image_id, box.tolist()), np.dot(self.coefs, f.reshape((f.size, 1))).ravel() + self.intercepts
hadoopy.counter('stats', 'num_boxes', box_num + 1)
print('ImageTime[%f]' % (time.time() - st))
def amap(self, key, feat):
feat = np.array([np.fromstring(feat, dtype=np.float32)])
if self.canopies.size:
nearest_dist = self.nn(feat, self.canopies)[1]
if nearest_dist > self.hard_dist:
hadoopy.counter('canopy_cluster', 'canopy_count')
self.canopies = np.concatenate((self.canopies, feat))
else:
hadoopy.counter('canopy_cluster', 'canopy_count')
self.canopies = feat
def map(self, name, image_or_data):
if isinstance(image_or_data, str):
try:
image = imfeat.image_fromstring(image_or_data)
except:
hadoopy.counter("DATA_ERRORS", "ImageLoadError")
return
else:
image = image_or_data
yield name, self._feat(image)
def _map(self, row, image_binary):
try:
if not image_binary:
raise ValueError
image = imfeat.image_fromstring(image_binary)
except:
hadoopy.counter('ERROR', 'FEATURE')
print('Error on row[%r]' % row)
else:
yield row, picarus.api.np_tostring(self._feat.compute_feature(image))
def reducer(k, vs):
for v in vs:
print(type(v))
if isinstance(v, dict):
metadata = v
else:
feature_hash = v
try:
yield k, (feature_hash, metadata)
except NameError:
hadoopy.counter("ERRORS", "JoinsFailed")
def map(self, url, value):
try:
data = download_file(url)
except Exception:
hadoopy.counter('FILE_DOWNLOADER', 'Exception')
else:
if self.output_type == 'meta':
yield url, (data, value)
elif self.output_type == 'image':
yield url, data
else:
raise ValueError('OutputType[%s]' % self.output_type)
def map(self, name, image_data):
try:
image = Image.open(StringIO.StringIO(image_data))
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
return
image = image.resize((self._image_length, self._image_length))
try:
yield name, np.asfarray(imfeat.compute(self._feat, image)[0])
except ValueError, e:
print(e)
hadoopy.counter('DATA_ERRORS', 'UnkImageType')
return
def train(classifier_name, classifier_extra, label_values):
import classipy
label_values = list(label_values)
hadoopy.counter('FeatureShape', str(len(label_values[0][1])))
if classifier_name == 'svmlinear':
return classipy.SVMLinear(options={'B': '1'}).train(label_values)
elif classifier_name == 'svm':
return classipy.SVM(options={'t': '2'}).train(label_values)
elif classifier_name == 'svm_hik':
return classipy.SVMScikit(kernel=classipy.kernels.histogram_intersection).train(label_values)
elif classifier_name == 'svmlinear_autotune':
def wrapped_optimizer(*args, **kw):
for x in pyram.exponential_grid(*args, **kw):
hadoopy.counter('X-Val', 'Rounds')
yield x
b = classipy.select_parameters(classipy.SVMLinear, label_values,
{'c': (10**-2, 10**1, 10)},
wrapped_optimizer,
options={'B': '1'})[1]
print(b)
return classipy.SVMLinear(b).train(label_values)
elif classifier_name == 'plslinearsvmxval':
num_dims = label_values[0][1].size
# Set the parameters by cross-validation
#,'pls__n_components': [x for x in [1, 8, 16, 32, 64, 128, 256] if x <= num_dims]
#('pls', sklearn.pls.PLSRegression(n_components=0)),
tuned_parameters = [{'svm__C': [.001, .01, .1, 1, 10, 100]}]
p = sklearn.pipeline.Pipeline([('svm', sklearn.svm.SVC(kernel=classipy.kernels.histogram_intersection, scale_C=True))]) # was cls
#p = sklearn.grid_search.GridSearchCV(cls, tuned_parameters, score_func=sklearn.metrics.f1_score)
num_neg = 0
num_pos = 0
import random
random.shuffle(label_values)
new_label_values = []
for l, v in label_values:
if l == 1:
if num_pos < 100:
new_label_values.append((l, v))
num_pos += 1
else:
if num_neg < 100:
new_label_values.append((l, v))
num_neg += 1
import sys
sys.stderr.write('Num Neg[%d] Pos[%d]\n' % (num_neg, num_pos))
p.fit(*zip(*new_label_values)[::-1])
return p # p.best_estimator_
else:
raise ValueError('Unknown classifier [%s]' % classifier_name)
def map(self, image_id, image_binary):
