def test_arr(self, mock_dat, mock_caffe):
# expected serialization of the test image
s = [
'\x08\x03\x10\x04\x18\x02"\x18\x01\x04\x07\n\r\x10\x13\x16\x02\x05\x08\x0b\x0e\x11\x14\x17\x03\x06\t\x0c\x0f\x12\x15\x18(\x00',
'\x08\x03\x10\x04\x18\x02"\x18\x02\x05\x08\x0b\x0e\x11\x14\x17\x03\x06\t\x0c\x0f\x12\x15\x18\x04\x07\n\r\x10\x13\x16\x19(\x00',
]
# mock caffe calls made by our module
mock_dat.return_value.SerializeToString = MagicMock(side_effect=s)
mock_caffe.io.array_to_datum.return_value = caffe.proto.caffe_pb2.Datum(
)
# use the module and test it
path_lmdb = os.path.join(self.dir_tmp, 'xarr2_lmdb')
tol.arrays_to_lmdb(self.arr, path_lmdb)
assert_true(os.path.isdir(path_lmdb), "failed to save LMDB")
count = 0
with lmdb.open(path_lmdb, readonly=True).begin() as txn:
for key, value in txn.cursor():
assert_equal(key, tol.IDX_FMT.format(count), "Unexpected key.")
assert_equal(value, s[count], "Unexpected content.")
count += 1
assert_equal(count, 2, "Unexpected number of samples.")
def setup_class(self):
self.dir_tmp = tempfile.mkdtemp()
x = np.array([[[1, 2, 3], [4, 5, 6]], [[7, 8, 9], [10, 11, 12]],
[
[13, 14, 15],
[16, 17, 18],
], [[19, 20, 21], [22, 23, 24]]])
tol.arrays_to_lmdb([y for y in x],
os.path.join(self.dir_tmp, 'x_lmdb'))
def setup_class(self):
self.dir_tmp = tempfile.mkdtemp()
x = np.array([[[ 0, 2, 3],
[ 4, 5, 6]
],
[[ 7, 8, 9],
[10, 11, 12]
],
[[13, 14, 15],
[16, 17, 18],
],
[[19, 20, 21],
[22, 23, 0]
]
])
tol.arrays_to_lmdb([y for y in x], os.path.join(self.dir_tmp, 'x_lmdb'))
def infer_to_lmdb(net, keys, n, dst_prefix):
"""
Run network inference for n batches and save results to an lmdb for each key.
Lower time complexity but much higher space complexity.
Not recommended for large datasets or large number of keys
See: infer_to_lmdb_cur() for slower alternative with less memory overhead
lmdb cannot preserve batches
"""
dc = {k:[] for k in keys}
for _ in range(n):
d = forward(net, keys)
for k in keys:
dc[k].extend(np.copy(d[k].astype(float)))
for k in keys:
to_lmdb.arrays_to_lmdb(dc[k], dst_prefix % (k,))
return [len(dc[k]) for k in keys]
def infer_to_lmdb(net, keys, n, dst_prefix):
"""
Run network inference for n batches and save results to an lmdb for each key.
Lower time complexity but much higher space complexity.
Not recommended for large datasets or large number of keys
See: infer_to_lmdb_cur() for slower alternative with less memory overhead
lmdb cannot preserve batches
"""
dc = {k: [] for k in keys}
for _ in range(n):
d = forward(net, keys)
for k in keys:
dc[k].extend(np.copy(d[k].astype(float)))
for k in keys:
to_lmdb.arrays_to_lmdb(dc[k], dst_prefix % (k, ))
return [len(dc[k]) for k in keys]
def nyudv2_to_lmdb(path_mat, dst_prefix, dir_dst, val_list=None):
val_list = val_list or []
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 nyudv2_to_lmdb(path_mat,
dst_prefix,
dir_dst,
val_list=None):
val_list = val_list or []
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
