How To Change Numpy Array Dtype And Reshape?

I have an array that I read from an HDF5 file, and it is a 1D array of tuples. Its dtype is: [('cycle', '

Solution 1:

The mix of dtypes makes this conversion trickier than usual. The answer at the end, copying fields to a target array has the combination of speed and generality.

Convert structured array to regular NumPy array - was suggested as a duplicate, but that case has all float fields.

Let's construct a sample:

In [850]: dt
Out[850]: dtype([('cycle', '<u2'), ('dxn', 'i1'), ('i (mA)', '<f4'), ('V', '<f4'), ('R(Ohm)', '<f4')])

In [851]: x=np.zeros((3,),dt)
In [852]: x['cycle']=[0,10,23]
In [853]: x['dxn']=[3,2,2]
In [854]: x['V']=[1,1,1]

In [855]: x
Out[855]: 
array([(0, 3, 0.0, 1.0, 0.0), (10, 2, 0.0, 1.0, 0.0),
       (23, 2, 0.0, 1.0, 0.0)], 
      dtype=[('cycle', '<u2'), ('dxn', 'i1'), ('i (mA)', '<f4'), ('V', '<f4'), ('R(Ohm)', '<f4')])

We can view the 3 float fields in ways suggested in that link:

In [856]: dt1=np.dtype([('f0','float32',(3))])

In [857]: y=x[list(x.dtype.names[2:])].view(dt1)
# or x[list(x.dtype.names[2:])].view((np.float32, 3))

In [858]: y
Out[858]: 
array([([0.0, 1.0, 0.0],), ([0.0, 1.0, 0.0],), ([0.0, 1.0, 0.0],)], 
      dtype=[('f0', '<f4', (3,))])

In [859]: y['f0']
Out[859]: 
array([[ 0.,  1.,  0.],
       [ 0.,  1.,  0.],
       [ 0.,  1.,  0.]], dtype=float32)

But I need to make y a copy if I want to change all the values. Writing to multiple fields at a time is not allowed.

In [863]: y=x[list(x.dtype.names[2:])].view(dt1).copy()
In [864]: y['f0']=np.arange(9.).reshape(3,3)

view with one dtype does not capture the row structure; we have to add that back with reshape. dt1 with a (3,) shape gets around that issue.

In [867]: x[list(x.dtype.names[2:])].view(np.float32)
Out[867]: array([ 0.,  1.,  0.,  0.,  1.,  0.,  0.,  1.,  0.], dtype=float32)

https://stackoverflow.com/a/5957455/901925 suggests going through a list.

In [868]: x.tolist()
Out[868]: [(0, 3, 0.0, 1.0, 0.0), (10, 2, 0.0, 1.0, 0.0), (23, 2, 0.0, 1.0, 0.0)]

In [869]: np.array(x.tolist())
Out[869]: 
array([[  0.,   3.,   0.,   1.,   0.],
       [ 10.,   2.,   0.,   1.,   0.],
       [ 23.,   2.,   0.,   1.,   0.]])

Individual fields can be converted with astype:

In[878]: x['cycle'].astype(np.float32)
Out[878]: array([  0.,  10.,  23.], dtype=float32)

In[879]: x['dxn'].astype(np.float32)
Out[879]: array([ 3.,  2.,  2.], dtype=float32)

but not multiple fields:

In [880]: x.astype(np.float32)
Out[880]: array([  0.,  10.,  23.], dtype=float32)

recfunctions help manipulated structured arrays (and recarrays)

from numpy.libimport recfunctions

Many of them construct a new empty structure, and copy values field by field. The equivalent in this case:

In [890]: z=np.zeros((3,5),np.float32)    
In [891]: for i in range(5):
   .....:     z[:,i] = x[x.dtype.names[i]]

In [892]: z
Out[892]: 
array([[  0.,   3.,   0.,   1.,   0.],
       [ 10.,   2.,   0.,   1.,   0.],
       [ 23.,   2.,   0.,   1.,   0.]], dtype=float32)

In this small case it is a bit slower than np.array(x.tolist()). But for 30000 records this is much faster.

Usually there are many more records than fields in a structured array, so iteration on fields is not slow.

Share