tokens = line.strip().split(',')
url = tokens[0]
im_name = get_fpath_filename(url)
im_path = join(im_dir, im_name)
url = url.replace('{','').replace('}','')
tokens[0] = url
# Can also use quality_utils to get some blurriness metrics
cv_im = cv2.imread(im_path, 0)
if cv_im is None:
continue
blurriness = get_blurriness(get_laplacian(cv_im))
x = load_and_preprocess_image(im_path, x_size, y_size, True)
if x is False:
continue
base_features = base_model.predict(x)
menu_preds = menu_model.predict(base_features)
menu_pred, menu_prob = _decode_predictions(menu_preds, mapping)[0]
tokens.extend([blurriness, menu_pred, str(menu_prob)])
output = ','.join([str(x) for x in tokens]) + '\n'
outf.write(output)
num_images += 1
def create_pairs_helper(pos_pairs, all_image_data, image_dir, year, month,
limit):
pairs = []
start_time = time.time()
total_count = 0
batch_size = 1600
# limit = 35200
output_dir = '/tigress/dchouren/thesis/resources/pairs'
output_file = join(output_dir,
year + '_' + month + '_' + str(limit) + '.h5')
print(len(pos_pairs))
with h5py.File(output_file, 'w') as f:
maxshape = (None, ) + (2, 3, 224, 224)
dset = f.create_dataset('pairs',
shape=(batch_size, 2, 3, 224, 224),
maxshape=maxshape,
compression='gzip')
row_count = 0
for i, (base_data, pos_data) in enumerate(pos_pairs):
base_filename = base_data[2].split('/')[-1]
if base_filename.endswith('zz=1'):
continue
base_image = vutils.load_and_preprocess_image(join(
image_dir, base_filename),
dataset='imagenet',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
if base_image is None:
continue
pos_filename = pos_data[2].split('/')[-1]
if pos_filename.endswith('zz=1'):
continue
pos_image = vutils.load_and_preprocess_image(join(
image_dir, pos_filename),
dataset='imagenet',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
if pos_image is None:
continue
possible_neg_image = all_image_data[np.random.choice(
all_image_data.shape[0])]
# ipdb.set_trace()
distance = meter_distance(possible_neg_image[:2], base_data[:2])
# print('finding negative')
i = 0
while distance < 2000:
# print(distance)
possible_neg_image = all_image_data[np.random.choice(
all_image_data.shape[0])]
distance = meter_distance(possible_neg_image[:2],
base_data[:2])
i += 1
neg_filename = possible_neg_image[2].split('/')[-1]
if neg_filename.endswith('zz=1'):
continue
neg_image = vutils.load_and_preprocess_image(join(
image_dir, neg_filename),
dataset='imagenet',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
if neg_image is None:
continue
pos_pair = [base_image, pos_image]
neg_pair = [base_image, neg_image]
# ipdb.set_trace()
pairs += [pos_pair]
pairs += [neg_pair]
# print('{}: {}'.format(len(pairs), int(time.time() - start_time)))
if len(pairs) == batch_size:
# print('here')
# ipdb.set_trace()
pairs = np.squeeze(np.asarray(pairs))
# print('squeezed pairs')
dset[row_count:] = pairs
# print('assigned dset')
row_count += pairs.shape[0]
dset.resize(row_count + pairs.shape[0], axis=0)
# print('resized dset')
# np.savez_compressed(open(output_file, 'wb'), pairs)
labels = []
pairs = []
total_count += batch_size
print(dset.shape)
if total_count >= limit:
# print('ending')
dset.resize(limit, axis=0)
print(dset.shape)
break
if dset.shape[0] == limit:
f.close()
print('{} | Saved to {}'.format(int(time.time() - start_time),
output_file))
return True
else:
del dset
del f
print('False')
return False
import vision_utils as vutils image_path = '/Users/daway/Documents/Princeton/Thesis2017/4392232665_90c251d33c.jpg' x = vutils.load_and_preprocess_image(image_path, preprocess=False, rescale=True) print(x.shape)
