python - 使用 h5py 创建大量数据集 - 无法注册数据类型原子(无法插入重复键)

Question

我正在尝试将大量 numpy 结构化数组作为数据集存储在 hdf5 文件中。
例如，
f['tree1'] = structured_array1
。
。
f['tree60000'] = structured_array60000(大约有 60000 棵树)，

读取文件的大约 70% 时，出现错误RuntimeError:无法注册数据类型原子(无法插入重复键)

此问题仅发生在非常大的 ascii 文件中(10e7 行，5gb)。如果文件大约(10e6 行，500mb)，则不会发生。如果我取出数据类型并仅存储为字符串的 numpy 数组，它也不会发生。

如果我中途停止读取文件，关闭我的终端，再次打开它，并继续从中途读取文件到最后(我保存我结束的行号)，我可以解决这个问题。我尝试在 python 函数本身中打开和关闭 hdf5 文件，但这没有用。

dt = [
('scale', 'f4'), 
('haloid', 'i8'), 
('scale_desc', 'f4'), 
('haloid_desc', 'i8'), 
('num_prog', 'i4'), 
('pid', 'i8'), 
('upid', 'i8'), 
('pid_desc', 'i8'), 
('phantom', 'i4'), 
('mvir_sam', 'f4'), 
('mvir', 'f4'), 
('rvir', 'f4'), 
('rs', 'f4'), 
('vrms', 'f4'), 
('mmp', 'i4'), 
('scale_lastmm', 'f4'), 
('vmax', 'f4'), 
('x', 'f4'), 
('y', 'f4'), 
('z', 'f4'), 
('vx', 'f4'), 
('vy', 'f4'), 
('vz', 'f4'), 
('jx', 'f4'), 
('jy', 'f4'), 
('jz', 'f4'), 
('spin', 'f4'), 
('haloid_breadth_first', 'i8'), 
('haloid_depth_first', 'i8'), 
('haloid_tree_root', 'i8'), 
('haloid_orig', 'i8'), 
('snap_num', 'i4'), 
('haloid_next_coprog_depthfirst', 'i8'), 
('haloid_last_prog_depthfirst', 'i8'), 
('haloid_last_mainleaf_depthfirst', 'i8'), 
('rs_klypin', 'f4'), 
('mvir_all', 'f4'), 
('m200b', 'f4'), 
('m200c', 'f4'), 
('m500c', 'f4'), 
('m2500c', 'f4'), 
('xoff', 'f4'), 
('voff', 'f4'), 
('spin_bullock', 'f4'), 
('b_to_a', 'f4'), 
('c_to_a', 'f4'), 
('axisA_x', 'f4'), 
('axisA_y', 'f4'), 
('axisA_z', 'f4'), 
('b_to_a_500c', 'f4'), 
('c_to_a_500c', 'f4'), 
('axisA_x_500c', 'f4'), 
('axisA_y_500c', 'f4'), 
('axisA_z_500c', 'f4'), 
('t_by_u', 'f4'), 
('mass_pe_behroozi', 'f4'), 
('mass_pe_diemer', 'f4')
]

def read_in_trees(self):
    """Store each tree as an hdf5 dataset.
    """  
    with open(self.fname) as ascii_file:
        with h5py.File(self.hdf5_name,"r+") as f:
            tree_id = ""  
            current_tree = []
            for line in ascii_file:
                if(line[0]=='#'): #new tree
                        arr = np.array(current_tree, dtype = dt)
                        f[tree_id] = arr
                        current_tree = []
                    tree_id = line[6:].strip('\n')
                else: #read in next tree element
                    current_tree.append(tuple(line.split()))
    return 

错误:

/Volumes/My Passport for Mac/raw_trees/bolshoi/rockstar/asciiReaderOne.py in read_in_trees(self)
    129                             arr = np.array(current_tree, dtype = dt)
    130                             # depth_sort =  arr['haloid_depth_first'].argsort()
--> 131                             f[tree_id] = arr
    132                             current_tree = []
    133                         first_line = False

/Library/Python/2.7/site-packages/h5py/_objects.so in h5py._objects.with_phil.wrapper (/Users/travis/build/MacPython/h5py-wheels/h5py/h5py/_objects.c:2458)()

/Library/Python/2.7/site-packages/h5py/_objects.so in h5py._objects.with_phil.wrapper (/Users/travis/build/MacPython/h5py-wheels/h5py/h5py/_objects.c:2415)()

/Library/Python/2.7/site-packages/h5py/_hl/group.pyc in __setitem__(self, name, obj)
    281 
    282         else:
--> 283             ds = self.create_dataset(None, data=obj, dtype=base.guess_dtype(obj))
    284             h5o.link(ds.id, self.id, name, lcpl=lcpl)
    285 