if self.num_inputs <= 0:
return
self.num_inputs -= 1
pyramid = np.zeros((len(self.ids), np.sum(self.num_bins_sqr)), dtype=np.int32)
num_boxes = 0
for (image_id, box), confs in super(Mapper, self).map(image_id, image_binary):
num_boxes += 1
cy = (box[2] + box[0]) / 2
cx = (box[1] + box[3]) / 2
offset = 0
cur_bins = []
for l in range(self.levels):
cur_bins.append(offset + int(cy * self.num_bins[l]) * self.num_bins[l] + int(cx * self.num_bins[l]))
offset += self.num_bins_sqr[l]
#if num_boxes < 1000 and num_boxes % 100:
# print((box, cy, cx, cur_bins))
inds = (confs >= 0).nonzero()[0]
hadoopy.counter('STATS', 'num_pos', inds.size)
hadoopy.counter('STATS', 'num_neg', confs.size - inds.size)
hadoopy.counter('STATS', 'total', confs.size)
if inds.size:
for cur_bin in cur_bins:
pyramid[inds, cur_bin] += 1
hadoopy.counter('STATS', 'sz-%s' % str(pyramid.shape))
if np.any(pyramid):
pyramid = pyramid * (self.bin_weight / float(num_boxes))
for exemplar_num, row in enumerate(pyramid):
yield exemplar_num, (image_id, row)
def map(self, key, feat):
stime = time.time()
feat = self._strto2d(feat)
self.ftime += time.time() - stime
stime = time.time()
if self.canopies.size:
nearest_dist = self.nn(feat, self.canopies)[1]
if nearest_dist > self.hard_dist:
hadoopy.counter('canopy_cluster', 'canopy_count')
self.canopies = np.concatenate((self.canopies, feat))
else:
hadoopy.counter('canopy_cluster', 'canopy_count')
self.canopies = feat
self.gtime += time.time() - stime
def map(self, image_name, image_label_points):
"""
Args:
image_name: A string (if not then we skip the input)
image_label_points: (image, [(label, points), ...]) where points is Nx2 (y, x)
"""
if not isinstance(image_name, str):
hadoopy.counter('SKIPPED_INPUTS', 'KeyNotString')
return
with self.timer('Build root_labels_image_points'):
image, label_points = image_label_points
root_labels_image_points = {}
if self.dp:
root_label_points = self.dp.group_lowest_nodes(image, label_points)
for root, label_points in root_label_points:
labels = np.array([x[0] for x in label_points], dtype=np.int32)
image_points = [(image, x[1]) for x in label_points]
root_labels_image_points[root] = labels, image_points
else:
labels = np.array([x[0] for x in label_points], dtype=np.int32)
image_points = [(image, x[1]) for x in label_points]
root_labels_image_points[0] = labels, image_points
with self.timer('Run train_map_hists and sum qlss/qrss'):
for root, (labels, image_points) in root_labels_image_points.items():
if self.level != int(np.floor(np.log2(root + 1))): # We are done processing this root
continue
qls, qrs = train_map_hists(labels, image_points, self.feats, self.num_classes)
qls, qrs = self.convert_matrix(qls), self.convert_matrix(qrs)
try:
self.root_count[root] += 1
except KeyError:
self.root_count[root] = 1
try:
try:
self.qlss[root] += qls
except NotImplementedError:
self.qlss[root] = self.qlss[root] + qls
try:
self.qrss[root] += qrs
except NotImplementedError:
self.qrss[root] = self.qrss[root] + qrs
self.num_images[root] += 1
except KeyError:
if self.max_root_buffer <= len(self.qlss):
for x in self.flush_node(root):
yield x
self.qlss[root] = qls
self.qrss[root] = qrs
self.num_images[root] = 1
def stop_time(self, name):
try:
dur = time.time() - self._pending_times[name]
except KeyError:
hadoopy.counter('stop_time', 'timer_failed')
else:
try:
time_stats = self._times[name]
# Min/Max/Sum/Count
self._times[name] = [min(time_stats[0], dur),
max(time_stats[1], dur),
time_stats[2] + dur,
time_stats[3] + 1]
except KeyError:
self._times[name] = [dur, dur, dur, 1]
def map(self, image_id, image_binary):
image = imfeat.image_fromstring(image_binary)
print(image.shape)
st = time.time()
box_num = -1
for box_num, (box, f) in enumerate(feature.image_patch_features_dense(image, normalize_box=True)):
scores = np.dot(self.coefs, f.reshape((f.size, 1))) + self.intercepts
