Related
I am trying to access a pytorch tensor by a matrix of indices and I recently found this bit of code that I cannot find the reason why it is not working.
The code below is split into two parts. The first half proves to work, whilst the second trips an error. I fail to see the reason why. Could someone shed some light on this?
import torch
import numpy as np
a = torch.rand(32, 16)
m, n = a.shape
xx, yy = np.meshgrid(np.arange(m), np.arange(m))
result = a[xx] # WORKS for a torch.tensor of size M >= 32. It doesn't work otherwise.
a = torch.rand(16, 16)
m, n = a.shape
xx, yy = np.meshgrid(np.arange(m), np.arange(m))
result = a[xx] # IndexError: too many indices for tensor of dimension 2
and if I change a = np.random.rand(16, 16) it does work as well.
To whoever comes looking for an answer: it looks like its a bug in pyTorch.
Indexing using numpy arrays is not well defined, and it works only if tensors are indexed using tensors. So, in my example code, this works flawlessly:
a = torch.rand(M, N)
m, n = a.shape
xx, yy = torch.meshgrid(torch.arange(m), torch.arange(m), indexing='xy')
result = a[xx] # WORKS
I made a gist to check it, and it's available here
First, let me give you a quick insight into the idea of indexing a tensor with a numpy array and another tensor.
Example: this is our target tensor to be indexed
numpy_indices = torch.tensor([[0, 1, 2, 7],
[0, 1, 2, 3]]) # numpy array
tensor_indices = torch.tensor([[0, 1, 2, 7],
[0, 1, 2, 3]]) # 2D tensor
t = torch.tensor([[1, 2, 3, 4], # targeted tensor
[5, 6, 7, 8],
[9, 10, 11, 12],
[13, 14, 15, 16],
[17, 18, 19, 20],
[21, 22, 23, 24],
[25, 26, 27, 28],
[29, 30, 31, 32]])
numpy_result = t[numpy_indices]
tensor_result = t[tensor_indices]
Indexing using a 2D numpy array: the index is read like pairs (x,y) tensor[row,column] e.g. t[0,0], t[1,1], t[2,2], and t[7,3].
print(numpy_result) # tensor([ 1, 6, 11, 32])
Indexing using a 2D tensor: walks through the index tensor in a row-wise manner and each value is an index of a row in the targeted tensor.
e.g. [ [t[0],t[1],t[2],[7]] , [[0],[1],[2],[3]] ] see the example below, the new shape of tensor_result after indexing is (tensor_indices.shape[0],tensor_indices.shape[1],t.shape[1])=(2,4,4).
print(tensor_result) # tensor([[[ 1, 2, 3, 4],
# [ 5, 6, 7, 8],
# [ 9, 10, 11, 12],
# [29, 30, 31, 32]],
# [[ 1, 2, 3, 4],
# [ 5, 6, 7, 8],
# [ 9, 10, 11, 12],
# [ 13, 14, 15, 16]]])
If you try to add a third row in numpy_indices, you will get the same error you have because the index will be represented by 3D e.g., (0,0,0)...(7,3,3).
indices = np.array([[0, 1, 2, 7],
[0, 1, 2, 3],
[0, 1, 2, 3]])
print(numpy_result) # IndexError: too many indices for tensor of dimension 2
However, this is not the case with indexing by tensor and the shape will be bigger (3,4,4).
Finally, as you see the outputs of the two types of indexing are completely different. To solve your problem, you can use
xx = torch.tensor(xx).long() # convert a numpy array to a tensor
What happens in the case of advanced indexing (rows of numpy_indices > 3 ) as your situation is still ambiguous and unsolved and you can check 1 , 2, 3.
Lets say I have a Python Numpy array a.
a = numpy.array([1,2,3,4,5,6,7,8,9,10,11])
I want to create a matrix of sub sequences from this array of length 5 with stride 3. The results matrix hence will look as follows:
numpy.array([[1,2,3,4,5],[4,5,6,7,8],[7,8,9,10,11]])
One possible way of implementing this would be using a for-loop.
result_matrix = np.zeros((3, 5))
for i in range(0, len(a), 3):
result_matrix[i] = a[i:i+5]
Is there a cleaner way to implement this in Numpy?
