def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.data_root = params['data_root']
self.im_shape = params['im_shape']
self.split = params['split']
self.mean_file = params['mean_file']
list_file = '../prepare_data/AFEW-VA/crop/{}_data.txt'.format(
self.split)
self.indexlist = [line.split(' ')[0] for line in open(list_file)]
# get list of image indexes.
# Read the mat file and assign to X
mat_contents = scipy.io.loadmat(
'../prepare_data/AFEW-VA/crop/{}_labels.mat'.format(self.split))
self.X = np.zeros(mat_contents['{}_labels'.format(self.split)].shape,
dtype=np.float16)
self.X[:, :] = mat_contents['{}_labels'.format(self.split)]
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer(self.mean_file)
print("BatchLoader initialized with {} images".format(
len(self.indexlist)))
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.number_classes = params['number_classes']
self.img_ext = params['img_ext']
self.dataset_root = params['dataset_root']
self.im_shape = params['im_shape']
# get list of image indexes.
list_file = params['split'] + '.csv'
self.indexlist = [
line.rstrip('\n').split(',')[0]
for line in open(osp.join(self.dataset_root, list_file))
]
self.labellist = [
line.rstrip('\n').split(',')[1]
for line in open(osp.join(self.dataset_root, list_file))
]
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
print("BatchLoader initialized with {} images".format(
len(self.indexlist)))
class BatchLoader(object):
"""
This class abstracts away the loading of images.
Images can either be loaded singly, or in a batch. The latter is used for
the asyncronous data layer to preload batches while other processing is
performed.
"""
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.pascal_root = params['pascal_root']
self.im_shape = params['im_shape']
self.im_mean = params['im_mean']
# get list of image indexes.
list_file = params['split'] + '.txt'
self.indexlist = [line.rstrip('\n') for line in open(
osp.join(self.pascal_root, 'ImageSets/Main', list_file))]
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
self.transformer.set_mean(self.im_mean)
print "BatchLoader initialized with {} images".format(
len(self.indexlist))
def load_next_image(self):
"""
Load the next image in a batch.
"""
# Did we finish an epoch?
if self._cur == len(self.indexlist):
self._cur = 0
shuffle(self.indexlist)
# Load an image
index = self.indexlist[self._cur] # Get the image index
image_file_name = index + '.jpg'
im = np.asarray(Image.open(
osp.join(self.pascal_root, 'JPEGImages', image_file_name)))
im = scipy.misc.imresize(im, self.im_shape) # resize
# do a simple horizontal flip as data augmentation
flip = np.random.choice(2)*2-1
im = im[:, ::flip, :]
# Load and prepare ground truth
multilabel = np.zeros(20).astype(np.float32)
anns = load_pascal_annotation(index, self.pascal_root)
for label in anns['gt_classes']:
# in the multilabel problem we don't care how MANY instances
# there are of each class. Only if they are present.
# The "-1" is b/c we are not interested in the background
# class.
multilabel[label - 1] = 1
self._cur += 1
return self.transformer.preprocess(im), multilabel
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
# get list of image indexs
self.indexlist = [
line.rstrip('\n') for line in open(params['data']).readlines()
]
self._cur = 0 # current image
# this class does some simple data-manipulation
self.transformer = SimpleTransformer()
print "BatchLoader initialized with {} images".format(
len(self.indexlist))
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.yelp_picture_root = params['yelp_picture_root']
self.yelp_csv_root = params['yelp_csv_root']
self.im_shape = params['im_shape']
self.split = params['split']
# get list of image indexes.
if self.split in ["train", "validation"]:
list_csv = self.yelp_csv_root + self.split + '_photo_to_biz_ids2.csv'
else:
list_csv = self.yelp_csv_root + self.split + '_photo_to_biz.csv'
self.image_key = []
with open(list_csv) as csv_file:
reader = csv.DictReader(csv_file)
for row in reader:
self.image_key.append(row)
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
print "BatchLoader initialized with {} images".format(
len(self.image_key))
if self.split in ["train", "validation", "poster"]:
attributes_csv = osp.join(self.yelp_csv_root, self.split + '2.csv')
self.attributes_dict = {}
if self.split == "poster":
attributes_csv = osp.join(self.yelp_picture_root,
self.split + '.csv')
with open(attributes_csv) as csv_file:
reader = csv.DictReader(csv_file)
for row in reader:
attr_string = row["labels"]
self.attributes_dict[row["business_id"]] = [
int(label) for label in row["labels"].split()
]
if self.split == 'validation':
lmdb_dir = self.yelp_picture_root + 'val_lmdb'
else:
lmdb_dir = self.yelp_picture_root + self.split + '_lmdb'
lmdb_env = lmdb.open(lmdb_dir)
self.lmdb_txn = lmdb_env.begin()
self.lmdb_cursor = self.lmdb_txn.cursor()
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.pascal_root = params['pascal_root']
self.im_shape = params['im_shape']
# get list of image indexes.
list_file = params['split'] + '.txt'
self.indexlist = [line.rstrip('\n') for line in open(
osp.join(self.pascal_root, 'ImageSets/Main', list_file))]
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
print "BatchLoader initialized with {} images".format(
len(self.indexlist))
def createPatch(x, y, label):
print x, y
patch = im[x - half_image_size:x + half_image_size,
y - half_image_size:y + half_image_size]
print np.mean(patch[:, :, 0]), np.mean(patch[:, :, 1]), np.mean(patch[:, :,
2])
st = SimpleTransformer()
patch = st.preprocess(patch) # no mean
savePatch(patch, x, y, label)
print np.mean(patch[:, :, 0]), np.mean(patch[:, :, 1]), np.mean(patch[:, :,
2])
bgrMean(np.mean(patch[:, :, 0]), np.mean(patch[:, :, 1]),
np.mean(patch[:, :, 2]))
return patch.tobytes()
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.dataRoot = params['dataRoot']
self.im_shape = params['im_shape']
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
# load cifar10 images into memory
# 32 x 32 x 3 channel images
imageLen = self.im_shape[0] * self.im_shape[1] * 3
self.data = np.empty(shape=[0,imageLen])
self.labelDense = []
for i in range(1,6):
data = self.LoadCifar10(self.dataRoot + '/data_batch_%d' % i)
self.data = np.concatenate((self.data, data['data']), axis=0)
self.labelDense += data['labels']
# self.label = np.concatenate((self.label, npLabel), axis=0)
self.dataCount = self.data.shape[0]
# self.labelOneHot = np.zeros(shape=[self.dataCount, 10])
self.labelOneHot = self.DenseToOneHot(np.array(self.labelDense), 10)
# print self.labelOneHot
self.indexList = range(self.dataCount) # shuffle()
def __init__(self, params, lexicon, channel):
self.im_shape = params['im_shape']
self.data_folder = params['data_folder']
self.lexicon = lexicon
self.channel = channel
self.indexlist = [
line.rstrip('\n').split(",") for line in open(
osp.join(self.data_folder, params['split'] +
'.txt'), 'r').readlines()
]
self._cur = 0
self.transformer = SimpleTransformer(channel)
print "BatchLoader initialized with {} images".format(
len(self.indexlist))
class BatchLoader(object):
"""
This class abstracts away the loading of images.
Images can either be loaded singly, or in a batch. The latter is used for
the asyncronous data layer to preload batches while other processing is
performed.
"""
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.pascal_root = params['pascal_root']
self.im_shape = params['im_shape']
# get list of image indexes.
list_file = params['split'] + '.txt'
self.indexlist = [line.rstrip('\n') for line in open(
osp.join(self.pascal_root, 'ImageSets/Main', list_file))]
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
print "BatchLoader initialized with {} images".format(
len(self.indexlist))
def load_next_image(self):
"""
Load the next image in a batch.
"""
# Did we finish an epoch?
if self._cur == len(self.indexlist):
self._cur = 0
shuffle(self.indexlist)
# Load an image
index = self.indexlist[self._cur] # Get the image index
image_file_name = index + '.jpg'
im = np.asarray(Image.open(
osp.join(self.pascal_root, 'JPEGImages', image_file_name)))
im = scipy.misc.imresize(im, self.im_shape) # resize
# do a simple horizontal flip as data augmentation
flip = np.random.choice(2)*2-1
im = im[:, ::flip, :]
# Load and prepare ground truth
multilabel = np.zeros(20).astype(np.float32)
anns = load_pascal_annotation(index, self.pascal_root)
for label in anns['gt_classes']:
# in the multilabel problem we don't care how MANY instances
# there are of each class. Only if they are present.
# The "-1" is b/c we are not interested in the background
# class.
multilabel[label - 1] = 1
self._cur += 1
return self.transformer.preprocess(im), multilabel
class BatchLoader(object):
"""
This class abstracts away the loading of images.
