np-xarr 0.1.3

Perform a numpy array transformation by giving examples.

np-xarr

Perform a numpy array transformation by giving examples.

Install

$ pip install np-xarr

Usage

>> from npxarr import X
>> import numpy as np
>> a = X('[1, 2, ...]', '[[1], [2], ...]') # [1, 2, ...] -> [[1], [2], ...]
>> a # print the equation from output index (x0, x1, ...) to input index (y0, y1, ...)
y0 = |_x0_| + |_x1_| # |_x_| means floor(x)
>> a(np.arange(6)) # apply the convertion to a certain numpy array
[[0], [1], [2], [3], [4], [5]]
>> a(np.arange(6).reshape(3, -1)) # 1, 2, in [1, 2, ...] can also represent a smaller array
[[[0 1]], [[2 3]], [[4 5]]]
>> a = X('[[1, 2, ...], [3, 4, ...], ...]', '[[1, 3, ...], [2, 4, ...], ...]') # transpose
>> a
y0 = |_x1_|, y1 = |_x0_|
>> a(np.arange(6).reshape(3, -1))
[[0 2 4]
 [1 3 5]]

Multiple inputs or outputs are supported.

>> a = X(['[[1], [2], ...]', '[a, b, ...]'],  # multiple input in a list
         '[[[1], [a]], [[2], [b]], ...]; [a, 1, b, 2, ...]') # or seperate by ;
>> a
out0: s = mod(|_x1_| + |_x2_|, 2) # s decides which input array to use
[
  in0: y0 = |_x0_| + |_x1_| + |_x2_|, y1 = |_x1_| + |_x2_| # (x0, x1, x2) -> (y0, y1)
  in1: y0 = |_x0_| + |_x1_| + |_x2_| -1
]
out1: s = mod(-|_mod(x0, 2)_| + 1, 2)
[
  in0: y0 = |_0.50*x0_|, y1 = 0
  in1: y0 = |_0.50*x0_|
]
>> a([np.array([[i] for i in range(6)]), -np.arange(10)]) # incompatible input shape here
(array([[[ 0], [ 0]],
       [[ 1], [-1]],
       [[ 2], [-2]],
       [[ 3], [-3]],
       [[ 4], [-4]],
       [[ 5], [-5]]]),
 array([ 0,  0, -1,  1, -2,  2, -3,  3, -4,  4, -5,  5, -6]))
# it will figure out the maximum valid output shape
>> a[1] # or just get the transformation for second output
s = mod(-|_mod(x0, 2)_| + 1, 2)
[
  in0: y0 = mod(|_0.50*x0_|, 6), y1 = 0
  in1: y0 = mod(|_0.50*x0_|, 10)
]
>> a[1]([np.array([[i] for i in range(6)]), -np.arange(10)])
[ 0,  0, -1,  1, -2,  2, -3,  3, -4,  4, -5,  5, -6]

Functions can be applied.

>> a = X('[1, 2, 3, 4, ...]', '[[times(2)], [neg(1)], [times(4)], [neg(3)], ...]', 
         f={'neg': lambda x: -x, 'times': lambda x: 10*x})

notice here the output with sequence [2, 1, 4, 3, ...]

>> a
y0 = |_x0_| + |_x1_| - |_2*mod(x0, 2)_| + 1
funcs: 
[
  times: mod(|_x1_| - |_mod(x0, 2)_| + 1, 2) = 1
  neg: mod(|_x1_| + |_mod(x0, 2)_|, 2) = 1
]
>> a(np.arange(6))
[[10], [ 0], [30], [-2], [50], [-4]]

You can provide output shape by hand

>> a = X('[1, 2, ...]', '[[1, 1, ...], [2, 2, ...], ...]')
>> a
y0 = |_x0_|
>> a(np.arange(6), outShapes=(-1, 3)) # or outShapes=[(-1, 3)], 
[[0 0 0]
 [1 1 1]
 [2 2 2]
 [3 3 3]
 [4 4 4]
 [5 5 5]]

And by providing parameter extraShapes...

>> a = X('[1, 2, 3, ...]', '[[1, 2], [2, 3], ...]')
>> a(np.arange(4), extraShapes=(1, 0)))
[[0 1]
 [1 2]
 [2 3]
 [3 0]]

Notes:

• Inefficient for large array

  The output array is built by code like np.array([inArrays[indexConverter(index)] for index <= outShape])

• Only support transformaton with formula $y_j = floor(a_ij*x_i) + b_j + floor(c_ij*mod(x_i, d_ij))$

Todo

• Test cover
• Support for illegal python variable name like a.1
• Try to deduce possible transformation using naive numpy function from calculated equation