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.
Related
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]]])
Is it possible to look up entries from an nd array without throwing an IndexError?
I'm hoping for something like:
>>> a = np.arange(10) * 2
>>> a[[-4, 2, 8, 12]]
IndexError
>>> wrap(a, default=-1)[[-4, 2, 8, 12]]
[-1, 4, 16, -1]
>>> wrap(a, default=-1)[200]
-1
Or possibly more like get_with_default(a, [-4, 2, 8, 12], default=-1)
Is there some builtin way to do this? Can I ask numpy not to throw the exception and return garbage, which I can then replace with my default value?
np.take with clip mode, sort of does this
In [155]: a
Out[155]: array([ 0, 2, 4, 6, 8, 10, 12, 14, 16, 18])
In [156]: a.take([-4,2,8,12],mode='raise')
...
IndexError: index 12 is out of bounds for size 10
In [157]: a.take([-4,2,8,12],mode='wrap')
Out[157]: array([12, 4, 16, 4])
In [158]: a.take([-4,2,8,12],mode='clip')
Out[158]: array([ 0, 4, 16, 18])
Except you don't have much control over the return value - here indexing on 12 return 18, the last value. And treated the -4 as out of bounds in the other direction, returning 0.
One way of adding the defaults is to pad a first
In [174]: a = np.arange(10) * 2
In [175]: ind=np.array([-4,2,8,12])
In [176]: np.pad(a, [1,1], 'constant', constant_values=-1).take(ind+1, mode='clip')
Out[176]: array([-1, 4, 16, -1])
Not exactly pretty, but a start.
This is my first post on any stack exchange site so forgive me for any stylistic errors (hopefully there are only stylistic errors). I am interested in the same feature but could not find anything from numpy better than np.take mentioned by hpaulj. Still np.take doesn't do exactly what's needed. Alfe's answer works but would need some elaboration in order to handle n-dimensional inputs. The following is another workaround that generalizes to the n-dimensional case. The basic idea is similar the one used by Alfe: create a new index with the out of bounds indices masked out (in my case) or disguised (in Alfe's case) and use it to index the input array without raising an error.
def take(a,indices,default=0):
#initialize mask; will broadcast to length of indices[0] in first iteration
mask = True
for i,ind in enumerate(indices):
#each element of the mask is only True if all indices at that position are in bounds
mask = mask & (0 <= ind) & (ind < a.shape[i])
#create in_bound indices
in_bound = [ind[mask] for ind in indices]
#initialize result with default value
result = default * np.ones(len(mask),dtype=a.dtype)
#set elements indexed by in_bound to their appropriate values in a
result[mask] = a[tuple(in_bound)]
return result
And here is the output from Eric's sample problem:
>>> a = np.arange(10)*2
>>> indices = (np.array([-4,2,8,12]),)
>>> take(a,indices,default=-1)
array([-1, 4, 16, -1])
You can restrict the range of the indexes to the size of your value array you want to index in using np.maximum() and np.minimum().
Example:
I have a heatmap like
h = np.array([[ 2, 3, 1],
[ 3, -1, 5]])
and I have a palette of RGB values I want to use to color the heatmap. The palette only names colors for the values 0..4:
p = np.array([[0, 0, 0], # black
[0, 0, 1], # blue
[1, 0, 1], # purple
[1, 1, 0], # yellow
[1, 1, 1]]) # white
Now I want to color my heatmap using the palette:
p[h]
Currently this leads to an error because of the values -1 and 5 in the heatmap:
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
IndexError: index 5 is out of bounds for axis 0 with size 5
But I can limit the range of the heatmap:
p[np.maximum(np.minimum(h, 4), 0)]
This works and gives me the result:
array([[[1, 0, 1],
[1, 1, 0],
[0, 0, 1]],
[[1, 1, 0],
[0, 0, 0],
[1, 1, 1]]])
If you really need to have a special value for the indexes which are out of bound, you could implement your proposed get_with_default() like this:
def get_with_default(values, indexes, default=-1):
return np.concatenate([[default], values, [default]])[
np.maximum(np.minimum(indexes, len(values)), -1) + 1]
a = np.arange(10) * 2
get_with_default(a, [-4, 2, 8, 12], default=-1)
Will return:
array([-1, 4, 16, -1])
as wanted.
in [http://deeplearning.net/tutorial/lenet.html#lenet] it says:
This will generate a matrix of shape (batch_size, nkerns[1] * 4 * 4),
# or (500, 50 * 4 * 4) = (500, 800) with the default values.
layer2_input = layer1.output.flatten(2)
when I use flatten function on a numpy 3d array I get a 1D array. but here it says I get a matrix. How does flatten(2) work in theano?
A similar example on numpy produces 1D array:
a= 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],
[25, 26, 27]]])
a.flatten(2)=array([ 1, 10, 19, 4, 13, 22, 7, 16, 25, 2, 11, 20, 5, 14, 23, 8, 17,
26, 3, 12, 21, 6, 15, 24, 9, 18, 27])
numpy doesn't support flattening only some dimensions but Theano does.
So if a is a numpy array, a.flatten(2) doesn't make any sense. It runs without error but only because the 2 is passed as the order parameter which seems to cause numpy to stick with the default order of C.
Theano's flatten does support axis specification. The documentation explains how it works.
Parameters:
x (any TensorVariable (or compatible)) – variable to be flattened
outdim (int) – the number of dimensions in the returned variable
Return type:
variable with same dtype as x and outdim dimensions
Returns:
variable with the same shape as x in the leading outdim-1 dimensions,
but with all remaining dimensions of x collapsed into the last dimension.
For example, if we flatten a tensor of shape (2, 3, 4, 5) with
flatten(x, outdim=2), then we’ll have the same (2-1=1) leading
dimensions (2,), and the remaining dimensions are collapsed. So the
output in this example would have shape (2, 60).
A simple Theano demonstration:
import numpy
import theano
import theano.tensor as tt
def compile():
x = tt.tensor3()
return theano.function([x], x.flatten(2))
def main():
a = numpy.arange(2 * 3 * 4).reshape((2, 3, 4))
f = compile()
print a.shape, f(a).shape
main()
prints
(2L, 3L, 4L) (2L, 12L)
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'.