"""
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.root_path = params['root_path']
# get list of image indexs
self.indexlist = [
line.rstrip('\n') for line in open(params['data']).readlines()
]
self._cur = 0 # current image
# this class does some simple data-manipulation
self.transformer = SimpleTransformer()
print "BatchLoader initialized with {} images".format(
len(self.indexlist))
def load_next_image(self):
"""
Load the next image in a batch
"""
# Did we finish an epoch?
if self._cur == len(self.indexlist):
self._cur = 0
shuffle(self.indexlist)
# Load images
index_line = self.indexlist[self._cur] # Get the image index
indexs = index_line.split()
images = []
#for index in indexs[:-1]:
im = np.asarray(Image.open(self.root_path + indexs[0]))
# do a simple horizontal flip as data augmentation
flip = np.random.choice(2) * 2 - 1
im = im[:, ::flip, :]
images.append(self.transformer.preprocess(im))
# Load and prepare ground truth
#label = np.zeros(self.gt_classes).astype(np.float32)
label = int(indexs[-1]) - 1
self._cur += 1
return images, label
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.pascal_root = params['pascal_root']
self.im_shape = params['im_shape']
# get list of image indexes.
list_file = params['split'] + '.txt'
self.indexlist = [line.rstrip('\n') for line in open(
osp.join(self.pascal_root, 'ImageSets/Main', list_file))]
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
print "BatchLoader initialized with {} images".format(
len(self.indexlist))
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.viper_root = params[
'viper_root'] #'/home/zf/caffe/multible_channel_identification/data/VIPeR'
self.im_shape_h = params['im_shape'][0]
self.im_shape_w = params['im_shape'][1]
# get list of image indexes.
list_file_a = params['split'][0] + '.txt' #split:cam_a或者cam_b
list_file_b = params['split'][1] + '.txt'
self.indexlist_a = [
line.rstrip('\n\r')
for line in open(osp.join(params['viper_root'][0], list_file_a))
] #0000到0631
self.indexlist_b = [
line.rstrip('\n\r')
for line in open(osp.join(params['viper_root'][1], list_file_b))
]
self.i_a = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
print "BatchLoader initialized with {} images".format(
len(self.indexlist_a)) #632
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.yelp_picture_root = params['yelp_picture_root']
self.yelp_csv_root = params['yelp_csv_root']
self.im_shape = params['im_shape']
self.split=params['split']
# get list of image indexes.
if self.split in ["train", "validation"]:
list_csv = self.yelp_csv_root + self.split + '_photo_to_biz_ids2.csv'
else:
list_csv = self.yelp_csv_root + self.split + '_photo_to_biz.csv'
self.image_key = []
with open(list_csv) as csv_file:
reader = csv.DictReader(csv_file)
for row in reader:
self.image_key.append(row)
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
print "BatchLoader initialized with {} images".format(
len(self.image_key))
if self.split in ["train", "validation", "poster"]:
attributes_csv = osp.join(self.yelp_csv_root, self.split + '2.csv')
self.attributes_dict = {}
if self.split == "poster":
attributes_csv = osp.join(self.yelp_picture_root, self.split + '.csv')
with open(attributes_csv) as csv_file:
reader = csv.DictReader(csv_file)
for row in reader:
attr_string = row["labels"]
self.attributes_dict[row["business_id"]] = [int(label) for label in row["labels"].split()]
if self.split == 'validation':
lmdb_dir = self.yelp_picture_root + 'val_lmdb'
else:
lmdb_dir = self.yelp_picture_root + self.split + '_lmdb'
lmdb_env = lmdb.open(lmdb_dir)
self.lmdb_txn = lmdb_env.begin()
self.lmdb_cursor = self.lmdb_txn.cursor()
class BatchLoader(object):
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.data_root = params['data_root']
self.im_shape = params['im_shape']
self.split = params['split']
self.mean_file = params['mean_file']
list_file = '../prepare_data/AFEW-VA/crop/{}_data.txt'.format(
self.split)
self.indexlist = [line.split(' ')[0] for line in open(list_file)]
# get list of image indexes.
# Read the mat file and assign to X
mat_contents = scipy.io.loadmat(
'../prepare_data/AFEW-VA/crop/{}_labels.mat'.format(self.split))
self.X = np.zeros(mat_contents['{}_labels'.format(self.split)].shape,
dtype=np.float16)
self.X[:, :] = mat_contents['{}_labels'.format(self.split)]
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer(self.mean_file)
print("BatchLoader initialized with {} images".format(
len(self.indexlist)))
def load_next_image(self):
if self._cur == len(self.indexlist):
self._cur = 0
# shuffle(self.indexlist)
# Load an image
image_file_name = self.indexlist[self._cur] # Get the image index
im = np.asarray(Image.open(osp.join(self.data_root, image_file_name)))
im = scipy.misc.imresize(im, self.im_shape) # resize
# do a simple horizontal flip as data augmentation
flip = np.random.choice(2) * 2 - 1
im = im[:, ::flip, :]
multilabel = np.asarray(self.X[self._cur], np.float16)
self._cur += 1
return self.transformer.preprocess(im), multilabel
class BatchLoader(object):
def __init__(self, params, lexicon, channel):
self.im_shape = params['im_shape']
self.data_folder = params['data_folder']
self.lexicon = lexicon
self.channel = channel
self.indexlist = [
line.rstrip('\n').split(",") for line in open(
osp.join(self.data_folder, params['split'] +
'.txt'), 'r').readlines()
]
self._cur = 0
self.transformer = SimpleTransformer(channel)
print "BatchLoader initialized with {} images".format(
len(self.indexlist))
def load_next_image(self):
if self._cur == len(self.indexlist):
self._cur = 0
shuffle(self.indexlist)
img_annotations = [
self.indexlist[self._cur][i]
for i in range(1, len(self.indexlist[self._cur]))
]
img_path = self.indexlist[self._cur][0]
im = scipy.misc.imread(img_path, flatten=(self.channel == 1))
im = scipy.misc.imresize(im, self.im_shape)
try:
multilabel = np.zeros(self.lexicon.getNumLabel())
for a in img_annotations:
multilabel[int(a)] = 1.0
except:
print "load error!!", self.indexlist[self._cur]
self._cur += 1
return self.transformer.preprocess(im), multilabel
def __init__(self, params):
self._image_dir = params['image_dir']
self._gt_dir = params['gt_dir']
self._file_list = params['file_list']
self._image_ext = params['image_ext']
self._nclasses = params['nclasses']
assert self._nclasses >= 1
self._batch_size = params['batch_size']
assert self._batch_size >= 1
self._bgr_mean = np.array(params['bgr_mean'])
# If _im_size is none then we assume that the batch size is 1
if 'im_size' in params:
self._im_size = np.array(params['im_size'])
else:
self._im_size = None
self._batch_size = 1
if 'mirror' in params:
self._mirror = params['mirror'] == 1
else:
self._mirror = False
if 'random_seed' in params:
rd.seed(params['random_seed'])
# Read the file list
fid = open(self._file_list, 'r')
self._list = [f.strip() for f in fid]
rd.shuffle(self._list)
self._transformer = SimpleTransformer(mean=self._bgr_mean)
self._idx = 0
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.data_root = params['data_root']
self.im_shape = params['im_shape']
self.n_classes = params['n_classes']
