I have a numpy array of one-hot vectors. I want to find the mode of these one-hot vectors. Note that this is not equivalent to finding the mode over the values.
e.g. for
x = [[0,0,0,1],
[0,0,0,1],
[0,0,1,0],
[0,1,0,0],
[1,0,0,0]]
assert vector_mode(x) == [0,0,0,1]
assert scipy.stats.mode(x) == [0,0,0,0]
What is the most efficient way to do this with numpy/scipy?
We are dealing with one-hot vectors as rows of the 2D input array. So, argmax of each row would be unique to each one-hot vector. Get those. Then, get their counts. Anyone of the rows with the max argmax count would be the desired mode row output. Let's pick the first off those with one more use of argmax and finally index into 2D input.
Hence, one implementation -
idx = np.argmax(x,1)
count = np.bincount(idx)
out = x[(idx==count.argmax()).argmax()]
If your vectors are one-hot, you can just use argmax to get the index of the hotspot and compute the mode of those:
hot = np.argmax(x, axis=1)
mode = scipy.stats.mode(hot).mode
In this case, mode is 3, meaning that the most common vector has a hotspot in index 3.
If you want to reinstate this into a one-hot vector, you can do:
vec = np.zeros(4)
vec[mode] = 1
scipy.stats.modehas a loop in it that compares every value found in the array to the entire array, which can cause surprisingly bad performance for an array with a lot of distinct values in it. For example,scipy.stats.mode(range(10**5))is appallingly slow.