class DDPNDataProvider:
def __init__(self, data_split, batchsize=1):
print('init DataProvider for %s : %s : %s' %
(cfg.IMDB_NAME, cfg.PROJ_NAME, data_split))
self.is_ss = cfg.FEAT_TYPE == 'ss'
self.ss_box_dir = cfg.SS_BOX_DIR
self.ss_feat_dir = cfg.SS_FEAT_DIR
self.feat_type = cfg.FEAT_TYPE
if 'refcoco' in cfg.IMDB_NAME or cfg.IMDB_NAME == 'refclef':
self.is_mscoco_prefix = True
else:
self.is_mscoco_prefix = False
self.use_kld = cfg.USE_KLD
# self.mscoco_prefix = cfg.MSCOCO_PREFIX
self.rpn_topn = cfg.RPN_TOPN
if self.is_ss:
self.bottomup_feat_dim = cfg.SS_FEAT_DIM
else:
self.bottomup_feat_dim = cfg.BOTTOMUP_FEAT_DIM
self.query_maxlen = cfg.QUERY_MAXLEN
# self.data_paths = cfg.DATA_PATHS
self.image_ext = '.jpg'
data_splits = data_split.split(cfg.SPLIT_TOK)
if 'train' in data_splits:
self.mode = 'train'
else:
self.mode = 'test'
self.batchsize = batchsize
self.image_dir = cfg.IMAGE_DIR
self.feat_dir = cfg.FEAT_DIR
self.dict_dir = cfg.QUERY_DIR # osp.join(cfg.DATA_DIR, cfg.IMDB_NAME, 'query_dict')
self.anno = self.load_data(data_splits)
self.qdic = Dictionary(self.dict_dir)
self.qdic.load()
self.index = 0
self.batch_len = None
self.num_query = len(self.anno)
def __getitem__(self, index):
if self.batch_len is None:
self.n_skipped = 0
qid_list = self.get_query_ids()
if self.mode == 'train':
random.shuffle(qid_list)
self.qid_list = qid_list
self.batch_len = len(qid_list)
self.epoch_counter = 0
print('mode %s has %d data' % (self.mode, self.batch_len))
qid = self.qid_list[index]
gt_bbox = np.zeros(4)
qvec = np.zeros(self.query_maxlen)
img_feat = np.zeros((self.rpn_topn, self.bottomup_feat_dim))
bbox = np.zeros((self.rpn_topn, 4))
img_shape = np.zeros(2)
spt_feat = np.zeros((self.rpn_topn, 5))
if self.use_kld:
query_label = np.zeros(self.rpn_topn)
query_label_mask = 0
query_bbox_targets = np.zeros((self.rpn_topn, 4))
query_bbox_inside_weights = np.zeros((self.rpn_topn, 4))
query_bbox_outside_weights = np.zeros((self.rpn_topn, 4))
valid_data = 1
t_qstr = self.anno[qid]['qstr']
t_qvec = self.str2list(t_qstr, self.query_maxlen)
qvec[...] = t_qvec
try:
t_gt_bbox = self.anno[qid]['boxes']
gt_bbox[...] = t_gt_bbox[0]
t_img_feat, t_num_bbox, t_bbox, t_img_shape = self.get_topdown_feat(
self.anno[qid]['iid'])
t_img_feat = t_img_feat.transpose((1, 0))
t_img_feat = (t_img_feat / np.sqrt((t_img_feat**2).sum()))
img_feat[:t_num_bbox, :] = t_img_feat
bbox[:t_num_bbox, :] = t_bbox
# spt feat
img_shape[...] = np.array(t_img_shape)
t_spt_feat = self.get_spt_feat(t_bbox, t_img_shape)
spt_feat[:t_num_bbox, :] = t_spt_feat
# query label, mask
t_gt_bbox = np.array(self.anno[qid]['boxes'])
t_query_label, t_query_label_mask, t_query_bbox_targets, t_query_bbox_inside_weights, t_query_bbox_outside_weights = \
self.get_labels(t_bbox, t_gt_bbox)
if self.use_kld:
query_label[:t_num_bbox] = t_query_label
query_label_mask = t_query_label_mask
else:
query_label = t_query_label
query_bbox_targets[:t_num_bbox, :] = t_query_bbox_targets
query_bbox_inside_weights[:
t_num_bbox, :] = t_query_bbox_inside_weights
query_bbox_outside_weights[:
t_num_bbox, :] = t_query_bbox_outside_weights
except Exception as e:
print(e)
valid_data = 0
if not self.use_kld:
query_label = -1
query_label_mask = 0
query_bbox_inside_weights[...] = 0
query_bbox_outside_weights[...] = 0
print('data not found for iid: %s' % str(self.anno[qid]['iid']))
if self.index >= self.batch_len - 1:
self.epoch_counter += 1
qid_list = self.get_query_ids()