# get list of image indexes.
list_file = params['split'] + '.json'
# indexlist is a list of all the photo_ids of the Yelp dataset
self.indexlist = [d['photo_id'] for d in json.load(open(osp.join(self.data_root, list_file)))]
# self.indexlist = [line.rstrip('\n') for line in
# open(osp.join(self.data_root, 'ImageSets/Main', list_file))]
self._cur = 0 # current image
# Get the corresponding attributes for the image file
# self.attrs = self.load_yelp_attributes(self.data_root, self.n_classes)
self.load_yelp_attributes(self.data_root, self.n_classes)
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
print "BatchLoader initialized with {} images".format(
len(self.indexlist))
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.anno_list = params['anno_list']
self.im_shape = params['im_shape']
self.attribute_list_path = params['attribute_list_path']
self.memimages = bool(params.get('memimages', 0))
self.img_transform = params.get('img_transform', 'resize')
self.ynorm = bool(params.get('ynorm', 0)) # y := y / ||y||_1
# this class does some simple data-manipulations
self.transformer = SimpleTransformer(mean=[104, 117, 123])
self.dictlist = [] # Annotation objects (image + labels) will be stored as dict here
# Load Attributes ----------------------------------------------------------------------------------------------
self.attr_id_to_name, self.attr_id_to_idx = load_attributes(self.attribute_list_path)
self.idx_to_attr_id = {v: k for k, v in self.attr_id_to_idx.iteritems()}
self.attr_id_list = self.attr_id_to_idx.keys()
self.n_attr = len(self.attr_id_list)
# Load Data ----------------------------------------------------------------------------------------------------
# Store the list of annotation files as indexlist
self.indexlist = [osp.join(DS_ROOT, line.rstrip('\n')) for line in open(self.anno_list)]
if self.memimages:
print "Loading images into memory"
print "Loading {} annotations".format(len(self.indexlist))
# Store each image-label object as a dict
# But, do not store the images. Only store the image file path
self.dictlist = [json.load(open(filename)) for filename in self.indexlist]
shuffle(self.dictlist)
self._cur = 0 # current image
# Add additional information to each dict
for idx, this_anno in enumerate(self.dictlist):
# Prepare the multilabel
# Get the list of attributes this corresponds to
attr_set = set(this_anno['labels'])
multilabel = labels_to_vec(attr_set, self.attr_id_to_idx)
if self.ynorm and np.sum(multilabel) > 0:
multilabel /= np.sum(multilabel)
assert np.sum(multilabel) > 0, 'Failed: np.sum(multilabel) > 0'
this_anno['label_vec'] = multilabel
this_anno['image_path'] = osp.join(DS_ROOT, this_anno['image_path'])
# Images can sometimes be huge (>5mb), which makes loading data extremely slow
# So, resize and stash them to enable quick loading
image_resized_path = this_anno['image_path'].replace('/images/', '/images_250/')
if os.path.exists(image_resized_path):
this_anno['image_path'] = image_resized_path
# To make training even faster, load the images into memory before it begins
if self.memimages:
im = imread(this_anno['image_path'])
if len(im.shape) == 2:
# This is a grayscale image
im = np.asarray(Image.open(this_anno['image_path']).convert('RGB'))
elif len(im.shape) == 3 and im.shape[2] == 4:
# CMYK Image
im = np.asarray(Image.open(this_anno['image_path']).convert('RGB'))
if self.img_transform == 'resize':
# Resize the image to the required shape
im = scipy.misc.imresize(im, self.im_shape)
this_anno['im'] = im
if idx % 100 == 0:
sys.stdout.write("processing %d/%d (%.2f%% done) \r" % (
idx, len(self.dictlist), idx * 100.0 / len(self.dictlist)))
sys.stdout.flush()
print "BatchLoader initialized with {} images".format(len(self.indexlist))
class BatchLoader(object):
"""
This class abstracts away the loading of images.
Images can either be loaded singly, or in a batch. The latter is used for
the asyncronous data layer to preload batches while other processing is
performed.
"""
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.anno_list = params['anno_list']
self.im_shape = params['im_shape']
self.attribute_list_path = params['attribute_list_path']
self.memimages = bool(params.get('memimages', 0))
self.img_transform = params.get('img_transform', 'resize')
self.ynorm = bool(params.get('ynorm', 0)) # y := y / ||y||_1
# this class does some simple data-manipulations
self.transformer = SimpleTransformer(mean=[104, 117, 123])
self.dictlist = [] # Annotation objects (image + labels) will be stored as dict here
# Load Attributes ----------------------------------------------------------------------------------------------
self.attr_id_to_name, self.attr_id_to_idx = load_attributes(self.attribute_list_path)
self.idx_to_attr_id = {v: k for k, v in self.attr_id_to_idx.iteritems()}
self.attr_id_list = self.attr_id_to_idx.keys()
self.n_attr = len(self.attr_id_list)
# Load Data ----------------------------------------------------------------------------------------------------
# Store the list of annotation files as indexlist
self.indexlist = [osp.join(DS_ROOT, line.rstrip('\n')) for line in open(self.anno_list)]
if self.memimages:
print "Loading images into memory"
print "Loading {} annotations".format(len(self.indexlist))
# Store each image-label object as a dict
# But, do not store the images. Only store the image file path
self.dictlist = [json.load(open(filename)) for filename in self.indexlist]
shuffle(self.dictlist)
self._cur = 0 # current image
# Add additional information to each dict
for idx, this_anno in enumerate(self.dictlist):
# Prepare the multilabel
# Get the list of attributes this corresponds to
attr_set = set(this_anno['labels'])
multilabel = labels_to_vec(attr_set, self.attr_id_to_idx)
if self.ynorm and np.sum(multilabel) > 0:
multilabel /= np.sum(multilabel)
assert np.sum(multilabel) > 0, 'Failed: np.sum(multilabel) > 0'
this_anno['label_vec'] = multilabel
this_anno['image_path'] = osp.join(DS_ROOT, this_anno['image_path'])
# Images can sometimes be huge (>5mb), which makes loading data extremely slow
# So, resize and stash them to enable quick loading
image_resized_path = this_anno['image_path'].replace('/images/', '/images_250/')
if os.path.exists(image_resized_path):
this_anno['image_path'] = image_resized_path
# To make training even faster, load the images into memory before it begins
if self.memimages:
im = imread(this_anno['image_path'])
if len(im.shape) == 2:
# This is a grayscale image
im = np.asarray(Image.open(this_anno['image_path']).convert('RGB'))
elif len(im.shape) == 3 and im.shape[2] == 4:
# CMYK Image
im = np.asarray(Image.open(this_anno['image_path']).convert('RGB'))
if self.img_transform == 'resize':
# Resize the image to the required shape
im = scipy.misc.imresize(im, self.im_shape)
this_anno['im'] = im
if idx % 100 == 0:
sys.stdout.write("processing %d/%d (%.2f%% done) \r" % (
idx, len(self.dictlist), idx * 100.0 / len(self.dictlist)))
sys.stdout.flush()
print "BatchLoader initialized with {} images".format(len(self.indexlist))
def load_next_image(self):
"""
Load the next image in a batch.
"""
# Sample image -------------------------------------------------------------------------------------------------
# Did we finish an epoch?
if self._cur == len(self.dictlist):
self._cur = 0
next_idx = self._cur
shuffle(self.dictlist)
else:
next_idx = self._cur
self._cur += 1
dct = self.dictlist[next_idx] # Get the anno
# Load image ---------------------------------------------------------------------------------------------------
image_path = dct['image_path']
multilabel = dct['label_vec']
# Load an image
if 'im' in dct: # Images can be preloaded before training with flag memimages
im = dct['im']
else:
im = imread(image_path)
if len(im.shape) == 2:
# This is a grayscale image
im = np.asarray(Image.open(image_path).convert('RGB'))
elif len(im.shape) == 3 and im.shape[2] == 4:
# CMYK Image
im = np.asarray(Image.open(image_path).convert('RGB'))
org_shape = im.shape
# Resize/Transform image ---------------------------------------------------------------------------------------
if self.img_transform == 'resize':
# Resize the image to the required shape
im = scipy.misc.imresize(im, self.im_shape)
elif self.img_transform == 'rand_crop':
# Take a random crop of size self.im_shape
# im.shape = [H, W, 3]
img_h, img_w, _ = im.shape
crop_h, crop_w = self.im_shape
if img_w < crop_w:
new_w = crop_w
new_h = int(np.round(img_h * (new_w / float(img_w)))) # Scale height to same aspect ratio
im = scipy.misc.imresize(im, (new_h, new_w))
img_w, img_h = new_w, new_h
# print 'New (w, h): ', (img_w, img_h)
if img_h < crop_h:
new_h = crop_h
new_w = int(np.round(img_w * (new_h / float(img_h))))
im = scipy.misc.imresize(im, (new_h, new_w))
img_w, img_h = new_w, new_h
# Sample (x1, y1) i.e, top-left point of the image
x1 = np.random.randint(low=0, high=(img_h - crop_h - 1))
y1 = np.random.randint(low=0, high=(img_w - crop_w - 1))
# Crop a window given this point
x2 = x1 + crop_h
y2 = y1 + crop_w
im = im[x1:x2, y1:y2, :]
# do a simple horizontal flip as data augmentation
flip = np.random.choice(2)*2-1
im = im[:, ::flip, :]
transformed_im = self.transformer.preprocess(im)
return transformed_im, multilabel
class BatchLoader(object):
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.yelp_picture_root = params['yelp_picture_root']
self.yelp_csv_root = params['yelp_csv_root']
self.im_shape = params['im_shape']
self.split = params['split']