/Library/Python/2.7/site-packages/h5py/_hl/group.pyc in create_dataset(self, name, shape, dtype, data, **kwds)
    101         """
    102         with phil:
--> 103             dsid = dataset.make_new_dset(self, shape, dtype, data, **kwds)
    104             dset = dataset.Dataset(dsid)
    105             if name is not None:

/Library/Python/2.7/site-packages/h5py/_hl/dataset.pyc in make_new_dset(parent, shape, dtype, data, chunks, compression, shuffle, fletcher32, maxshape, compression_opts, fillvalue, scaleoffset, track_times)
    124 
    125     if data is not None:
--> 126         dset_id.write(h5s.ALL, h5s.ALL, data)
    127 
    128     return dset_id

/Library/Python/2.7/site-packages/h5py/_objects.so in h5py._objects.with_phil.wrapper (/Users/travis/build/MacPython/h5py-wheels/h5py/h5py/_objects.c:2458)()

/Library/Python/2.7/site-packages/h5py/_objects.so in h5py._objects.with_phil.wrapper (/Users/travis/build/MacPython/h5py-wheels/h5py/h5py/_objects.c:2415)()

/Library/Python/2.7/site-packages/h5py/h5d.so in h5py.h5d.DatasetID.write (/Users/travis/build/MacPython/h5py-wheels/h5py/h5py/h5d.c:3260)()

/Library/Python/2.7/site-packages/h5py/h5t.so in h5py.h5t.py_create (/Users/travis/build/MacPython/h5py-wheels/h5py/h5py/h5t.c:15314)()

/Library/Python/2.7/site-packages/h5py/h5t.so in h5py.h5t.py_create (/Users/travis/build/MacPython/h5py-wheels/h5py/h5py/h5t.c:14903)()

/Library/Python/2.7/site-packages/h5py/h5t.so in h5py.h5t._c_compound (/Users/travis/build/MacPython/h5py-wheels/h5py/h5py/h5t.c:14192)()

/Library/Python/2.7/site-packages/h5py/h5t.so in h5py.h5t.py_create (/Users/travis/build/MacPython/h5py-wheels/h5py/h5py/h5t.c:15314)()

/Library/Python/2.7/site-packages/h5py/h5t.so in h5py.h5t.py_create (/Users/travis/build/MacPython/h5py-wheels/h5py/h5py/h5t.c:14749)()

/Library/Python/2.7/site-packages/h5py/h5t.so in h5py.h5t._c_float (/Users/travis/build/MacPython/h5py-wheels/h5py/h5py/h5t.c:12379)()

RuntimeError: Unable to register datatype atom (Can't insert duplicate key)

Answer 1

你得到错误堆栈了吗？指示代码中何处产生错误？

您报告:错误 RuntimeError: Unable to register datatype atom (Can't insert duplicate key)

在/usr/lib/python3/dist-packages/h5py/_hl/datatype.py

class Datatype(HLObject):
    # Represents an HDF5 named datatype stored in a file.
    # >>> MyGroup["name"] = numpy.dtype("f")
    def __init__(self, bind):
        """ Create a new Datatype object by binding to a low-level TypeID.

我在这里抛出一个猜测。您的 dt 有 57 个术语。我怀疑每次您向文件添加一个树时，它都会将每个字段注册为一个新的数据类型。

In [71]: (57*10e7*.7)/(2**32)
Out[71]: 0.9289942681789397

57 * 10e7 的 70% 接近于 2*32。如果 Python/numpy 使用 int32 作为 dtype id，那么你可能会达到这个限制。

我们必须在 h5py 或 numpy 代码中进行更多挖掘，才能找到发出此错误消息的人。

通过向文件添加一个数组:

f[tree_id] = arr

您将每个数组放入新组 中的数据集中。如果每个 Dataset 都有一个数据类型，或数组的每个字段的数据类型，您可以轻松获得 2*32 数据类型。

另一方面，如果您可以将多个 arr 存储到一个组或数据集中，您可能会避免注册数千种数据类型。我对 h5py 不够熟悉，无法建议您如何操作。

---

我想知道这个序列是否可以重用多个数据集的数据类型:

dt1=np.dtype(dt)
gg= f.create_group('testgroup')
gg['xdtype']=dt1
# see h5py.Datatype doc
xdtype=gg['xdtype']
x=np.zeros((10,),dtype=xdtype)
gg['tree1']=x
x=np.ones((10,),dtype=xdtype)
gg['tree2']=x

根据 Datatype 文档，我尝试注册一个命名数据类型，并将其用于添加到组中的每个数据集。

In [117]: isinstance(xdtype, h5py.Datatype)
Out[117]: True
In [118]: xdtype.id
Out[118]: <h5py.h5t.TypeCompoundID at 0xb46e0d4c>

因此，如果我正确阅读 def make_new_dset，这将绕过 py_create 调用。