Approach #1 : Using broadcasting -
def broadcasting_app(a, L, S ): # Window len = L, Stride len/stepsize = S
nrows = ((a.size-L)//S)+1
return a[S*np.arange(nrows)[:,None] + np.arange(L)]
Approach #2 : Using more efficient NumPy strides -
def strided_app(a, L, S ): # Window len = L, Stride len/stepsize = S
nrows = ((a.size-L)//S)+1
n = a.strides[0]
return np.lib.stride_tricks.as_strided(a, shape=(nrows,L), strides=(S*n,n))
Sample run -
In [143]: a
Out[143]: array([ 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11])
In [144]: broadcasting_app(a, L = 5, S = 3)
Out[144]:
array([[ 1, 2, 3, 4, 5],
[ 4, 5, 6, 7, 8],
[ 7, 8, 9, 10, 11]])
In [145]: strided_app(a, L = 5, S = 3)
Out[145]:
array([[ 1, 2, 3, 4, 5],
[ 4, 5, 6, 7, 8],
[ 7, 8, 9, 10, 11]])
Starting in Numpy 1.20, we can make use of the new sliding_window_view to slide/roll over windows of elements.
And coupled with a stepping [::3], it simply becomes:
from numpy.lib.stride_tricks import sliding_window_view
# values = np.array([1,2,3,4,5,6,7,8,9,10,11])
sliding_window_view(values, window_shape = 5)[::3]
# array([[ 1, 2, 3, 4, 5],
# [ 4, 5, 6, 7, 8],
# [ 7, 8, 9, 10, 11]])
where the intermediate result of the sliding is:
sliding_window_view(values, window_shape = 5)
# array([[ 1, 2, 3, 4, 5],
# [ 2, 3, 4, 5, 6],
# [ 3, 4, 5, 6, 7],
# [ 4, 5, 6, 7, 8],
# [ 5, 6, 7, 8, 9],
# [ 6, 7, 8, 9, 10],
# [ 7, 8, 9, 10, 11]])
Modified version of #Divakar's code with checking to ensure that memory is contiguous and that the returned array cannot be modified. (Variable names changed for my DSP application).
def frame(a, framelen, frameadv):
"""frame - Frame a 1D array
a - 1D array
framelen - Samples per frame
frameadv - Samples between starts of consecutive frames
Set to framelen for non-overlaping consecutive frames
Modified from Divakar's 10/17/16 11:20 solution:
https://stackoverflow.com/questions/40084931/taking-subarrays-from-numpy-array-with-given-stride-stepsize
CAVEATS:
Assumes array is contiguous
Output is not writable as there are multiple views on the same memory
"""
if not isinstance(a, np.ndarray) or \
not (a.flags['C_CONTIGUOUS'] or a.flags['F_CONTIGUOUS']):
raise ValueError("Input array a must be a contiguous numpy array")
# Output
nrows = ((a.size-framelen)//frameadv)+1
oshape = (nrows, framelen)
# Size of each element in a
n = a.strides[0]
# Indexing in the new object will advance by frameadv * element size
ostrides = (frameadv*n, n)
return np.lib.stride_tricks.as_strided(a, shape=oshape,
strides=ostrides, writeable=False)
I'm trying to figure out how to do shifting on a tensor that has b (batch size), d (depth), h (hight) and w (width) represented as following:
b, d, h, w = tensor.size()
So, I need to find the subtract between the shifted tensor and the tensor itself.
I'm thinking of using torch.narrow or torch.concat to do it for each side (shift the right, left, up then down side) and at each time I subtract from the same tensor side (tensor itself side), then at the end I will add/sum the differences/subtractions of each side (so I will have the final subtraction between the shifted and the tensor itself.
I'm new to PyTorch, it's easy to understand but struggling to implemented and maybe there is a simpler way (directly do the subtraction rather than working on each side and so on .....)
Any help on that please?
Basically, you can split the tensor first, and then cat them in reverse order. I write a function to implement your thoughts. The shift should be a non-negative number and less than or equal to the size of dim.
def tensor_shift(t, dim, shift):
"""
t (tensor): tensor to be shifted.
dim (int): the dimension apply shift.
shift (int): shift distance.
