CODEWITHSUNDEEP
pythonnumpymultidimensional-arraydata-science
thedlade
thedlade

Reputation: 53

Find indices of duplicated rows in python ndarray

I coded the for loop to enumerate a multidimensional ndarray containing n rows of 28x28 pixel values.

I am looking for the index of each row that is duplicated and the indices of the duplicates without redundancies.

I found this code here (thanks unutbu) and modified it to read the ndarray, it works 70% of the time, however 30% of the time it is identifying the wrong images as duplicates.

How can it be improved to detect the correct rows?

def overlap_same(arr):
seen = []
dups = collections.defaultdict(list)
for i, item in enumerate(arr):
    for j, orig in enumerate(seen):
        if np.array_equal(item, orig):
            dups[j].append(i)
            break
    else:
        seen.append(item)
return dups

e.g. return overlap_same(train) returns:

defaultdict(<type 'list'>, {34: [1388], 35: [1815], 583: [3045], 3208:
[4426], 626: [824], 507: [4438], 188: [338, 431, 540, 757, 765, 806,
808, 834, 882, 1515, 1539, 1715, 1725, 1789, 1841, 2038, 2081, 2165,
2170, 2300, 2455, 2683, 2733, 2957, 3290, 3293, 3311, 3373, 3446, 3542,
3565, 3890, 4110, 4197, 4206, 4364, 4371, 4734, 4851]})

plotting some samples of the correct case on matplotlib gives:

fig = plt.figure()
a=fig.add_subplot(1,2,1)
plt.imshow(train[35])
a.set_title('train[35]')
a=fig.add_subplot(1,2,2)
plt.imshow(train[1815])
a.set_title('train[1815]')
plt.show

train data 35 vs 1815

which is correct

However:

fig = plt.figure()
a=fig.add_subplot(1,2,1)
plt.imshow(train[3208])
a.set_title('train[3208]')
a=fig.add_subplot(1,2,2)
plt.imshow(train[4426])
a.set_title('train[4426]')
plt.show

enter image description here

is incorrect as they do not match

Sample data (train[:3])

array([[[-0.5       , -0.5       , -0.5       , ...,  0.48823529,
      0.5       ,  0.17058824],
    [-0.5       , -0.5       , -0.5       , ...,  0.48823529,
      0.5       , -0.0372549 ],
    [-0.5       , -0.5       , -0.5       , ...,  0.5       ,
      0.47647059, -0.24509804],
    ..., 
    [-0.49215686,  0.34705883,  0.5       , ..., -0.5       ,
     -0.5       , -0.5       ],
    [-0.31176472,  0.44901961,  0.5       , ..., -0.5       ,
     -0.5       , -0.5       ],
    [-0.11176471,  0.5       ,  0.49215686, ..., -0.5       ,
     -0.5       , -0.5       ]],

   [[-0.24509804,  0.2764706 ,  0.5       , ...,  0.5       ,
      0.25294119, -0.36666667],
    [-0.5       , -0.47254902, -0.02941176, ...,  0.20196079,
     -0.46862745, -0.5       ],
    [-0.49215686, -0.5       , -0.5       , ..., -0.47647059,
     -0.5       , -0.49607843],
    ..., 
    [-0.49215686, -0.49607843, -0.5       , ..., -0.5       ,
     -0.5       , -0.49215686],
    [-0.5       , -0.5       , -0.26862746, ...,  0.13137256,
     -0.46470588, -0.5       ],
    [-0.30000001,  0.11960784,  0.48823529, ...,  0.5       ,
      0.28431374, -0.24117647]],

   [[-0.5       , -0.5       , -0.5       , ..., -0.5       ,
     -0.5       , -0.5       ],
    [-0.5       , -0.5       , -0.5       , ..., -0.5       ,
     -0.5       , -0.5       ],
    [-0.5       , -0.5       , -0.5       , ..., -0.5       ,
     -0.5       , -0.5       ],
    ..., 
    [-0.5       , -0.5       , -0.5       , ...,  0.48431373,
      0.5       ,  0.31568629],
    [-0.5       , -0.49215686, -0.5       , ...,  0.49215686,
      0.5       ,  0.04901961],
    [-0.5       , -0.5       , -0.5       , ...,  0.04117647,
     -0.17450981, -0.45686275]]], dtype=float32)

Upvotes: 3

Views: 1208

Answers (1)

Eelco Hoogendoorn
Eelco Hoogendoorn

Reputation: 10769

The numpy_indexed package has a lot of functionality to solve these type of problems efficiently.

For instance, (unlike numpy's builtin unique) this will find your unique images:

import numpy_indexed as npi
unique_training_images = npi.unique(train)

Or if you want to find all the indices of each unique group, you can use:

indices = npi.group_by(train).split(np.arange(len(train)))

Note that these functions do not have quadratic time complexity, like in your original post, and are fully vectorized, and thus in all likelihood a lot more efficient. Also, unlike pandas it does not have a preferred data format, and is fully nd-array capable, so acting on arrays with shape [n_images, 28, 28] 'just works'.

Upvotes: 2

Related Questions