CODEWITHSUNDEEP
pythonpandasnumpy
Rob
Rob

Reputation: 478

Multiplying leading diagonal values in Pandas?

I am trying to multiply the leading diagonal in a pandas dataframe and I am not sure how to proceed in a computationally reasonable way.

df =       [ 3  4  5
             6  7  8
             9 10 11]

ouput_df = [231 32 5
            60  77 8
             9  10 11]

Explanation - lookoing to 3 * 7 * 11 for the first element, 4 * 8 for the second element, 7 * 11 for the fifth element etc.,

Note: The matrix I am working on is not a square matrix, but a rectangular matrix.

Upvotes: 3

Views: 354

Answers (4)

brentertainer
brentertainer

Reputation: 2188

Here's a method that operates on the DataFrame in place.

df = pd.DataFrame(data=[[3, 4, 5], [6, 7, 8], [9, 10, 11]])
m, n = df.shape
for i in range(-m + 1, n):
    ri, rj = max(-i, 0), min(m - 1, n - i - 1)
    ci, cj = max( i, 0), min(n - 1, m + i - 1)
    np.fill_diagonal(df.values[ri:rj+1,ci:cj+1],
                     df.values.diagonal(i)[::-1].cumprod()[::-1])
print(df)

Result:

     0   1   2
0  231  32   5
1   60  77   8
2    9  10  11

Upvotes: -1

Divakar
Divakar

Reputation: 221524

Here's one based on NumPy -

def cumprod_upper_diag(a):
    m,n = a.shape
    mask = ~np.tri(m,n, dtype=bool)
    p = np.ones((m,n),dtype=a.dtype)
    p[mask[:,::-1]] = a[mask]
    a[mask] = p[::-1].cumprod(0)[::-1][mask[:,::-1]]
    return a

a = df.to_numpy(copy=False) # For older versions : a = df.values
out = a.copy()
cumprod_upper_diag(out)
cumprod_upper_diag(out.T)
out.ravel()[::a.shape[1]+1] = out.ravel()[::out.shape[1]+1][::-1].cumprod()[::-1]
out_df = pd.DataFrame(out)

Upvotes: 3

user3483203
user3483203

Reputation: 51165

You can use a sparse diagonal matrix here with some finnicking. This assumes all non-zero elements in your original matrix, or else this will not work.

from scipy import sparse

a = df.to_numpy()
b = sparse.dia_matrix(a)
c = b.data[:, ::-1]
cp = np.cumprod(np.where(c != 0, c, 1), axis=1)
b.data = cp[:, ::-1]
b.A


array([[231,  32,   5],
       [ 60,  77,   8],
       [  9,  10,  11]], dtype=int64)

Upvotes: 2

Quang Hoang
Quang Hoang

Reputation: 150735

As Chris mentioned, this is cumprod in reverse order:

# stack for groupby
new_df = df.stack().reset_index()[::-1]

# diagonals meaning col_num - row_num are the same
diags = new_df['level_0']-new_df['level_1']

# groupby diagonals
new_df['out'] = new_df.groupby(diags)[0].cumprod()

# pivot to get the original shape
new_df.pivot('level_0', 'level_1', 'out')

output:

level_1    0   1   2
level_0             
0        231  32   5
1         60  77   8
2          9  10  11

Upvotes: 1

Related Questions