def nyudv2_to_lmdb(path_mat, dst_prefix, dir_dst, val_list=[]):
if not os.path.isfile(path_mat):
raise IOError("Path is not a regular file (%s)" % path_mat)
_, ext = os.path.splitext(path_mat)
if ext != ".mat" and ext != ".h5" and ext != ".hdf5":
raise IOError("Invalid file type, expecting mat/h5/hdf5 file (%s)" % path_mat)
try:
data = io.loadmat(path_mat)
except (ValueError, NotImplementedError):
data = h5py.File(path_mat) # support version >= 7.3 matfile HDF5 format
pass
lmdb_info = []
train_idx = None
for typ in [NYUDV2DataType.IMAGES, NYUDV2DataType.LABELS, NYUDV2DataType.DEPTHS]:
if typ == NYUDV2DataType.IMAGES:
dat = [mu.cwh_to_chw(x).astype(np.float) for x in data[typ]]
elif typ == NYUDV2DataType.LABELS:
dat = np.expand_dims(data[typ], axis=1).astype(int)
dat = big_arr_to_arrs(dat)
elif typ == NYUDV2DataType.DEPTHS:
dat = np.expand_dims(data[typ], axis=1).astype(np.float)
dat = big_arr_to_arrs(dat)
else:
raise ValueError("unknown NYUDV2DataType")
if train_idx is None:
train_idx, val_idx = get_train_val_split_from_idx(len(dat), val_list)
shuffle(train_idx)
print (train_idx)
# # len(ndarray) same as ndarray.shape[0]
# if len(labels) != len(imgs):
# raise ValueError("No. of images != no. of labels. (%d) != (%d)",
# len(imgs), len(labels))
#
# if len(labels) != len(depths):
# raise ValueError("No. of depths != no. of labels. (%d) != (%d)",
# len(depths), len(labels))
print typ, len(dat), dat[0].shape
fpath_lmdb = os.path.join(dir_dst, "%s%s_train_lmdb" % (dst_prefix, typ))
to_lmdb.arrays_to_lmdb([dat[i] for i in train_idx], fpath_lmdb)
lmdb_info.append((len(train_idx), fpath_lmdb))
fpath_lmdb = os.path.join(dir_dst, "%s%s_val_lmdb" % (dst_prefix, typ))
to_lmdb.arrays_to_lmdb([dat[i] for i in val_idx], fpath_lmdb)
lmdb_info.append((len(val_idx), fpath_lmdb))
return lmdb_info
def big_arr_to_arrs(a):
"""
Turn NxCxWxH array into list of CxHxW for Caffe
"""
return [mu.cwh_to_chw(x) for x in a]