# get list of image indexes.
if self.split in ["train", "validation"]:
list_csv = self.yelp_csv_root + self.split + '_photo_to_biz_ids2.csv'
else:
list_csv = self.yelp_csv_root + self.split + '_photo_to_biz.csv'
self.image_key = []
with open(list_csv) as csv_file:
reader = csv.DictReader(csv_file)
for row in reader:
self.image_key.append(row)
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
print "BatchLoader initialized with {} images".format(
len(self.image_key))
if self.split in ["train", "validation", "poster"]:
attributes_csv = osp.join(self.yelp_csv_root, self.split + '2.csv')
self.attributes_dict = {}
if self.split == "poster":
attributes_csv = osp.join(self.yelp_picture_root,
self.split + '.csv')
with open(attributes_csv) as csv_file:
reader = csv.DictReader(csv_file)
for row in reader:
attr_string = row["labels"]
self.attributes_dict[row["business_id"]] = [
int(label) for label in row["labels"].split()
]
if self.split == 'validation':
lmdb_dir = self.yelp_picture_root + 'val_lmdb'
else:
lmdb_dir = self.yelp_picture_root + self.split + '_lmdb'
lmdb_env = lmdb.open(lmdb_dir)
self.lmdb_txn = lmdb_env.begin()
self.lmdb_cursor = self.lmdb_txn.cursor()
def load_next_image(self):
"""
Load the next image in a batch.
"""
# Did we finish an epoch?
if self._cur == len(self.image_key):
self._cur = 0
if not self.lmdb_cursor.next():
self.lmdb_cursor = self.lmdb_txn.cursor()
self.lmdb_cursor.next()
business_id = self.image_key[self._cur]["business_id"]
photo_id = self.image_key[self._cur]["photo_id"] # Get the image index
# Load and prepare ground truth
multilabel = np.zeros(9).astype(np.float32)
if self.split in ["train", "validation", "poster"]:
anns = self.load_yelp_attributes(business_id)
for label in anns:
# convert label information to a 1/0 array.
multilabel[label] = 1
self._cur += 1
datum = caffe.proto.caffe_pb2.Datum()
key, value = self.lmdb_cursor.item()
datum.ParseFromString(value)
label = datum.label
# TODO(prad): Add a check for test as well.
if self.split != 'test' and self.split != 'poster' and str(
label) != business_id:
print "Houston, we have a problem." + str(label) + ":" + str(
business_id)
data = caffe.io.datum_to_array(datum)
im = scipy.misc.imresize(data, self.im_shape) # resize
# do a simple horizontal flip as data augmentation
flip = np.random.choice(2) * 2 - 1
im = im[:, ::flip, :]
# The datum above already does BGR transformation. But we are redoing it using the transformer below.
im = im[:, :, ::-1]
return self.transformer.preprocess(im), multilabel, photo_id, label
def load_yelp_attributes(self, business_id, mapping=None):
return self.attributes_dict[business_id]
class BatchLoader(object):
"""
This class abstracts away the loading of images.
Images can either be loaded singly, or in a batch. The latter is used for
the asyncronous data layer to preload batches while other processing is
performed.
"""
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.dataRoot = params['dataRoot']
self.im_shape = params['im_shape']
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
# load cifar10 images into memory
# 32 x 32 x 3 channel images
imageLen = self.im_shape[0] * self.im_shape[1] * 3
self.data = np.empty(shape=[0,imageLen])
self.labelDense = []
for i in range(1,6):
data = self.LoadCifar10(self.dataRoot + '/data_batch_%d' % i)
self.data = np.concatenate((self.data, data['data']), axis=0)
self.labelDense += data['labels']
# self.label = np.concatenate((self.label, npLabel), axis=0)
self.dataCount = self.data.shape[0]
# self.labelOneHot = np.zeros(shape=[self.dataCount, 10])
self.labelOneHot = self.DenseToOneHot(np.array(self.labelDense), 10)
# print self.labelOneHot
self.indexList = range(self.dataCount) # shuffle()
# print 'Batchloader initialized with %d images' % (self.dataCount)
# print self.indexList
# label_dense : m x 1
# output : [m x num_classes] one hot
def DenseToOneHot(self, labels_dense, num_classes):
"""Convert class labels from scalars to one-hot vectors."""
labelOneHot = np.zeros((labels_dense.shape[0], num_classes))
labelOneHot[np.arange(labels_dense.shape[0]), np.int8(labels_dense)] = 1
return np.int8(labelOneHot)
def LoadCifar10(self, file):
import cPickle
with open(file, 'rb') as fo:
dict = cPickle.load(fo)
return dict
def load_next_image(self):
"""
Load the next image in a batch.
"""
# Did we finish an epoch?
if self._cur == self.dataCount:
self._cur = 0
shuffle(self.indexList)
index = self.indexList[self._cur] # Get the image index
im = self.data[index,:].reshape(self.im_shape[2],self.im_shape[0],self.im_shape[1]).transpose([1,2,0])
# plt.ioff()
# plt.imshow(im)
# plt.title(self.labelDense[index])
# plt.show()
# im = scipy.misc.imresize(im, self.im_shape) # resize
# do a simple horizontal flip as data augmentation
# flip = np.random.choice(2)*2-1
# im = im[:, ::flip, :]
# Load and prepare ground truth
# multilabel = np.zeros(20).astype(np.float32)
self._cur += 1
# return self.transformer.preprocess(im), self.labelOneHot[index,:]
return self.transformer.preprocess(im), self.labelDense[index]
class BatchLoader(object):
"""
This class abstracts away the loading of images.
Images can either be loaded singly, or in a batch. The latter is used for
the asyncronous data layer to preload batches while other processing is
performed.
"""
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.anno_list = params['anno_list']
self.im_shape = params['im_shape']
self.label_shortlist_path = params['label_shortlist']
self.num_labels = params.get('nlabels', NUM_LABELS)
self.memimages = bool(params.get('memimages', 0))
self.img_transform = params.get('img_transform', 'resize')
self.ynorm = bool(params.get('ynorm', 0)) # y := y / ||y||_1
self.wloss = bool(params.get('wloss', 0))
self.user_prefs = params.get('user_prefs', None)
self.scale_user_pref = bool(params.get('scale_user_pref', 0))
# Possible options:
# 'uniform' (Default) : Sample uniformly
# 'weighted': Sample a weight uniformly (usually between 1-5). Then sample an example from on of these.
self.sampling = params.get('sampling', 'uniform')
if self.label_shortlist_path is not None:
self.attr_id_to_idx = dict()
self.attr_id_to_weight = dict()
with open(self.label_shortlist_path, 'r') as f:
f.readline() # Skip header line
for line in f:
idx, attr_id, count, weight = line.strip().split('\t')
idx = int(idx)
count = int(count)
weight = float(weight)
self.attr_id_to_idx[attr_id] = idx
self.attr_id_to_weight[attr_id] = weight
else:
assert False, "Not Supported"
# self.attr_id_to_idx = load_attr_to_idx()
self.attr_id_list = self.attr_id_to_idx.keys()
self.n_attr = len(self.attr_id_list)
self.user_mat = None
if self.user_prefs is not None:
'''
This is a file of the format:
<attribute_id> <attribute_name> <score_1> <score_2> .... <score_U>
where U = # of users
score_U indicates how important this attribute is to him/her
'''
with open(self.user_prefs) as uf:
uf.readline() # Skip header line
pref_dct = dict(
) # Store mapping: attr_id = [..., score_i, ...]