import vision_utils as vutils
import time
im_path = '/scratch/network/dchouren/images/2012/01/8331121208_daaf6fa41c.jpg'
start_time = time.time()
im = vutils.load_and_preprocess_image(im_path,
dataset='imagenet',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
def load_pairs(same_pairs, diff_pairs, image_dir, output_file):
pairs = []
labels = []
total_count = 0
batch_size = 1600
start_time = time.time()
with h5py.File(output_file, 'w') as f:
maxshape = (None, ) + (2, 3, 224, 224)
dset = f.create_dataset('pairs',
shape=(batch_size, 2, 3, 224, 224),
maxshape=maxshape)
labels_dset = f.create_dataset('labels',
shape=(batch_size, 1),
maxshape=(None, 1))
row_count = 0
# ipdb.set_trace()
data = np.empty((same_pairs.shape[0] * 2, same_pairs.shape[1]),
dtype=same_pairs.dtype)
# ipdb.set_trace()
data[0::2] = same_pairs
data[1::2] = diff_pairs
# ipdb.set_trace()
for i, (base_data, pair_data) in enumerate(data):
# ipdb.set_trace()
base_filename = base_data[2].split('/')[-1]
year, month = base_data[3].split('-')[0], base_data[3].split(
'-')[1]
if base_filename.endswith('zz=1'):
continue
base_image = vutils.load_and_preprocess_image(join(
image_dir, year, month, month, base_filename),
dataset='imagenet',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
if base_image is None:
# print(i)
continue
pair_filename = pair_data[2].split('/')[-1]
year, month = pair_data[3].split('-')[0], pair_data[3].split(
'-')[1]
if pair_filename.endswith('zz=1'):
continue
pair_image = vutils.load_and_preprocess_image(join(
image_dir, year, month, month, pair_filename),
dataset='imagenet',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
if pair_image is None:
# print(i)
continue
pair = [base_image, pair_image]
if i % 2 == 0:
label = 1
else:
label = 0
pairs += [pair]
labels += [label]
if len(pairs) == batch_size:
pairs = np.squeeze(np.asarray(pairs))
dset[row_count:] = pairs
labels = np.array(labels)
labels_dset[row_count:] = np.reshape(labels,
(labels.shape[0], 1))
row_count += pairs.shape[0]
dset.resize(row_count + pairs.shape[0], axis=0)
labels_dset.resize(row_count + pairs.shape[0], axis=0)
labels = []
pairs = []
total_count += batch_size
if dset.shape[0] >= data.shape[0]:
dset.resize(data.shape[0], axis=0)
print('{} | Saved to {}'.format(int(time.time() - start_time),
output_file))
def load_middlebury(match_pairs, image_dir):
middlebury_pairs = []
middlebury_dir = '/tigress/dchouren/thesis/resources/stereo/images'
filenames = sorted(os.listdir(middlebury_dir))
quads = list(
zip(filenames[::4], filenames[1::4], filenames[2::4], filenames[3::4]))
all_pairs = []
stereo_pairs = []
for quad in quads:
pairs = [(quad[0], quad[1]), (quad[0], quad[2]), (quad[0], quad[3])]
all_pairs.extend(pairs)
stereo_pairs.extend([(quad[0], quad[1])])
all_middlebury_images = []
all_stereo_images = []
for base_file, pos_file in all_pairs:
base_image = vutils.load_and_preprocess_image(join(
middlebury_dir, base_file),
dataset='imagenet',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
pos_image = vutils.load_and_preprocess_image(join(
middlebury_dir, pos_file),
dataset='imagenet',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
all_middlebury_images.append([base_image, pos_image])
for base_file, pos_file in stereo_pairs:
base_image = vutils.load_and_preprocess_image(join(
middlebury_dir, base_file),
dataset='imagenet',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
pos_image = vutils.load_and_preprocess_image(join(
middlebury_dir, pos_file),
dataset='imagenet',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