random.shuffle(qid_list)
self.qid_list = qid_list
print('a epoch passed')
return gt_bbox, qvec, img_feat, bbox, img_shape, spt_feat, query_label, query_label_mask, \
query_bbox_targets, query_bbox_inside_weights, query_bbox_outside_weights, valid_data, int(self.anno[qid]['iid'])
def __len__(self):
return self.num_query
def get_image_ids(self):
qid_list = self.get_query_ids()
iid_list = set()
for qid in qid_list:
iid_list.add(self.anno[qid]['iid'])
return list(iid_list)
def get_query_ids(self):
return self.anno.keys()
def get_num_query(self):
return self.num_query
def load_data(self, data_splits):
anno = {}
for data_split in data_splits:
# data_path = osp.join(cfg.DATA_DIR, cfg.IMDB_NAME, 'format_%s.pkl'%str(data_split))
data_path = cfg.ANNO_PATH % str(data_split)
t_anno = load(data_path)
anno.update(t_anno)
return anno
def get_vocabsize(self):
return self.qdic.size()
def get_iid(self, qid):
return self.anno[qid]['iid']
def get_img_path(self, iid):
if self.is_mscoco_prefix:
return os.path.join(
self.image_dir,
'COCO_train2014_' + str(iid).zfill(12) + self.image_ext)
else:
return os.path.join(self.image_dir, str(iid) + self.image_ext)
def str2list(self, qstr, query_maxlen):
q_list = qstr.split()
qvec = np.zeros(query_maxlen, dtype=np.int64)
# cvec = np.zeros(query_maxlen, dtype=np.int64)
for i, _ in enumerate(range(query_maxlen)):
# if i < query_maxlen - len(q_list):
# cvec[i] = 0
# else:
w = q_list[i - (query_maxlen - len(q_list))]
# is the word in the vocabulary?
# if self.qdic.has_token(w) is False:
# w = cfg.UNK_WORD #'<unk>'
qvec[i] = self.qdic.lookup(w)
# cvec[i] = 0 if i == query_maxlen - len(q_list) else 1
# return qvec, cvec
return qvec
def load_ss_box(self, ss_box_path):
boxes = np.loadtxt(ss_box_path)
if len(boxes) == 0:
raise Exception("boxes is None!")
boxes = boxes - 1
boxes[:, [0, 1]] = boxes[:, [1, 0]]
boxes[:, [2, 3]] = boxes[:, [3, 2]]
return boxes
def get_topdown_feat(self, iid):
try:
if self.is_ss:
img_path = self.get_img_path(iid)
im = skimage.io.imread(img_path)
img_h = im.shape[0]
img_w = im.shape[1]
feat_path = os.path.join(self.ss_feat_dir, str(iid) + '.npz')
ss_box_path = os.path.join(self.ss_box_dir, str(iid) + '.txt')
bbox = self.load_ss_box(ss_box_path)
num_bbox = bbox.shape[0]
img_feat = np.transpose(np.load(feat_path)['x'], (1, 0))
else:
if self.is_mscoco_prefix: # zfill(12) insert 0 before the str
feat_path = os.path.join(
self.feat_dir, 'COCO_train2014_' + str(iid).zfill(12) +
self.image_ext + '.npz')
else:
feat_path = os.path.join(
self.feat_dir,
str(iid) + self.image_ext + '.npz')
feat_dict = np.load(feat_path)
img_feat = feat_dict['x']
num_bbox = feat_dict['num_bbox']
bbox = feat_dict['bbox']
img_h = feat_dict['image_h']
img_w = feat_dict['image_w']
return img_feat, num_bbox, bbox, (img_h, img_w)
except Exception as e:
print(e)
raise Exception("UnkownError")
def create_batch_rpn(self, iid):
img_path = self.get_img_path(iid)
# img = cv2.imread(img_path)
img_feat, num_bbox, bbox, img_shape = self.get_topdown_feat(iid)
return num_bbox, bbox, img_path
def create_batch_recall(self, qid):
iid = self.anno[qid]['iid']
gt_bbox = self.anno[qid]['boxes']
img_path = self.get_img_path(iid)
# img = cv2.imread(img_path)
img_feat, num_bbox, bbox, img_shape = self.get_topdown_feat(iid)
return num_bbox, bbox, gt_bbox, img_path
def compute_targets(self, ex_rois, gt_rois, query_label):
"""Compute bounding-box regression targets for an image."""