for line in uf:
if line.strip() == '':
continue
tokens = line.strip().split('\t')
attr_id = tokens[0]
attr_name = tokens[1]
scores = [float(s) for s in tokens[2:]]
if attr_id in self.attr_id_to_idx:
pref_dct[attr_id] = scores
# Check n_users is consistent
n_users = len(pref_dct.values()[0])
assert all([n_users == len(x) for x in pref_dct.values()
]), Counter([len(x) for x in pref_dct.values()])
# Manually fill-in safe
pref_dct[SAFE_ATTR_ID] = np.ones(n_users) * SAFE_WEIGHT
# Make sure we have preferences for all attributes that we need
assert all([
pref_attr_id in self.attr_id_to_idx
for pref_attr_id in pref_dct.keys()
])
# Represent as a matrix: A x U
# Where col_j represents attribute preferences for user j
n_attr = len(self.attr_id_to_idx)
self.user_mat = np.zeros((n_attr, n_users))
for attr_id, idx in self.attr_id_to_idx.iteritems():
attr_scores = pref_dct[attr_id]
self.user_mat[idx] = attr_scores
print 'User preferences: '
print self.user_mat
print 'user_mat.shape = ', self.user_mat.shape
# Normalize user_mat
if self.scale_user_pref:
self.user_mat -= 2.5 # Assuming mean of scores = 2.5, so scale to [-2.5, 2.5]
self.user_mat /= 2.5 # Scale to [-1, 1]
# Store the list of annotation files as indexlist
self.indexlist = [line.rstrip('\n') for line in open(self.anno_list)]
if self.memimages:
print "Loading images into memory"
print "Loading {} annotations".format(len(self.indexlist))
# Store each image-label object as a dict
# But, do not store the images. Only store the image file path
self.dictlist = [json.load(open(aidx)) for aidx in self.indexlist]
shuffle(self.dictlist)
# Create a weight vector
self.idx_to_attr_id = {
v: k
for k, v in self.attr_id_to_idx.iteritems()
}
self.idx_to_weight = np.ones(68)
if self.wloss:
for idx in sorted(self.idx_to_attr_id.keys()):
attr_id = self.idx_to_attr_id[idx]
self.idx_to_weight[idx] = self.attr_id_to_weight[attr_id]
print 'Class weights: '
print self.idx_to_weight
if self.sampling == 'weighted':
'''
1. Create a mapping of WEIGHT (int) -> attribute_idx
weight of example = max(weight of attribute i in example)
2. When sampling next image:
a. Sample weight ~ [1, 2, 3, 4, 5]
b. Sample an example corresponding to this weight
Maintain a dict:
{
1: [3, 10, 4, ...],
2: [45, 11, 90, ...],
...
}
and pop an idx from the list when asked for next image
'''
self.weight_to_idx_list = get_w2idx(self.dictlist,
self.attr_id_to_weight)
# Maintain a copy of this, because it will mutate in each iteration (pop() to consume)
self.org_weight_to_idx_list = copy.deepcopy(
self.weight_to_idx_list)
elif self.sampling == 'class_weighted':
'''
1. Create a mapping of LABEL (attr_id) -> DICT_IDX
2. When sampling next image:
a. Sample class ~ [attr_1, attr_2, ..., attr_L]
b. Sample an example corresponding to this label
Maintain a dict:
{
attr_1: [3, 10, 4, ...],
attr_2: [45, 11, 90, ...],
...
}
and pop an idx from the list when asked for next image
'''
self.class_to_idx_list = get_class2idx(self.dictlist)
# Maintain a copy of this, because it will mutate in each iteration (pop() to consume)
self.org_class_to_idx_list = copy.deepcopy(self.class_to_idx_list)
else:
self._cur = 0 # current image
self.weight_to_idx_list = None
self.class_to_idx_list = None
# Add to each dict the label vector
for idx, this_anno in enumerate(self.dictlist):
# Prepare the multilabel
# Get the list of attributes this corresponds to
if 'labels' in this_anno:
attr_set = set(this_anno['labels'])
else:
this_attr_list = []
for categ_id, attr_id_list in this_anno[
'attributes'].iteritems():
this_attr_list += attr_id_list
attr_set = set(this_attr_list)
multilabel = attribute_set_to_vec(self.attr_id_to_idx,
attr_set,
is_safe=this_anno['safe'])
if self.ynorm and np.sum(multilabel) > 0:
multilabel /= np.sum(multilabel)
assert np.sum(multilabel) > 0, 'Failed: np.sum(multilabel) > 0'
this_anno['label_vec'] = multilabel
image_path = this_anno['image_path']
image_resized_path = image_path.replace('/images_chunks/',
'/images_chunks_resized/')
if os.path.exists(image_resized_path):
this_anno['image_path'] = image_resized_path
if self.memimages:
im = imread(this_anno['image_path'])
if len(im.shape) == 2:
# This is a grayscale image
im = np.asarray(
Image.open(this_anno['image_path']).convert('RGB'))
elif len(im.shape) == 3 and im.shape[2] == 4:
# CMYK Image
im = np.asarray(
Image.open(this_anno['image_path']).convert('RGB'))
if self.img_transform == 'resize':
# Resize the image to the required shape
im = scipy.misc.imresize(im, self.im_shape)
this_anno['im'] = im
if idx % 100 == 0:
sys.stdout.write("processing %d/%d (%.2f%% done) \r" %
(idx, len(self.dictlist),
idx * 100.0 / len(self.dictlist)))
sys.stdout.flush()
print 'multilabel.shape = ', multilabel.shape
# this class does some simple data-manipulations
self.transformer = SimpleTransformer(mean=[104, 117, 123])
print "BatchLoader initialized with {} images".format(
len(self.indexlist))
def get_weights(self):
return self.idx_to_weight
def get_user_prefs(self):
return self.user_mat.copy()
def get_attr_id_list(self):
return self.attr_id_list
def load_next_image(self):
"""
Load the next image in a batch.
"""
# Sample image -------------------------------------------------------------------------------------------------
# Choose which idx in dctlist to read
# The next block should fill this in
if self.sampling == 'weighted':
# 1. Sample a weight
this_w = np.random.choice(self.weight_to_idx_list.keys())
# 2.a. Is an image available for this weight. If not,
if len(self.weight_to_idx_list[this_w]) == 0:
# Copy from the original mapping
self.weight_to_idx_list = copy.deepcopy(
self.org_weight_to_idx_list)
# Shuffle indices
for w in sorted(self.weight_to_idx_list.keys()):
shuffle(self.weight_to_idx_list[w])
# 2.b. Get the next index
next_idx = self.weight_to_idx_list[this_w].pop()
elif self.sampling == 'class_weighted':
# 1. Sample a label
this_attr_id = np.random.choice(self.class_to_idx_list.keys())
# 2a. Is there a training example available for this weight? If not,
if len(self.class_to_idx_list[this_attr_id]) == 0:
# Copy from original mapping
self.class_to_idx_list = copy.deepcopy(
self.org_class_to_idx_list)
# Shuffle them
for ai in self.class_to_idx_list:
shuffle(self.class_to_idx_list[ai])
# 2b. Get next index
next_idx = self.class_to_idx_list[this_attr_id].pop()
else:
# Did we finish an epoch?
if self._cur == len(self.dictlist):
self._cur = 0
next_idx = self._cur
shuffle(self.dictlist)
else:
next_idx = self._cur
self._cur += 1
dct = self.dictlist[next_idx] # Get the anno
# Load image ---------------------------------------------------------------------------------------------------
image_path = dct['image_path']
multilabel = dct['label_vec']
assert multilabel.shape[
0] == self.num_labels, 'multilabel.shape[0] ({}) != self.num_labels ({})'.format(
multilabel.shape[0], self.num_labels)
# Load an image
if 'im' in dct:
im = dct['im']
else:
im = imread(image_path)
if len(im.shape) == 2:
# This is a grayscale image
im = np.asarray(Image.open(image_path).convert('RGB'))
elif len(im.shape) == 3 and im.shape[2] == 4:
# CMYK Image
im = np.asarray(Image.open(image_path).convert('RGB'))
org_shape = im.shape
# Resize/Transform image ---------------------------------------------------------------------------------------
if self.img_transform == 'resize':
# Resize the image to the required shape
im = scipy.misc.imresize(im, self.im_shape)
elif self.img_transform == 'rand_crop':
# Take a random crop of size self.im_shape
# im.shape = [H, W, 3]
img_h, img_w, _ = im.shape
crop_h, crop_w = self.im_shape
# print 'Processing file: ', image_path
# print 'Old (w, h): ', (img_w, img_h)
if img_w < crop_w:
new_w = crop_w
new_h = int(np.round(
img_h *
(new_w /
float(img_w)))) # Scale height to same aspect ratio
im = scipy.misc.imresize(im, (new_h, new_w))
img_w, img_h = new_w, new_h
# print 'New (w, h): ', (img_w, img_h)
if img_h < crop_h:
new_h = crop_h
new_w = int(np.round(img_w * (new_h / float(img_h))))
im = scipy.misc.imresize(im, (new_h, new_w))
img_w, img_h = new_w, new_h
# print 'New (w, h): ', (img_w, img_h)
# Sample (x1, y1) i.e, top-left point of the image
x1 = np.random.randint(low=0, high=(img_h - crop_h - 1))
y1 = np.random.randint(low=0, high=(img_w - crop_w - 1))
# Crop a window given this point
x2 = x1 + crop_h
y2 = y1 + crop_w
im = im[x1:x2, y1:y2, :]
# print '(x1, y1) = ', (x1, x2)
# print 'Cropped (w, h): ', (x2-x1, y2-y1)
# print 'im.shape = ', im.shape
# do a simple horizontal flip as data augmentation
flip = np.random.choice(2) * 2 - 1
im = im[:, ::flip, :]
transformed_im = self.transformer.preprocess(im)
return transformed_im, multilabel
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.anno_list = params['anno_list']
self.im_shape = params['im_shape']
self.label_shortlist_path = params['label_shortlist']
self.num_labels = params.get('nlabels', NUM_LABELS)
self.memimages = bool(params.get('memimages', 0))
self.img_transform = params.get('img_transform', 'resize')
self.ynorm = bool(params.get('ynorm', 0)) # y := y / ||y||_1
self.wloss = bool(params.get('wloss', 0))
self.user_prefs = params.get('user_prefs', None)
self.scale_user_pref = bool(params.get('scale_user_pref', 0))
# Possible options:
# 'uniform' (Default) : Sample uniformly
# 'weighted': Sample a weight uniformly (usually between 1-5). Then sample an example from on of these.
self.sampling = params.get('sampling', 'uniform')
if self.label_shortlist_path is not None:
self.attr_id_to_idx = dict()
self.attr_id_to_weight = dict()
with open(self.label_shortlist_path, 'r') as f:
f.readline() # Skip header line
for line in f:
idx, attr_id, count, weight = line.strip().split('\t')
idx = int(idx)
count = int(count)
weight = float(weight)
self.attr_id_to_idx[attr_id] = idx
self.attr_id_to_weight[attr_id] = weight
else:
assert False, "Not Supported"
# self.attr_id_to_idx = load_attr_to_idx()
self.attr_id_list = self.attr_id_to_idx.keys()
self.n_attr = len(self.attr_id_list)
self.user_mat = None
if self.user_prefs is not None:
'''
This is a file of the format:
<attribute_id> <attribute_name> <score_1> <score_2> .... <score_U>
where U = # of users
score_U indicates how important this attribute is to him/her
'''
with open(self.user_prefs) as uf:
uf.readline() # Skip header line
pref_dct = dict(
) # Store mapping: attr_id = [..., score_i, ...]
for line in uf:
if line.strip() == '':
continue
tokens = line.strip().split('\t')
attr_id = tokens[0]
attr_name = tokens[1]
scores = [float(s) for s in tokens[2:]]
if attr_id in self.attr_id_to_idx:
pref_dct[attr_id] = scores
# Check n_users is consistent
n_users = len(pref_dct.values()[0])
assert all([n_users == len(x) for x in pref_dct.values()
]), Counter([len(x) for x in pref_dct.values()])
# Manually fill-in safe
pref_dct[SAFE_ATTR_ID] = np.ones(n_users) * SAFE_WEIGHT
# Make sure we have preferences for all attributes that we need
assert all([
pref_attr_id in self.attr_id_to_idx
for pref_attr_id in pref_dct.keys()
])
# Represent as a matrix: A x U
# Where col_j represents attribute preferences for user j
n_attr = len(self.attr_id_to_idx)
self.user_mat = np.zeros((n_attr, n_users))
for attr_id, idx in self.attr_id_to_idx.iteritems():
attr_scores = pref_dct[attr_id]
self.user_mat[idx] = attr_scores
print 'User preferences: '
print self.user_mat
print 'user_mat.shape = ', self.user_mat.shape
# Normalize user_mat
if self.scale_user_pref:
self.user_mat -= 2.5 # Assuming mean of scores = 2.5, so scale to [-2.5, 2.5]
self.user_mat /= 2.5 # Scale to [-1, 1]
# Store the list of annotation files as indexlist
self.indexlist = [line.rstrip('\n') for line in open(self.anno_list)]
if self.memimages:
print "Loading images into memory"
print "Loading {} annotations".format(len(self.indexlist))
# Store each image-label object as a dict
# But, do not store the images. Only store the image file path
self.dictlist = [json.load(open(aidx)) for aidx in self.indexlist]
shuffle(self.dictlist)
# Create a weight vector
self.idx_to_attr_id = {
v: k
for k, v in self.attr_id_to_idx.iteritems()
}
self.idx_to_weight = np.ones(68)
if self.wloss:
for idx in sorted(self.idx_to_attr_id.keys()):
attr_id = self.idx_to_attr_id[idx]
self.idx_to_weight[idx] = self.attr_id_to_weight[attr_id]
print 'Class weights: '
print self.idx_to_weight
if self.sampling == 'weighted':
'''
1. Create a mapping of WEIGHT (int) -> attribute_idx
weight of example = max(weight of attribute i in example)
2. When sampling next image:
a. Sample weight ~ [1, 2, 3, 4, 5]
b. Sample an example corresponding to this weight
Maintain a dict:
{
1: [3, 10, 4, ...],
2: [45, 11, 90, ...],
...
}
and pop an idx from the list when asked for next image
'''
self.weight_to_idx_list = get_w2idx(self.dictlist,
self.attr_id_to_weight)
# Maintain a copy of this, because it will mutate in each iteration (pop() to consume)
self.org_weight_to_idx_list = copy.deepcopy(
self.weight_to_idx_list)
elif self.sampling == 'class_weighted':
'''
1. Create a mapping of LABEL (attr_id) -> DICT_IDX
2. When sampling next image:
a. Sample class ~ [attr_1, attr_2, ..., attr_L]
b. Sample an example corresponding to this label
Maintain a dict:
{
attr_1: [3, 10, 4, ...],
attr_2: [45, 11, 90, ...],
...
}
and pop an idx from the list when asked for next image
'''
self.class_to_idx_list = get_class2idx(self.dictlist)
# Maintain a copy of this, because it will mutate in each iteration (pop() to consume)
self.org_class_to_idx_list = copy.deepcopy(self.class_to_idx_list)
else:
self._cur = 0 # current image
self.weight_to_idx_list = None
self.class_to_idx_list = None
# Add to each dict the label vector
for idx, this_anno in enumerate(self.dictlist):
# Prepare the multilabel
# Get the list of attributes this corresponds to
if 'labels' in this_anno:
attr_set = set(this_anno['labels'])
else:
this_attr_list = []
for categ_id, attr_id_list in this_anno[
'attributes'].iteritems():
this_attr_list += attr_id_list
attr_set = set(this_attr_list)
multilabel = attribute_set_to_vec(self.attr_id_to_idx,
attr_set,
is_safe=this_anno['safe'])
if self.ynorm and np.sum(multilabel) > 0:
multilabel /= np.sum(multilabel)
assert np.sum(multilabel) > 0, 'Failed: np.sum(multilabel) > 0'
this_anno['label_vec'] = multilabel
image_path = this_anno['image_path']
image_resized_path = image_path.replace('/images_chunks/',
'/images_chunks_resized/')
if os.path.exists(image_resized_path):
this_anno['image_path'] = image_resized_path
if self.memimages:
im = imread(this_anno['image_path'])
if len(im.shape) == 2:
# This is a grayscale image
im = np.asarray(
Image.open(this_anno['image_path']).convert('RGB'))
elif len(im.shape) == 3 and im.shape[2] == 4:
# CMYK Image
im = np.asarray(
Image.open(this_anno['image_path']).convert('RGB'))
if self.img_transform == 'resize':
# Resize the image to the required shape
im = scipy.misc.imresize(im, self.im_shape)
this_anno['im'] = im
if idx % 100 == 0:
sys.stdout.write("processing %d/%d (%.2f%% done) \r" %
(idx, len(self.dictlist),
idx * 100.0 / len(self.dictlist)))
sys.stdout.flush()
print 'multilabel.shape = ', multilabel.shape
# this class does some simple data-manipulations
self.transformer = SimpleTransformer(mean=[104, 117, 123])
print "BatchLoader initialized with {} images".format(
len(self.indexlist))
class BatchLoader(object):
"""
This class abstracts away the loading of images.
