I have two matrices in Python 2.7: one dense A_dense and the another sparse matrix A_sparse. I am interested in computing element-wise multiplication followed by sum. There are two ways to do it: use numpy's multiplication or scipy sparse multiplication. I expect them to give exactly same result with difference in execution time. But I find that they give different results for certain matrix sizes.
import numpy as np
from scipy import sparse
L=2000
np.random.seed(2)
rand_x=np.random.rand(L)
A_sparse_init=np.diag(rand_x, -1)+np.diag(rand_x, 1)
A_sparse=sparse.csr_matrix(A_sparse_init)
A_dense=np.random.rand(L+1,L+1)
print np.sum(A_sparse.multiply(A_dense))-np.sum(np.multiply(A_dense[A_sparse.nonzero()], A_sparse.data))
Output:
1.1368683772161603e-13
If I choose L=2001, then output is:
0.0
To check the size dependence of the difference using two different multiplication method, I wrote:
L=100
np.random.seed(2)
N_loop=100
multiply_diff_arr=np.zeros(N_loop)
for i in xrange(N_loop):
rand_x=np.random.rand(L)
A_sparse_init=np.diag(rand_x, -1)+np.diag(rand_x, 1)
A_sparse=sparse.csr_matrix(A_sparse_init)
A_dense=np.random.rand(L+1,L+1)
multiply_diff_arr[i]=np.sum(A_sparse.multiply(A_dense))-np.sum(np.multiply(A_dense[A_sparse.nonzero()], A_sparse.data))
L+=1
I got the following plot:
Can anyone help me understand what's happening? Don't we expect the difference between two methods to be at least 1e-18 rather than 1e-13?
I don't have a full answer, but this might help find the answer:
Under the hood, scipy.sparse will convert to coo format and do this:
ret = self.tocoo()
if self.shape == other.shape:
data = np.multiply(ret.data, other[ret.row, ret.col])
The question is then why these two operations give different results:
ret = A_sparse.tocoo()
c = np.multiply(ret.data, A_dense[ret.row, ret.col])
ret.data = c.view(type=np.ndarray)
c.sum() - ret.sum()
-1.1368683772161603e-13
Edit:
The difference stems from different defaults on which axis to add.reduce first.
E.g.:
A_sparse.multiply(A_dense).sum(axis=1).sum()
A_sparse.multiply(A_dense).sum(axis=0).sum()
Numpy defaults to 0 first.
I am trying to calculate the first and second order moments for a portfolio of stocks (i.e. expected return and standard deviation).
expected_returns_annual
Out[54]:
ticker
adj_close CNP 0.091859
F -0.007358
GE 0.095399
TSLA 0.204873
WMT -0.000943
dtype: float64
type(expected_returns_annual)
Out[55]: pandas.core.series.Series
weights = np.random.random(num_assets)
weights /= np.sum(weights)
returns = np.dot(expected_returns_annual, weights)
So normally the expected return is calculated by
(x1,...,xn' * (R1,...,Rn)
with x1,...,xn are weights with a constraint that all the weights have to sum up to 1 and ' means that the vector is transposed.
Now I am wondering a bit about the numpy dot function, because
returns = np.dot(expected_returns_annual, weights)
and
returns = np.dot(expected_returns_annual, weights.T)
give the same results.
I tested also the shape of weights.T and weights.
weights.shape
Out[58]: (5,)
weights.T.shape
Out[59]: (5,)
The shape of weights.T should be (,5) and not (5,), but numpy displays them as equal (I also tried np.transpose, but there is the same result)
Does anybody know why numpy behave this way? In my opinion the np.dot product automatically shape the vector the right why so that the vector product work well. Is that correct?
Best regards
Tom
The semantics of np.dot are not great
As Dominique Paul points out, np.dot has very heterogenous behavior depending on the shapes of the inputs. Adding to the confusion, as the OP points out in his question, given that weights is a 1D array, np.array_equal(weights, weights.T) is True (array_equal tests for equality of both value and shape).
Recommendation: use np.matmul or the equivalent # instead
If you are someone just starting out with Numpy, my advice to you would be to ditch np.dot completely. Don't use it in your code at all. Instead, use np.matmul, or the equivalent operator #. The behavior of # is more predictable than that of np.dot, while still being convenient to use. For example, you would get the same dot product for the two 1D arrays you have in your code like so:
returns = expected_returns_annual # weights
You can prove to yourself that this gives the same answer as np.dot with this assert:
assert expected_returns_annual # weights == expected_returns_annual.dot(weights)
Conceptually, # handles this case by promoting the two 1D arrays to appropriate 2D arrays (though the implementation doesn't necessarily do this). For example, if you have x with shape (N,) and y with shape (M,), if you do x # y the shapes will be promoted such that:
x.shape == (1, N)
y.shape == (M, 1)
Complete behavior of matmul/#
Here's what the docs have to say about matmul/# and the shapes of inputs/outputs:
If both arguments are 2-D they are multiplied like conventional matrices.