pred_common = [image_id, box.tolist(), f.tolist()]
for score, preds in zip(scores, self.preds):
pred = self.output_formatter([float(score[0])] + pred_common)
if len(preds) >= self.max_hard:
heapq.heappushpop(preds, pred)
else:
heapq.heappush(preds, pred)
hadoopy.counter("stats", "num_boxes", box_num + 1)
print("ImageTime[%f]" % (time.time() - st))
def mapper(key, value):
"""
Args:
key, value:
either of - the output of clustering (/partition or /samples),
- the output of video keyframing (/allframes)
- (hash,image_data) input
- (hash,image_metadata) input
Env vars:
IMAGE_TYPE: record, cluster, kv, frame
THUMB_SIZE: longest dimension for the thumbnail images
Yields:
image_hash, image_data: serialized jpeg data for thumbnail
"""
if os.environ['IMAGE_TYPE'] == 'record':
# Image input format
image_hash, image_metadata = key, value
image_file = picarus.io._record_to_file(image_metadata)
elif os.environ['IMAGE_TYPE'] == 'cluster':
# Cluster output /partition or /sample
cluster_index, (image_hash, image_data) = key, value
image_file = StringIO.StringIO(image_data)
elif os.environ['IMAGE_TYPE'] == 'kv':
# hash, image bytes
image_hash, image_data = key, value
image_file = StringIO.StringIO(image_data)
elif os.environ['IMAGE_TYPE'] == 'frame':
image_hash, image_metadata = key, value
image_file = StringIO.StringIO(image_metadata['image_data'])
thumb_size = int(os.environ['THUMB_SIZE'])
try:
image = Image.open(image_file)
image.thumbnail((thumb_size, thumb_size))
except:
hadoopy.counter('INPUT_ERROR', 'IMAGE_DATA_ERROR')
image = image.convert('RGB')
s = StringIO.StringIO()
image.save(s, 'JPEG')
s.seek(0)
# Output type: kv
yield image_hash, s.buf
def map(self, event_filename, video_data):
"""
Args:
event_filename: Tuple of (event, filename)
video_data: Binary video data
Yields:
A tuple in the form of ((event, filename, frame_num, frame_time), frame_data)
"""
ext = '.' + event_filename[1].rsplit('.')[1]
event, filename = event_filename
heap = [(float('-inf'), None)] * self.max_outputs
with tempfile.NamedTemporaryFile(suffix=ext) as fp:
fp.write(video_data)
fp.flush()
sys.stderr.write('Prevideo\n')
try:
for frame_num, frame_time, frame in viderator.frame_iter(fp.name,
frozen=True,
frame_skip=self.frame_skip):
sys.stderr.write('FrameNum[%d]\n' % frame_num)
if frame_num >= self.max_frames_per_video:
break
frame_orig = frame
if self.remove_bars:
sz = self.remove_bars.find_bars(frame)
frame = frame[sz[0]:sz[1], sz[2]:sz[3], :]
if not frame.size: # Empty
continue
st = time.time()
c = self._feat(frame)[0]
sys.stderr.write('FrameTime[%f]\n' % (time.time() - st))
print('FrameTime[%f]' % (time.time() - st))
if c > heap[0][0]:
if self.output_frame:
heapq.heappushpop(heap,
(c, ((event, filename, frame_num, frame_time), imfeat.image_tostring(frame_orig, 'JPEG'))))
else:
heapq.heappushpop(heap,
(c, ((event, filename, frame_num, frame_time), '')))
except IOError:
hadoopy.counter('PICARUS', 'CantProcessVideo')
for x in heap[-self.max_outputs_per_video:]:
heapq.heappushpop(self.heap, x)
def map(self, event_filename, video_data):
"""
Args:
event_filename: Tuple of (event, filename)
video_data: Binary video data
Yields:
A tuple in the form of ((event, filename), value) where value is a dict
with contents
prev_frame_time:
prev_frame_num:
prev_frame:
frame_time:
frame_num:
frame:
"""
ext = '.' + event_filename[1].rsplit('.')[1]
with tempfile.NamedTemporaryFile(suffix=ext) as fp:
fp.write(video_data)
fp.flush()
prev_frame_time = None
prev_frame_num = None
prev_frame = None
try:
for (frame_num, frame_time, frame), iskeyframe in self.kf(viderator.frame_iter(fp.name,
frame_skip=self.frame_skip,
frozen=True)):
if self.max_time < frame_time:
break
if iskeyframe and prev_frame is not None:
yield event_filename, {'prev_frame_time': prev_frame_time,
'prev_frame_num': prev_frame_num,
'prev_frame': imfeat.image_tostring(prev_frame, 'JPEG'),