Images can either be loaded singly, or in a batch. The latter is used for
the asyncronous data layer to preload batches while other processing is
performed.
"""
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.viper_root = params[
'viper_root'] #'/home/zf/caffe/multible_channel_identification/data/VIPeR'
self.im_shape_h = params['im_shape'][0]
self.im_shape_w = params['im_shape'][1]
# get list of image indexes.
list_file_a = params['split'][0] + '.txt' #split:cam_a或者cam_b
list_file_b = params['split'][1] + '.txt'
self.indexlist_a = [
line.rstrip('\n\r')
for line in open(osp.join(params['viper_root'][0], list_file_a))
] #0000到0631
self.indexlist_b = [
line.rstrip('\n\r')
for line in open(osp.join(params['viper_root'][1], list_file_b))
]
self.i_a = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
print "BatchLoader initialized with {} images".format(
len(self.indexlist_a)) #632
def load_next_three_images(self):
"""
Load the next image in a batch.
"""
# Did we finish an epoch?
if self.i_a == len(self.indexlist_a):
self.i_a = 0
shuffle(self.indexlist_a)
# Load three images
index_a = self.indexlist_a[self.i_a] # Get the image index of cam_a
i_b = 0
while (self.indexlist_b[i_b] != index_a):
i_b = i_b + 1
if (i_b == len(self.indexlist_a)):
print "there is some error in the data"
return
index_b = self.indexlist_b[i_b] # Get the image index of cam_b
i_c = random.randint(0, 631)
if (i_c == self.i_a):
i_c = random.randint(0, 631)
index_c = self.indexlist_a[i_c] #Get the positive
image_file_a_name = index_a + '.bmp'
image_file_b_name = index_b + '.bmp'
image_file_c_name = index_c + '.bmp'
im = np.zeros((3, 230, 80, 3))
im[0] = np.asarray(
Image.open(osp.join(self.viper_root[0], image_file_a_name)))
im[1] = np.asarray(
Image.open(osp.join(self.viper_root[1], image_file_b_name)))
im[2] = np.asarray(
Image.open(osp.join(self.viper_root[0], image_file_c_name)))
#print "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!"
pic = np.zeros((3, 3, 230, 80))
pic[0] = self.transformer.preprocess(im[0])
pic[1] = self.transformer.preprocess(im[1])
pic[2] = self.transformer.preprocess(im[2])
return pic
class BatchLoader():
"""
Class definition for loading a batch of data. Use this class with a worker
thread
"""
def __init__(self, params):
self._image_dir = params['image_dir']
self._gt_dir = params['gt_dir']
self._file_list = params['file_list']
self._image_ext = params['image_ext']
self._nclasses = params['nclasses']
assert self._nclasses >= 1
self._batch_size = params['batch_size']
assert self._batch_size >= 1
self._bgr_mean = np.array(params['bgr_mean'])
# If _im_size is none then we assume that the batch size is 1
if 'im_size' in params:
self._im_size = np.array(params['im_size'])
else:
self._im_size = None
self._batch_size = 1
if 'mirror' in params:
self._mirror = params['mirror'] == 1
else:
self._mirror = False
if 'random_seed' in params:
rd.seed(params['random_seed'])
# Read the file list
fid = open(self._file_list, 'r')
self._list = [f.strip() for f in fid]
rd.shuffle(self._list)
self._transformer = SimpleTransformer(mean=self._bgr_mean)
self._idx = 0
def load_one_data_point(self, fname):
"""
Load a single data point and preprocess it.
Parameters
----------
fname: the file for which we need to load the data
Returns
-------
im: Processed Image
gt: Processed Semantic Labeling
"""
im_name = os.path.join(self._image_dir,
'{}.{}'.format(fname, self._image_ext))
gt_name = os.path.join(self._gt_dir, '{}.txt'.format(fname))
im = skimage.io.imread(im_name)
gt = np.loadtxt(gt_name, delimiter=' ')
[h, w, _] = im.shape
assert h > self._im_size[0]
assert w > self._im_size[1]
h_start_idx = rd.randint(0, h - self._im_size[0])
h_end_idx = h_start_idx + self._im_size[0]
w_start_idx = rd.randint(0, w - self._im_size[1])
w_end_idx = w_start_idx + self._im_size[1]
final_im = im[h_start_idx:h_end_idx, w_start_idx:w_end_idx, :]
final_gt = gt[h_start_idx:h_end_idx, w_start_idx:w_end_idx]
if self._mirror and rd.randint(0, 1) == 1:
final_im = final_im[:, ::-1, :]
final_gt = final_gt[:, ::-1]
final_im = self._transformer.preprocess(final_im)
return final_im, final_gt
def get_fname(self):
"""
Randomly select next file to process.
Get the next file in the list to load in the CNN. If it finishes
the list then it randomly shuffles the list. This process ensures
that we look at each file in the list in each epoch and we randomly
select files rather than a fixed order.
Returns
-------
fname: A randomly Selected filename
"""
if self._idx >= len(self._list):
rd.shuffle(self._list)
self._idx = 0
fname = self._list[self._idx]
self._idx += 1
return fname
def next_batch(self):
"""
Get the next batch to process
Returns
-------
ims: numpy array of the images
gts: numpy array of labels
"""
if self._batch_size == 1:
fname = self.get_fname()
(ims, gts) = self.load_one_data_point(fname)
else:
ims = np.zeros((self._batch_size, 3, self._im_size[0],
self._im_size[1]))
gts = np.zeros((self._batch_size, 1, self._im_size[0],
self._im_size[1]))
for iter in range(self._batch_size):
fname = self.get_fname()
(im, gt) = self.load_one_data_point(fname)
ims[iter, :] = im
gts[iter, :] = gt
return ims, gts
class BatchLoader(object):
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.yelp_picture_root = params['yelp_picture_root']
self.yelp_csv_root = params['yelp_csv_root']
self.im_shape = params['im_shape']
self.split=params['split']
# get list of image indexes.
if self.split in ["train", "validation"]:
list_csv = self.yelp_csv_root + self.split + '_photo_to_biz_ids2.csv'
else:
list_csv = self.yelp_csv_root + self.split + '_photo_to_biz.csv'
self.image_key = []
with open(list_csv) as csv_file:
reader = csv.DictReader(csv_file)
for row in reader:
self.image_key.append(row)
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
print "BatchLoader initialized with {} images".format(
len(self.image_key))
if self.split in ["train", "validation", "poster"]:
attributes_csv = osp.join(self.yelp_csv_root, self.split + '2.csv')
self.attributes_dict = {}
if self.split == "poster":
attributes_csv = osp.join(self.yelp_picture_root, self.split + '.csv')
with open(attributes_csv) as csv_file:
reader = csv.DictReader(csv_file)
for row in reader:
attr_string = row["labels"]
self.attributes_dict[row["business_id"]] = [int(label) for label in row["labels"].split()]
if self.split == 'validation':
lmdb_dir = self.yelp_picture_root + 'val_lmdb'
else:
lmdb_dir = self.yelp_picture_root + self.split + '_lmdb'
lmdb_env = lmdb.open(lmdb_dir)
self.lmdb_txn = lmdb_env.begin()
self.lmdb_cursor = self.lmdb_txn.cursor()
def load_next_image(self):
"""
Load the next image in a batch.