If either argument is N-D, N > 2, it is treated as a stack of matrices residing in the last two indexes and broadcast accordingly.
If the first argument is 1-D, it is promoted to a matrix by prepending a 1 to its dimensions. After matrix multiplication the prepended 1 is removed.
If the second argument is 1-D, it is promoted to a matrix by appending a 1 to its dimensions. After matrix multiplication the appended 1 is removed.
Notes: the arguments for using # over dot
As hpaulj points out in the comments, np.array_equal(x.dot(y), x # y) for all x and y that are 1D or 2D arrays. So why do I (and why should you) prefer #? I think the best argument for using # is that it helps to improve your code in small but significant ways:
# is explicitly a matrix multiplication operator. x # y will raise an error if y is a scalar, whereas dot will make the assumption that you actually just wanted elementwise multiplication. This can potentially result in a hard-to-localize bug in which dot silently returns a garbage result (I've personally run into that one). Thus, # allows you to be explicit about your own intent for the behavior of a line of code.
Because # is an operator, it has some nice short syntax for coercing various sequence types into arrays, without having to explicitly cast them. For example, [0,1,2] # np.arange(3) is valid syntax.
To be fair, while [0,1,2].dot(arr) is obviously not valid, np.dot([0,1,2], arr) is valid (though more verbose than using #).
When you do need to extend your code to deal with many matrix multiplications instead of just one, the ND cases for # are a conceptually straightforward generalization/vectorization of the lower-D cases.
I had the same question some time ago. It seems that when one of your matrices is one dimensional, then numpy will figure out automatically what you are trying to do.
The documentation for the dot function has a more specific explanation of the logic applied:
If both a and b are 1-D arrays, it is inner product of vectors
(without complex conjugation).
If both a and b are 2-D arrays, it is matrix multiplication, but using
matmul or a # b is preferred.
If either a or b is 0-D (scalar), it is equivalent to multiply and
using numpy.multiply(a, b) or a * b is preferred.
If a is an N-D array and b is a 1-D array, it is a sum product over
the last axis of a and b.
If a is an N-D array and b is an M-D array (where M>=2), it is a sum
product over the last axis of a and the second-to-last axis of b:
In NumPy, a transpose .T reverses the order of dimensions, which means that it doesn't do anything to your one-dimensional array weights.
This is a common source of confusion for people coming from Matlab, in which one-dimensional arrays do not exist. See Transposing a NumPy Array for some earlier discussion of this.
np.dot(x,y) has complicated behavior on higher-dimensional arrays, but its behavior when it's fed two one-dimensional arrays is very simple: it takes the inner product. If we wanted to get the equivalent result as a matrix product of a row and column instead, we'd have to write something like
np.asscalar(x # y[:, np.newaxis])
adding a trailing dimension to y to turn it into a "column", multiplying, and then converting our one-element array back into a scalar. But np.dot(x,y) is much faster and more efficient, so we just use that.
Edit: actually, this was dumb on my part. You can, of course, just write matrix multiplication x # y to get equivalent behavior to np.dot for one-dimensional arrays, as tel's excellent answer points out.
The shape of weights.T should be (,5) and not (5,),
suggests some confusion over the shape attribute. shape is an ordinary Python tuple, i.e. just a set of numbers, one for each dimension of the array. That's analogous to the size of a MATLAB matrix.
(5,) is just the way of displaying a 1 element tuple. The , is required because of older Python history of using () as a simple grouping.
In [22]: tuple([5])
Out[22]: (5,)
Thus the , in (5,) does not have a special numpy meaning, and
In [23]: (,5)
File "<ipython-input-23-08574acbf5a7>", line 1
(,5)
^
SyntaxError: invalid syntax
A key difference between numpy and MATLAB is that arrays can have any number of dimensions (upto 32). MATLAB has a lower boundary of 2.
The result is that a 5 element numpy array can have shapes (5,), (1,5), (5,1), (1,5,1)`, etc.
The handling of a 1d weight array in your example is best explained the np.dot documentation. Describing it as inner product seems clear enough to me. But I'm also happy with the
sum product over the last axis of a and the second-to-last axis of b
description, adjusted for the case where b has only one axis.
(5,) with (5,n) => (n,) # 5 is the common dimension
(n,5) with (5,) => (n,)
(n,5) with (5,1) => (n,1)
In:
(x1,...,xn' * (R1,...,Rn)
are you missing a )?
(x1,...,xn)' * (R1,...,Rn)
And the * means matrix product? Not elementwise product (.* in MATLAB)? (R1,...,Rn) would have size (n,1). (x1,...,xn)' size (1,n). The product (1,1).
By the way, that raises another difference. MATLAB expands dimensions to the right (n,1,1...). numpy expands them to the left (1,1,n) (if needed by broadcasting). The initial dimensions are the outermost ones. That's not as critical a difference as the lower size 2 boundary, but shouldn't be ignored.