'frame_time': frame_time,
'frame_num': frame_num,
'frame': imfeat.image_tostring(frame, 'JPEG')}
prev_frame_num = frame_num
prev_frame = frame
except Exception, e:
print(e)
hadoopy.counter('VIDEO_ERROR', 'FFMPEGCantParse')
def map(self, key, value):
"""
Args:
key: Image name
value: Image as jpeg byte data
Yields:
A tuple in the form of (key, value)
key: Image name
value: (image, faces) where image is the input value and faces is
a list of ((x, y, w, h), n)
"""
try:
image = self._load_cv_image(value)
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
return
dist = self._compute_face_distance(image)
yield dist, (key, value)
def map(self, image_hash, feature):
"""
Args:
image_hash: Unique image string
feature: Numpy image feature
Yields:
A tuple in the form of (classifier_name, label_value)
classifier_name: String representing the classifier
label_value: (label, feature) where label is an int
"""
try:
class_labels = self._hash_class_labels[image_hash]
except KeyError:
hadoopy.counter('DATA_ERRORS', 'UNKNOWN_IMAGE_HASH')
return
for classifier_name, label in class_labels:
yield classifier_name, (label, feature)
def _image_from_str(s):
"""Load from string, crop to a square, resize to _initial_image_size
Args:
s: String of bytes representing a JPEG image
Returns:
RGB Image with height/width as _initial_image_size
Raises:
ValueError: Image is height/width too small (< _initial_image_size)
or mode isn't RGB
IOError: Image is unreadable
"""
try:
img = Image.open(StringIO.StringIO(s))
except IOError, e:
hadoopy.counter('Stats', 'IMG_BAD')
raise e
def wrapped_optimizer(*args, **kw):
for x in pyram.exponential_grid(*args, **kw):
hadoopy.counter('X-Val', 'Rounds')
yield x
def _map(self, row, image_binary):
try:
image = imfeat.image_fromstring(image_binary)
yield row, imfeat.image_tostring(imfeat.resize_image_max_side(image, self.max_side), 'jpg')
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
def reduce(self, exemplar_num, id_rows):
out = np.hstack([x[1].ravel()
for x in sorted(id_rows, key=lambda x: x[0])])
hadoopy.counter('STATS', 'sz-%s' % str(out.shape))
yield exemplar_num, out
def test_counter(self):
def err(x):
self.assertEqual('reporter:counter:a,b,5\n', x)
hadoopy.counter('a', 'b', 5, err=err)
def _map(self, row, image_binary):
try:
image = imfeat.image_fromstring(image_binary)
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
yield row, picarus.api.np_tostring(self._feat(image))
def reducer(event, filename_predicates):
yield event, dict(filename_predicates)
hadoopy.counter('SkippingTaskCounters', 'ReduceProcessedGroups')
class Mapper(object):
def __init__(self):
_target_image = cv2.imread('target.jpg')
_target_image = cv2.resize(
_target_image,
(_target_image.shape[1] // _tile_length * _tile_length,
_target_image.shape[0] // _tile_length * _tile_length))
self.target_tiles = {}
ytiles = _target_image.shape[0] / _tile_length
xtiles = _target_image.shape[1] / _tile_length
print('Xtiles[%d] Ytiles[%d]' % (xtiles, ytiles))
assert xtiles > 0 and ytiles > 0
xsubtiles = xtiles * _subtiles_per_tile_length
ysubtiles = ytiles * _subtiles_per_tile_length
self.min_dists = {}
self.dist = lambda x, y: np.sum(np.abs(x - y)) # distpy.L2Sqr().dist
for y in xrange(ysubtiles):
for x in xrange(xsubtiles):
# Defines which tile the subtile is in
#'%.6d_%.6d' %
tile_id = (x / _subtiles_per_tile_length,
y / _subtiles_per_tile_length)
# Defines which position it is in within the tile
#'%.6d_%.6d' %
subtile_id = (x % _subtiles_per_tile_length,
y % _subtiles_per_tile_length)
#'\t'.join(
key = (tile_id[0], tile_id[1], subtile_id[0], subtile_id[1])
yp = ysubtiles - y - 1 # NOTE(brandyn): Flip coordinates for y axis
tile = _target_image[(yp * _subtile_length):((yp + 1) *
_subtile_length),
(x * _subtile_length):(x + 1) *
_subtile_length, :]
self.target_tiles[key] = np.asfarray(tile)
@staticmethod