assert ex_rois.shape[1] == 4
assert gt_rois.shape[1] == 4
targets = bbox_transform(ex_rois, gt_rois)
if cfg.BBOX_NORMALIZE_TARGETS_PRECOMPUTED:
# Optionally normalize targets by a precomputed mean and stdev
targets = ((targets - np.array(cfg.BBOX_NORMALIZE_MEANS)) /
np.array(cfg.BBOX_NORMALIZE_STDS))
query_bbox_target_data = np.hstack(
(query_label[:, np.newaxis], targets)).astype(np.float32,
copy=False)
return query_bbox_target_data
def get_query_bbox_regression_labels(self, query_bbox_target_data):
query_label = query_bbox_target_data[:, 0]
query_bbox_targets = np.zeros((query_label.size, 4), dtype=np.float32)
query_bbox_inside_weights = np.zeros(query_bbox_targets.shape,
dtype=np.float32)
inds = np.where(query_label > 0)[0]
if len(inds) != 0:
for ind in inds:
query_bbox_targets[ind, :] = query_bbox_target_data[ind, 1:]
if query_label[ind] == 1:
query_bbox_inside_weights[ind, :] = cfg.BBOX_INSIDE_WEIGHTS
elif query_label[ind] == 2:
query_bbox_inside_weights[ind, :] = 0.2
return query_bbox_targets, query_bbox_inside_weights
# 获取 query score和 bbox regression 的 label, mask
def get_labels(self, rpn_rois, gt_boxes):
# overlaps: (rois x gt_boxes)
# overlaps = bbox_overlaps(np.ascontiguousarray(rpn_rois, dtype=np.float), np.ascontiguousarray(gt_boxes[:, :4], dtype=np.float))
overlaps = bbox_overlaps(
np.ascontiguousarray(rpn_rois, dtype=np.float),
np.ascontiguousarray(gt_boxes[:, :4], dtype=np.float))
if self.use_kld:
query_label = np.zeros(rpn_rois.shape[0])
query_label_mask = 0
bbox_label = np.zeros(rpn_rois.shape[0])
# keep_inds = []
# 找出 query = 1 的 gt_box 的 index
query_gt_ind = 0
query_overlaps = overlaps[:, query_gt_ind].reshape(-1)
if self.use_kld:
# kld: 根据 iou 设置权重
if query_overlaps.max() >= 0.5:
query_label_mask = 1
query_inds = np.where(query_overlaps >= cfg.THRESHOLD)[0]
for ind in query_inds:
query_label[ind] = query_overlaps[ind]
if query_label.sum() == 0:
print(query_overlaps.max())
query_label = query_label / float(query_label.sum())
else:
# softmax
if query_overlaps.max() >= 0.5:
query_label = int(query_overlaps.argmax())
else:
query_label = -1
rois = rpn_rois
gt_assignment = overlaps.argmax(axis=1)
gt_target_boxes = gt_boxes[gt_assignment, :4]
bbox_label[np.where(overlaps.max(axis=1) >= 0.5)[0]] = 2
if query_overlaps.max() >= 0.5:
query_inds = np.where(query_overlaps >= cfg.THRESHOLD)[0]
bbox_label[query_inds] = 1
gt_target_boxes[query_inds] = gt_boxes[query_gt_ind, :4]
bbox_target_data = self.compute_targets(rois, gt_target_boxes,
bbox_label)
query_bbox_targets, query_bbox_inside_weights = self.get_query_bbox_regression_labels(
bbox_target_data)
query_bbox_outside_weights = np.array(
query_bbox_inside_weights > 0).astype(np.float32)
return query_label, query_label_mask, query_bbox_targets, query_bbox_inside_weights, query_bbox_outside_weights
def get_spt_feat(self, bbox, img_shape):
spt_feat = np.zeros((bbox.shape[0], 5), dtype=np.float)
spt_feat[:, 0] = bbox[:, 0] / float(img_shape[1])
spt_feat[:, 1] = bbox[:, 1] / float(img_shape[0])
spt_feat[:, 2] = bbox[:, 2] / float(img_shape[1])
spt_feat[:, 3] = bbox[:, 3] / float(img_shape[0])
spt_feat[:, 4] = (bbox[:, 2] - bbox[:, 0]) * (
bbox[:, 3] - bbox[:, 1]) / float(img_shape[0] * img_shape[1])
return spt_feat
class DataProvider(object):
def __init__(self, data_split, batchsize=1):
print 'init DataProvider for %s : %s : %s' % (cfg.IMDB_NAME, cfg.PROJ_NAME, data_split)
self.is_ss = cfg.FEAT_TYPE == 'ss'