all_stereo_images.append([base_image, pos_image])
all_middlebury_images = np.squeeze(np.array(all_middlebury_images))
all_stereo_images = np.squeeze(np.array(all_stereo_images))
# ipdb.set_trace()
pairs = []
labels = []
total_count = 0
batch_size = 1600
# ipdb.set_trace()
print(match_pairs.shape[0])
start_time = time.time()
with h5py.File(output_file, 'w') as f:
maxshape = (None, ) + (2, 3, 224, 224)
dset = f.create_dataset('pairs',
shape=(batch_size, 2, 3, 224, 224),
maxshape=maxshape)
labels_dset = f.create_dataset('labels',
shape=(batch_size, 1),
maxshape=(None, 1))
row_count = 0
for i, (base_data, pair_data) in enumerate(
np.vstack(zip(match_pairs, match_pairs))):
if i % 2 == 0:
pair = all_stereo_images[np.random.choice(
all_stereo_images.shape[0])]
label = 1
else:
base_filename = base_data[2].split('/')[-1]
year, month = base_data[3].split('-')[0], base_data[3].split(
'-')[1]
base_image = vutils.load_and_preprocess_image(join(
image_dir, year, month, month, base_filename),
dataset='flickr',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
if base_image is None:
continue
pair_filename = pair_data[2].split('/')[-1]
year, month = pair_data[3].split('-')[0], pair_data[3].split(
'-')[1]
pair_image = vutils.load_and_preprocess_image(join(
image_dir, year, month, month, pair_filename),
dataset='flickr',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
if pair_image is None:
continue
pair = np.squeeze(np.array([base_image, pair_image]))
label = 0
pairs += [pair]
labels += [label]
if len(pairs) == batch_size:
# print('here')
# pairs = np.squeeze(np.asarray(pairs))
# ipdb.set_trace()
pairs = np.asarray(pairs)
dset[row_count:] = pairs
labels = np.array(labels)
labels_dset[row_count:] = np.reshape(labels,
(labels.shape[0], 1))
row_count += pairs.shape[0]
dset.resize(row_count + pairs.shape[0], axis=0)
labels_dset.resize(row_count + pairs.shape[0], axis=0)
labels = []
pairs = []
total_count += batch_size
print('{} | Saved to {}'.format(int(time.time() - start_time),
output_file))
def create_pairs_helper(pos_pairs, all_image_data, image_dir, output_dir):
pairs = []
labels = []
# pos_pair_images = []
# neg_pair_images = []
# image_dir = '/scratch/network/dchouren/images/2014/01_new/01_new'
last_time = time.time()
count = 0
for i, (base_data, pos_data) in enumerate(pos_pairs):
base_filename = base_data[2].split('/')[-1]
base_image = vutils.load_and_preprocess_image(join(
image_dir, base_filename),
dataset='imagenet',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
pos_filename = pos_data[2].split('/')[-1]
pos_image = vutils.load_and_preprocess_image(join(
image_dir, pos_filename),
dataset='imagenet',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
possible_neg_image = all_image_data[np.random.choice(
all_image_data.shape[0])]
while meter_distance(possible_neg_image[:2], base_data[:2]) < 2000:
possible_neg_image = all_image_data[np.random.choice(
all_image_data.shape[0])]
neg_filename = possible_neg_image[2].split('/')[-1]
neg_image = vutils.load_and_preprocess_image(join(
image_dir, neg_filename),
dataset='imagenet',
x_size=224,
y_size=224,
preprocess=False,
rescale=True)
if base_image is None or pos_image is None or neg_image is None:
continue
pos_pair = [base_image, pos_image]
neg_pair = [base_image, neg_image]
# ipdb.set_trace()
pairs += [pos_pair]
pairs += [neg_pair]
labels += [1, 0]
if i % 100 == 0:
print(i, time.time() - last_time)
if len(labels) == 10000:
np.save(open(join(output_dir, 'pairs_' + str(count) + '.npy')),
pairs)
labels = []
pairs = []
