使用 h5py 从大文件中读取而不将整个文件加载到内存中

Question

Does the following read from a dataset without loading the entire thing at once into memory [the whole thing will not fit into memory] and get the size of the dataset without loading the data using h5py in python? if not, how?

h5 = h5py.File('myfile.h5', 'r')
mydata = h5.get('matirx') # are all data loaded into memory by using h5.get?
part_of_mydata= mydata[1000:11000,:]
size_data =  mydata.shape 

Thanks.

Answer 1

get (or indexing) fetches a reference to the Dataset on the file, but does not load any data.

In [789]: list(f.keys())
Out[789]: ['dset', 'dset1', 'vset']
In [790]: d=f['dset1']
In [791]: d
Out[791]: <HDF5 dataset "dset1": shape (2, 3, 10), type "<f8">
In [792]: d.shape         # shape of dataset
Out[792]: (2, 3, 10)
In [793]: arr=d[:,:,:5]    # indexing the set fetches part of the data
In [794]: arr.shape
Out[794]: (2, 3, 5)
In [795]: type(d)
Out[795]: h5py._hl.dataset.Dataset
In [796]: type(arr)
Out[796]: numpy.ndarray

d the Dataset is array like, but not actually a numpy array.

Fetch the whole Dataset with:

In [798]: arr = d[:]
In [799]: type(arr)
Out[799]: numpy.ndarray

Exactly how of the file it has to read to fetch yourslice depends on the slicing, data layout, chunking, and other things that generally aren't under your control, and shouldn't worry you.

Note also that when reading one dataset I'm not loading the others. Same would apply to groups.

http://docs.h5py.org/en/latest/high/dataset.html#reading-writing-data