self.ss_box_dir = cfg.SS_BOX_DIR
self.ss_feat_dir = cfg.SS_FEAT_DIR
self.feat_type = cfg.FEAT_TYPE
if 'refcoco' in cfg.IMDB_NAME or cfg.IMDB_NAME == 'refclef':
self.is_mscoco_prefix = True
else:
self.is_mscoco_prefix = False
self.use_kld = cfg.USE_KLD
# self.mscoco_prefix = cfg.MSCOCO_PREFIX
self.rpn_topn = cfg.RPN_TOPN
if self.is_ss:
self.bottomup_feat_dim = cfg.SS_FEAT_DIM
else:
self.bottomup_feat_dim = cfg.BOTTOMUP_FEAT_DIM
self.query_maxlen = cfg.QUERY_MAXLEN
# self.data_paths = cfg.DATA_PATHS
self.image_ext = '.jpg'
data_splits = data_split.split(cfg.SPLIT_TOK)
if 'train' in data_splits:
self.mode = 'train'
else:
self.mode = 'test'
self.batchsize = batchsize
self.image_dir = cfg.IMAGE_DIR
self.feat_dir = cfg.FEAT_DIR
self.dict_dir = cfg.QUERY_DIR #osp.join(cfg.DATA_DIR, cfg.IMDB_NAME, 'query_dict')
self.anno = self.load_data(data_splits)
self.qdic = Dictionary(self.dict_dir)
self.qdic.load()
self.index = 0
self.batch_len = None
self.num_query = len(self.anno)
def get_image_ids(self):
qid_list = self.get_query_ids()
iid_list = set()
for qid in qid_list:
iid_list.add(self.anno[qid]['iid'])
return list(iid_list)
def get_query_ids(self):
return self.anno.keys()
def get_num_query(self):
return self.num_query
def load_data(self, data_splits):
anno = {}
for data_split in data_splits:
#data_path = osp.join(cfg.DATA_DIR, cfg.IMDB_NAME, 'format_%s.pkl'%str(data_split))
data_path = cfg.ANNO_PATH%str(data_split)
t_anno = load(data_path)
anno.update(t_anno)
return anno
def get_vocabsize(self):
return self.qdic.size()
def get_iid(self, qid):
return self.anno[qid]['iid']
def get_img_path(self, iid):
if self.is_mscoco_prefix:
return os.path.join(self.image_dir, 'COCO_train2014_' + str(iid).zfill(12) + self.image_ext)
else:
return os.path.join(self.image_dir, str(iid) + self.image_ext)
def str2list(self, qstr, query_maxlen):
q_list = qstr.split()
qvec = np.zeros(query_maxlen, dtype=np.int64)
cvec = np.zeros(query_maxlen, dtype=np.int64)
for i,_ in enumerate(xrange(query_maxlen)):
if i < query_maxlen - len(q_list):
cvec[i] = 0
else:
w = q_list[i-(query_maxlen-len(q_list))]
# is the word in the vocabulary?
# if self.qdic.has_token(w) is False:
# w = cfg.UNK_WORD #'<unk>'
qvec[i] = self.qdic.lookup(w)
cvec[i] = 0 if i == query_maxlen - len(q_list) else 1
return qvec, cvec
def load_ss_box(self, ss_box_path):
boxes = np.loadtxt(ss_box_path)
if len(boxes) == 0:
raise Exception("boxes is None!")
boxes = boxes - 1
boxes[:,[0,1]] = boxes[:,[1,0]]
boxes[:,[2,3]] = boxes[:,[3,2]]
return boxes
def get_topdown_feat(self, iid):
try:
if self.is_ss:
img_path = self.get_img_path(iid)
im = skimage.io.imread(img_path)
img_h = im.shape[0]
img_w = im.shape[1]
feat_path = os.path.join(self.ss_feat_dir, str(iid) + '.npz')
ss_box_path = os.path.join(self.ss_box_dir, str(iid) + '.txt')
bbox = self.load_ss_box(ss_box_path)
num_bbox = bbox.shape[0]
img_feat = np.transpose(np.load(feat_path)['x'], (1,0))
else:
if self.is_mscoco_prefix:
feat_path = os.path.join(self.feat_dir, 'COCO_train2014_' + str(iid).zfill(12) + self.image_ext + '.npz')
else:
feat_path = os.path.join(self.feat_dir, str(iid) + self.image_ext + '.npz')
feat_dict = np.load(feat_path)
img_feat = feat_dict['x']
num_bbox = feat_dict['num_bbox']
bbox = feat_dict['bbox']
img_h = feat_dict['image_h']
img_w = feat_dict['image_w']
return img_feat, num_bbox, bbox, (img_h, img_w)
except Exception, e:
print e
raise Exception("UnkownError")