def _crop_image_from_str(s):
"""Load from string, crop to a square, resize to _initial_image_size
Args:
s: String of bytes representing a JPEG image
Returns:
RGB Image with height/width as _initial_image_size
Raises:
ValueError: Image is height/width too small (< _initial_image_size)
or mode isn't RGB
IOError: Image is unreadable
"""
if isinstance(s, tuple):
s = s[0]
try:
img = imfeat.image_fromstring(s)
except IOError, e:
hadoopy.counter('Stats', 'IMG_BAD')
raise e
min_side = min(img.shape[:2])
if min_side < _initial_image_size:
hadoopy.counter('Stats', 'IMG_TOO_SMALL')
raise ValueError
if img.ndim != 3:
hadoopy.counter('Stats', 'IMG_WRONG_MODE')
raise ValueError
return imfeat.resize_image(img, _initial_image_size,
_initial_image_size)
def map(self, event_filename, video_data):
"""
Args:
event_filename: Tuple of (event, filename)
video_data: Binary video data
Yields:
A tuple in the form of ((event, filename), features) where features is a dict
frame_features: List of frame features
file_size: Size in bytes
where each frame feature is a dictionary of
frame_time: Time in seconds
frame_num: Frame number
prev_frame_num: Previous frame number (useful if there is a frame skip)
keyframe: Boolean True/False
surf: List of surf points (see impoint)
face_widths:
face_heights:
predictions: Dictionary of predictions
"""
sys.stderr.write('In Raw:%s\n' % str(event_filename))
print(event_filename)
ext = '.' + event_filename[1].rsplit('.')[1]
with tempfile.NamedTemporaryFile(suffix=ext) as fp:
with self.timer('Writing video data'):
fp.write(video_data)
fp.flush()
kf = keyframe.DecisionTree(min_interval=0)
kf.load()
prev_frame = None
prev_frame_num = 0
all_out = []
sz = len(video_data)
self.timer.start('KF')
try:
for (frame_num, frame_time, frame), iskeyframe in kf(
viderator.frame_iter(fp.name, frozen=True)):
hadoopy.counter('RawFeatures', 'NumFrames')
self.timer.stop('KF')
print(frame_time)
if frame_num > self._max_frames:
break
if frame_num % 100 == 0:
with self.timer('Computing face features'):
faces = _detect_faces(
imfeat.convert_image(frame,
[('opencv', 'gray', 8)]),
self.cascade)
else:
faces = {}
out = {
'frame_time': frame_time,
'frame_num': frame_num,
'prev_frame_num': prev_frame_num,
'keyframe': iskeyframe,
'surf': kf.prev_vec['surf']
}
if faces: # If any faces
face_heights = np.array([x[0][3] for x in faces
]) / float(frame.height)
face_widths = np.array([x[0][2] for x in faces
]) / float(frame.width)
out['face_widths'] = face_widths
out['face_heights'] = face_heights
# Output the cur and previous frames if this is a keyframe
if iskeyframe and np.random.random(
) < self._frame_output_prob:
out['prev_frame'] = cv_to_jpg(prev_frame)
out['frame'] = cv_to_jpg(frame)
# Compute scene features
with self.timer('Computing scene classifier features'):
frame_res = cv.fromarray(
cv2.resize(
np.asarray(cv.GetMat(frame)),
(self._image_width, self._image_height)))
feature = self._feat(frame_res)
out['predictions'] = dict(
(classifier_name, classifier.predict(feature)) for
classifier_name, classifier in self._classifiers)
# Output JPEG with match lines from the SURF feature
if np.random.random(
) < self._match_line_prob and prev_frame:
out['surf_image'] = cv_to_jpg(
plot_matches(prev_frame,
kf.surf_debug['matches'],
kf.surf_debug['points0'],
kf.surf_debug['points1'],
max_feat_width=kf.max_feat_width))
# Output data buffer
all_out.append(out)
if len(all_out) >= self._block_size:
with self.timer('Yield'):
yield event_filename, {
'frame_features': all_out,
'file_size': sz
}
all_out = []
prev_frame = frame
prev_frame_num = frame_num
self.timer.start('KF')
except viderator.FPSParseException: # NOTE(brandyn): This will disregard videos with this error
hadoopy.counter('SkippedVideos', 'FPSParseException')
return
if all_out:
with self.timer('Yield'):
yield event_filename, {
'frame_features': all_out,
'file_size': sz
}