My goal is to to turn a row vector into a column vector and vice versa. The documentation for numpy.ndarray.transpose says:
For a 1-D array, this has no effect. (To change between column and row vectors, first cast the 1-D array into a matrix object.)
However, when I try this:
my_array = np.array([1,2,3])
my_array_T = np.transpose(np.matrix(myArray))
I do get the wanted result, albeit in matrix form (matrix([[66],[640],[44]])), but I also get this warning:
PendingDeprecationWarning: the matrix subclass is not the recommended way to represent matrices or deal with linear algebra (see https://docs.scipy.org/doc/numpy/user/numpy-for-matlab-users.html). Please adjust your code to use regular ndarray.
my_array_T = np.transpose(np.matrix(my_array))
How can I properly transpose an ndarray then?
A 1D array is itself once transposed, contrary to Matlab where a 1D array doesn't exist and is at least 2D.
What you want is to reshape it:
my_array.reshape(-1, 1)
Or:
my_array.reshape(1, -1)
Depending on what kind of vector you want (column or row vector).
The -1 is a broadcast-like, using all possible elements, and the 1 creates the second required dimension.
If your array is my_array and you want to convert it to a column vector you can do:
my_array.reshape(-1, 1)
For a row vector you can use
my_array.reshape(1, -1)
Both of these can also be transposed and that would work as expected.
IIUC, use reshape
my_array.reshape(my_array.size, -1)
I have two np.ndarrays, data with shape (8000, 500) and sample with shape (1, 500).
What I am trying to achieve is measure various types of metrics between every row in data to sample.
When using from sklearn.metrics.pairwise.cosine_distances I was able to take advantage of numpy's broadcasting executing the following line
x = cosine_distances(data, sample)
But when I tried to use the same procedure with scipy.spatial.distance.cosine I got the error
ValueError: Input vector should be 1-D.
I guess this is a broadcasting issue and I'm trying to find a way to get around it.
My ultimate goal is to iterate over all of the distances available in scipy.spatial.distance that can accept two vectors and apply them to the data and the sample.
How can I replicate the broadcasting that automatically happens in sklearn's in my scipy version of the code?
OK, looking at the docs, http://scikit-learn.org/stable/modules/generated/sklearn.metrics.pairwise.cosine_distances.html
With (800,500) and (1,500) inputs ((samples, features)), you should get back a (800,1) result ((samples1, samples2)).
I wouldn't describe that as broadcasting. It's more like dot product, that performs some sort calculation (norm) over features (the 500 shape), reducing that down to one value. It's more like np.dot(data, sample.T) in its handling of dimensions.
https://docs.scipy.org/doc/scipy/reference/generated/scipy.spatial.distance.cosine.html is Computes the Cosine distance between 1-D arrays, more like
for row in data:
for s in sample:
d = cosine(row, s)
or since sample has only one row
distances = np.array([cosine(row, sample[0]) for row in data])
In other words, the sklearn version does the pairwise iteration (maybe in compiled code), while the spartial just evaluates the distance for one pair.
pairwise.cosine_similarity does
# K(X, Y) = <X, Y> / (||X||*||Y||)
K = safe_sparse_dot(X_normalized, Y_normalized.T, dense_output=dense_output)
That's the dot like behavior that I mentioned earlier, but with the normalization added.
I have a large scipy.sparse.csc_matrix and would like to normalize it. That is subtract the column mean from each element and divide by the column standard deviation (std)i.
scipy.sparse.csc_matrix has a .mean() but is there an efficient way to compute the variance or std?
You can calculate the variance yourself using the mean, with the following formula:
E[X^2] - (E[X])^2
E[X] stands for the mean. So to calculate E[X^2] you would have to square the csc_matrix and then use the mean function. To get (E[X])^2 you simply need to square the result of the mean function obtained using the normal input.
Sicco has the better answer.
However, another way is to convert the sparse matrix to a dense numpy array one column at a time (to keep the memory requirements lower compared to converting the whole matrix at once):
# mat is the sparse matrix
# Get the number of columns
cols = mat.shape[1]
arr = np.empty(shape=cols)
for i in range(cols):
arr[i] = np.var(mat[:, i].toarray())
The most efficient way I know of is to use StandardScalar from scikit:
from sklearn.preprocessing import StandardScaler
scalar = StandardScaler(with_mean=False)
scalar.fit(X)
Then the variances are in the attribute var_:
X_var = scalar.var_
The curious thing though, is that when I densified first using pandas (which is very slow) my answer was off by a few percent. I don't know which is more accurate.
The efficient way is actually to densify the entire matrix, then standardize it in the usual way with
X = X.toarray()
X -= X.mean()
X /= X.std()
As #Sebastian has noted in his comments, standardizing destroys the sparsity structure (introduces lots of non-zero elements) in the subtraction step, so there's no use keeping the matrix in a sparse format.