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.
Dear community I have a challenge with regard to tensor indexing in PyTorch. The problem is very simple. Given a tensor create an index tensor to index its maximum values per column.
x = T.tensor([[0, 3, 0, 5, 9, 8, 2, 0],
[0, 4, 9, 6, 7, 9, 1, 0]])
Given this tensor I would like to build a boolean mask for indexing its maximum values per colum. To be specific I do not need its maximum values, torch.max(x, dim=0), nor its indices, torch.argmax(x, dim=0), but a boolean mask for indexing other tensor based on this tensor max values. My ideal output would be:
# Input tensor
x
tensor([[0, 3, 0, 5, 9, 8, 2, 0],
[0, 4, 9, 6, 7, 9, 1, 0]])
# Ideal output bool mask tensor
idx
tensor([[1, 0, 0, 0, 1, 0, 1, 1],
[0, 1, 1, 1, 0, 1, 0, 0]])
I know that values_max = x[idx] and values_max = x.max(dim=0) are equivalent but I am not looking for values_max but for idx.
I have built a solution around it but it just seem to complex and I am sure torch have an optimized way to do this. I have tried to use torch.index_select with the output of x.argmax(dim=0) but failed so I built a custom solution that seems to cumbersome to me so I am asking for help to do this in a vectorized / tensorial / torch way.
You can perform this operation by first extracting the index of the maximum value column-wise of your tensor with torch.argmax, setting keepdim to True
>>> x.argmax(0, keepdim=True)
tensor([[0, 1, 1, 1, 0, 1, 0, 0]])
Then you can use torch.scatter to place 1s in a zero tensor at the designated indices:
>>> torch.zeros_like(x).scatter(0, x.argmax(0,True), value=1)
tensor([[1, 0, 0, 0, 1, 0, 1, 1],
[0, 1, 1, 1, 0, 1, 0, 0]])
I want to create a subset of my data by applying tf.data.Dataset filter operation. I have this data:
data = tf.convert_to_tensor([[1, 2, 1, 1, 5, 5, 9, 12], [1, 2, 3, 8, 4, 5, 9, 12]])
dataset = tf.data.Dataset.from_tensor_slices(data)
I want to retrieve a subset of 'dataset' which corresponds to all elements whose first column is equal to 1. So, result should be:
[[1, 1, 1], [1, 3, 8]] # dtype : dataset
I tried this:
subset = dataset.filter(lambda x: tf.equal(x[0], 1))
But I don't get the correct result, since it sends me back x[0]
Someone to help me ?
I finally resolved it:
a = tf.convert_to_tensor([1, 2, 1, 1, 5, 5, 9, 12])
b = tf.convert_to_tensor([1, 2, 3, 8, 4, 5, 9, 12])
data_set = tf.data.Dataset.from_tensor_slices((a, b))
subset = data_set.filter(lambda x, y: tf.equal(x, 1))
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.
I have a 1d array of values
i = np.arange(0,7,1)
and a function
# Returns a column matrix
def fn(i):
return np.matrix([[i*2,i*3]]).T
fnv = np.vectorize(fn)
then writing
fnv(i)
gives me an error
File "<stdin>", line 1, in <module>
File "c:\Python33\lib\site-packages\numpy\lib\function_base.py",
line 1872, in __call__
return self._vectorize_call(func=func, args=vargs)
File "c:\Python33\lib\site-packages\numpy\lib\function_base.py",
line 1942, in _vectorize_call
copy=False, subok=True, dtype=otypes[0])
ValueError: setting an array element with a sequence.
The result I am looking for is a matrix with two rows and as many columns as in the input array. What is the best notation in numpy to achieve this?
For example i would equal
[1,2,3,4,5,6]
and the output would equal
[[2,4,6,8,10,12],
[3,6,9,12,15,18]]
EDIT
You should try to avoid using vectorize, because it gives the illusion of numpy efficiency, but inside it's all python loops.
If you really have to deal with user supplied functions that take ints and return a matrix of shape (2, 1) then there probably isn't much you can do. But that seems like a really weird use case. If you can replace that with a list of functions that take an int and return an int, and that use ufuncs when needed, i.e. np.sin instead of math.sin, you can do the following
def vectorize2(funcs) :
def fnv(arr) :
return np.vstack([f(arr) for f in funcs])
return fnv
f2 = vectorize2((lambda x : 2 * x, lambda x : 3 * x))
>>> f2(np.arange(10))
array([[ 0, 2, 4, 6, 8, 10, 12, 14, 16, 18],
[ 0, 3, 6, 9, 12, 15, 18, 21, 24, 27]])
Just for your reference, I have timed this vectorization against your proposed one:
f = vectorize(fn)
>>> timeit.timeit('f(np.arange(10))', 'from __main__ import np, f', number=1000)
0.28073329263679625
>>> timeit.timeit('f2(np.arange(10))', 'from __main__ import np, f2', number=1000)
0.023139129945661807
>>> timeit.timeit('f(np.arange(10000))', 'from __main__ import np, f', number=10)
2.3620706288432984
>>> timeit.timeit('f2(np.arange(10000))', 'from __main__ import np, f2', number=10)
0.002757072593169596
So there is an order of magnitude in speed even for small arrays, that grows to a x1000 speed up, available almost for free, for larger arrays.
ORIGINAL ANSWER
Don't use vectorize unless there is no way around it, it's slow. See the following examples
>>> a = np.array(range(7))
>>> a
array([0, 1, 2, 3, 4, 5, 6])
>>> np.vstack((a, a+1))
array([[0, 1, 2, 3, 4, 5, 6],
[1, 2, 3, 4, 5, 6, 7]])
>>> np.vstack((a, a**2))
array([[ 0, 1, 2, 3, 4, 5, 6],
[ 0, 1, 4, 9, 16, 25, 36]])
Whatever your function is, if it can be constructed with numpy's ufuncs, you can do something like np.vstack((a, f(a))) and get what you want
A simple reimplementation of vectorize gives me what I want
def vectorize( fn):
def do_it (array):
return np.column_stack((fn(p) for p in array))
return do_it
If this is not performant or there is a better way then let me know.