"""
# Did we finish an epoch?
if self._cur == len(self.image_key):
self._cur = 0
if not self.lmdb_cursor.next():
self.lmdb_cursor = self.lmdb_txn.cursor()
self.lmdb_cursor.next()
business_id = self.image_key[self._cur]["business_id"]
photo_id = self.image_key[self._cur]["photo_id"] # Get the image index
# Load and prepare ground truth
multilabel = np.zeros(9).astype(np.float32)
if self.split in ["train", "validation","poster"]:
anns = self.load_yelp_attributes(business_id)
for label in anns:
# convert label information to a 1/0 array.
multilabel[label] = 1
self._cur += 1
datum = caffe.proto.caffe_pb2.Datum()
key, value = self.lmdb_cursor.item()
datum.ParseFromString(value)
label = datum.label
# TODO(prad): Add a check for test as well.
if self.split != 'test' and self.split != 'poster' and str(label) != business_id:
print "Houston, we have a problem." + str(label) + ":" + str(business_id)
data = caffe.io.datum_to_array(datum)
im = scipy.misc.imresize(data, self.im_shape) # resize
# do a simple horizontal flip as data augmentation
flip = np.random.choice(2)*2-1
im = im[:, ::flip, :]
# The datum above already does BGR transformation. But we are redoing it using the transformer below.
im = im[:, :, ::-1]
return self.transformer.preprocess(im), multilabel, photo_id, label
def load_yelp_attributes(self, business_id, mapping=None):
return self.attributes_dict[business_id]
class BatchLoader(object):
"""
This class abstracts away the loading of images.
Images can either be loaded singly, or in a batch. The latter is used for
the asyncronous data layer to preload batches while other processing is
performed.
"""
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.data_root = params['data_root']
self.im_shape = params['im_shape']
self.n_classes = params['n_classes']
# get list of image indexes.
list_file = params['split'] + '.json'
# indexlist is a list of all the photo_ids of the Yelp dataset
self.indexlist = [d['photo_id'] for d in json.load(open(osp.join(self.data_root, list_file)))]
# self.indexlist = [line.rstrip('\n') for line in
# open(osp.join(self.data_root, 'ImageSets/Main', list_file))]
self._cur = 0 # current image
# Get the corresponding attributes for the image file
# self.attrs = self.load_yelp_attributes(self.data_root, self.n_classes)
self.load_yelp_attributes(self.data_root, self.n_classes)
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
print "BatchLoader initialized with {} images".format(
len(self.indexlist))
def load_next_image(self):
"""
Load the next image in a batch.
"""
# Did we finish an epoch?
if self._cur == len(self.indexlist):
self._cur = 0
shuffle(self.indexlist)
#print "Loading image #{}".format(self._cur)
# Load an image
index = self.indexlist[self._cur] # Get the image index
image_file_name = index + '.jpg'
im = np.asarray(Image.open(
osp.join(self.data_root, 'images', image_file_name)))
im = scipy.misc.imresize(im, self.im_shape) # resize
# do a simple horizontal flip as data augmentation
flip = np.random.choice(2)*2-1
im = im[:, ::flip, :]
# Load and prepare ground truth
multilabel = np.zeros(self.n_classes).astype(np.float32)
# retrieve the attributes of the business to which this photo belongs
attrs = self.get_business_attributes(index)
# Always sort dict and return list of sorted key-value pairs to ensure feature vector is in always the same order
attrs_list = sorted(attrs.iteritems())
# print attrs
for i, label in enumerate(attrs):
# in the multilabel problem we don't care how MANY instances
# there are of each class. Only if they are present.
# The "-1" is b/c we are not interested in the background
# class.
multilabel[i] = attrs[label]
self._cur += 1
return self.transformer.preprocess(im), multilabel
def load_yelp_attributes(self, data_root, n_classes):
"""
This code is borrowed from Ross Girshick's FAST-RCNN code
(https://github.com/rbgirshick/fast-rcnn).
See publication for further details: (http://arxiv.org/abs/1504.08083).
Thanks Ross!
"""
classes = ("Accepts Credit Cards", "Alcohol",
"Ambience_casual", "Ambience_classy", "Ambience_divey", "Ambience_hipster",
"Ambience_intimate", "Ambience_romantic", "Ambience_touristy", "Ambience_trendy",
"Ambience_upscale", "Attire_casual", "Attire_dressy", "Attire_formal", "Caters",
"Delivery", "Dietary_Restrictions_dairy-free", "Dietary_Restrictions_gluten-free",
"Dietary_Restrictions_halal", "Dietary_Restrictions_kosher", "Dietary_Restrictions_soy-free",
"Dietary_Restrictions_vegan", "Dietary_Restrictions_vegetarian", "Drive-Thru",
"Good_For_breakfast", "Good_For_brunch", "Good_For_dessert", "Good_For_dinner",
"Good_For_latenight", "Good_For_lunch", "Good For Dancing", "Good For Groups",
"Good for Kids", "Happy Hour", "Has TV", 'Music_background_music',
'Music_dj', 'Music_jukebox', 'Music_karaoke', 'Music_live', 'Music_video',
"Noise_Level_average", "Noise_Level_loud", "Noise_Level_quiet", "Noise_Level_very_loud",
"Outdoor Seating",
"Parking_garage", "Parking_lot", "Parking_street", "Parking_valet", "Parking_validated",
"Price_Range_1", "Price_Range_2", "Price_Range_3", "Price_Range_4",
'Smoking_outdoor', 'Smoking_yes', 'Smoking_no',
"Take-out",
"Takes Reservations",
"Waiter Service",)
class_to_ind = dict(zip(classes, xrange(n_classes)))
attr_filename = osp.join(data_root, "business_attributes.json")
# print 'Loading: {}'.format(filename)
# Load the list of business attributes
self.business_attributes = json.load(open(attr_filename))
gt_classes = np.zeros((n_classes), dtype=np.int32)
self.photo_2_business = json.load( open(osp.join(data_root, "photo_id_to_business_id.json")) )
return self.business_attributes
def get_business_attributes(self, index):
"""
Get the attributes JSON for the image with name `index`
"""
# Look for the business which the photo_2_business json points the image to.
business = [b for b in self.photo_2_business if b["photo_id"] == index][0]
#print sorted(self.business_attributes[0].keys())
# Given the business from photo_2_business, use the business_id to find the attributes of the business
attrs = [a for a in self.business_attributes if a["id"] == business["business_id"]][0]
attributes = dict(attrs)
# remove the id from the list of attributes
attributes.pop("id")
return attributes
class BatchLoader(object):
"""
This class abstracts away the loading of images.
Images can either be loaded singly, or in a batch. The latter is used for
the asyncronous data layer to preload batches while other processing is
performed.
"""
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.number_classes = params['number_classes']
self.img_ext = params['img_ext']
self.dataset_root = params['dataset_root']
self.im_shape = params['im_shape']
# get list of image indexes.
list_file = params['split'] + '.csv'
self.indexlist = [
line.rstrip('\n').split(',')[0]
for line in open(osp.join(self.dataset_root, list_file))
]
self.labellist = [
line.rstrip('\n').split(',')[1]
for line in open(osp.join(self.dataset_root, list_file))
]
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
print("BatchLoader initialized with {} images".format(
len(self.indexlist)))
def load_next_image(self):
"""
Load the next image in a batch.
"""
# Did we finish an epoch?
if self._cur == len(self.indexlist):
self._cur = 0
randnum = random.randint(0, 100)
random.seed(randnum)
shuffle(self.indexlist)
random.seed(randnum)
shuffle(self.labellist)
# Load an image
index = self.indexlist[self._cur] # Get the image index
image_file_name = index + self.img_ext
im = np.asarray(
Image.open(osp.join(self.dataset_root, 'images', image_file_name)))
im = scipy.misc.imresize(im, self.im_shape) # resize
# do a simple horizontal flip as data augmentation
flip = np.random.choice(2) * 2 - 1
im = im[:, ::flip, :]
# Load and prepare ground truth
multilabel = np.zeros(self.number_classes).astype(np.float32)
for label in self.labellist[self._cur].split(' '):
multilabel[int(label)] = 1
self._cur += 1
return self.transformer.preprocess(im), multilabel