"""
assert 0 <= shift <= t.size(dim), "shift distance should be smaller than or equal to the dim length."
overflow = t.index_select(dim, torch.arange(t.size(dim)-shift, t.size(dim)))
remain = t.index_select(dim, torch.arange(t.size(dim)-shift))
return torch.cat((overflow, remain),dim=dim)
Here are some test results.
a = torch.arange(1,13).view(-1,3)
a
#tensor([[ 1, 2, 3],
# [ 4, 5, 6],
# [ 7, 8, 9],
# [10, 11, 12]])
shift(a, 0, 1) # shift 1 unit along dim=0
#tensor([[10, 11, 12],
# [ 1, 2, 3],
# [ 4, 5, 6],
# [ 7, 8, 9]])
b = torch.arange(1,13).view(-1,2,3)
b
#tensor([[[ 1, 2, 3],
# [ 4, 5, 6]],
#
# [[ 7, 8, 9],
# [10, 11, 12]]])
shift(b, 1, 1) # shift 1 unit along dim=1
#tensor([[[ 4, 5, 6],
# [ 1, 2, 3]],
#
# [[10, 11, 12],
# [ 7, 8, 9]]])
If I create a numpy array, and another to serve as a selective index into it:
>>> x
array([[ 2, 3, 4],
[ 5, 6, 7],
[ 6, 7, 8],
[11, 12, 13]])
>>> nz
array([ True, True, False, True], dtype=bool)
then direct use of nz returns a view of the original array:
>>> x[nz,:]
array([[ 2, 3, 4],
[ 5, 6, 7],
[11, 12, 13]])
>>> x[nz,:] += 2
>>> x
array([[ 4, 5, 6],
[ 7, 8, 9],
[ 6, 7, 8],
[13, 14, 15]])
however, naturally, an assignment makes a copy:
>>> v = x[nz,:]
Any operation on v is on the copy, and has no effect on the original array.
Is there any way to create a named view, from x[nz,:], simply to abbreviate code, or which I can pass around, so operations on the named view will affect only the selected elements of x?
Numpy has masked_array, which might be what you are looking for:
import numpy as np
x = np.asarray([[ 2, 3, 4],[ 5, 6, 7],[ 6, 7, 8],[11, 12, 13]])
nz = np.asarray([ True, True, False, True], dtype=bool)
mx = np.ma.masked_array(x, ~nz.repeat(3)) # True means masked, so "~" is needed
mx += 2
# x changed as well because it is the base of mx
print(x)
print(x is mx.base)
Assume we have an array with NxMxD shape. I want to get a list with D NxM arrays.
The correct way of doing it would be:
np.dsplit(myarray, D)
However, this returns D NxMx1 arrays.
I can achieve the desired result by doing something like:
[myarray[..., i] for i in range(D)]
Or:
[np.squeeze(subarray) for subarray in np.dsplit(myarray, D)]
However, I feel like it is a bit redundant to need to perform an additional operation. Am I missing any numpy function that returns the desired result?
Try D.swapaxes(1,2).swapaxes(1,0)
>>>import numpy as np
>>>a = np.arange(24).reshape(2,3,4)
>>>a
array([[[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]],
[[12, 13, 14, 15],
[16, 17, 18, 19],
[20, 21, 22, 23]]])
>>>[a[:,:,i] for i in range(4)]
[array([[ 0, 4, 8],
[12, 16, 20]]),
array([[ 1, 5, 9],
[13, 17, 21]]),
array([[ 2, 6, 10],
[14, 18, 22]]),
array([[ 3, 7, 11],
[15, 19, 23]])]
>>>a.swapaxes(1,2).swapaxes(1,0)
array([[[ 0, 4, 8],
[12, 16, 20]],
[[ 1, 5, 9],
[13, 17, 21]],
[[ 2, 6, 10],
[14, 18, 22]],
[[ 3, 7, 11],
[15, 19, 23]]])
Edit: As pointed out by ajcr (thanks again), the transpose command is more convenient since the two swaps can be done in one step by using
D.transpose(2,0,1)
np.dsplit uses np.array_split, the core of which is:
sub_arys = []
sary = _nx.swapaxes(ary, axis, 0)
for i in range(Nsections):
st = div_points[i]; end = div_points[i+1]
sub_arys.append(_nx.swapaxes(sary[st:end], axis, 0))
with axis=-1, this is equivalent to:
[x[...,i:(i+1)] for i in np.arange(x.shape[-1])] # or
[x[...,[i]] for i in np.arange(x.shape[-1])]
which accounts for the singleton dimension.
So there's nothing wrong or inefficient about your
[x[...,i] for i in np.arange(x.shape[-1])]
Actually in quick time tests, any use of dsplit is slow. It's generality costs. So adding squeeze is relatively cheap.
But by accepting the other answer, it looks like you are really looking for an array of the correct shape, rather than a list of arrays. For many operations that makes sense. split is more useful when the subarrays have more than one 'row' along the split axis, or even an uneven number of 'rows'.