numpy - 如何组合多个索引（将多个一对一矩阵访问替换为一次访问）

Question

Update

The implementation did not consider multiple occurrences of a same word, and self word occurrences.

For instance when stride=2 and the word at the position is W, co-occurrence of X needs +2, self-co-occurrence of W needs +1.

X|Y|W|X|W

Question

To update the m * m matrix ( co_occurance_matrix ), currently accessing row by row with a loop. The entire code is at the bottom.

How can I remove the loop and update the multiple rows all at once? I believe there should be a way to combine each index into one matrix that replaces the loop with one vectorized update.

Please advice possible approaches.

Current implementation

for position in range(0, n):       
    co_ccurrence_matrix[
        sequence[position],                                                # position  to the word
        sequence[max(0, position-stride) : min((position+stride),n-1) +1]  # positions to co-occurrence words
    ] += 1

1. Loop over an array of word indices sequence (word index is an integer code for each word).
2. For each word at the position in the loop, check the co-occurring words on both sides within a stride distance.
   This is a N-gram context window as in the purple box in the diagram . N = context_size = stride*2 + 1 .
3. Increment the count of each co-occurrence word in the co_occurrence_matrix as per blue lines in the diagram .

Attempts

It seems the Integer array indexing may be a way to access multiple rows at the same time.

x = np.array([[ 0,  1,  2],
              [ 3,  4,  5],
              [ 6,  7,  8],
              [ 9, 10, 11]])
rows = np.array([[0, 0],
                 [3, 3]], dtype=np.intp)
columns = np.array([[0, 2],
                    [0, 2]], dtype=np.intp)
x[rows, columns]
---
array([[ 0,  2],
       [ 9, 11]])

Create a multi-dimensional indices by combining each index in the loop, but it does not work with the error. Please advise the cause and the mistakes, or if the attempt does not make sense.

    indices = np.array([
        [
            sequence[0],                                         # position  to the word
            sequence[max(0, 0-stride) : min((0+stride),n-1) +1]  # positions to co-occurrence words
        ]]
    )
    assert n > 1
    for position in range(1, n):
        co_occurrence_indices = np.array([
            [
                sequence[position],                                                # position  to the word
                sequence[max(0, position-stride) : min((position+stride),n-1) +1]  # positions to co-occurrence words
            ]]
        )
        indices = np.append(
            indices,
            co_occurrence_indices,
            axis=0
        )

    print("Updating the co_occurrence_matrix: indices \n{} \nindices.dtype {}".format(
        indices,
        indices.dtype
    ))
    co_ccurrence_matrix[  
        indices              <---- Error
    ] += 1
 

Output

Updating the co_occurrence_matrix: indices 
[[0 array([0, 1])]
 [1 array([0, 1, 2])]
 [2 array([1, 2, 3])]
 [3 array([2, 3, 0])]
 [0 array([3, 0, 1])]
 [1 array([0, 1, 4])]
 [4 array([1, 4, 5])]
 [5 array([4, 5, 6])]
 [6 array([5, 6, 7])]
 [7 array([6, 7])]] 
indices.dtype object

<ipython-input-88-d9b081bf2f1a>:48: VisibleDeprecationWarning: Creating an ndarray from ragged nested sequences (which is a list-or-tuple of lists-or-tuples-or ndarrays with different lengths or shapes) is deprecated. If you meant to do this, you must specify 'dtype=object' when creating the ndarray
  indices = np.array([
<ipython-input-88-d9b081bf2f1a>:56: VisibleDeprecationWarning: Creating an ndarray from ragged nested sequences (which is a list-or-tuple of lists-or-tuples-or ndarrays with different lengths or shapes) is deprecated. If you meant to do this, you must specify 'dtype=object' when creating the ndarray
  co_occurrence_indices = np.array([

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-88-d9b081bf2f1a> in <module>
     84 sequence, word_to_id, id_to_word = preprocess(corpus)
     85 vocabrary_size = max(word_to_id.values()) + 1
---> 86 create_cooccurrence_matrix(sequence, vocabrary_size , 3)

<ipython-input-88-d9b081bf2f1a> in create_cooccurrence_matrix(sequence, vocabrary_size, context_size)
     70         indices.dtype
     71     ))
---> 72     co_ccurrence_matrix[
     73         indices
     74     ] += 1

IndexError: arrays used as indices must be of integer (or boolean) type

---

Current code

import numpy as np
 
def preprocess(text):
    """
    Args:
        text: A string including sentences to process. corpus
    Returns:
        sequence:
            A numpy array of word indices to every word in the original text as they appear in the text.
            The objective of corpus is to preserve the original text but as numerical indices.
        word_to_id: A dictionary to map a word to a word index
        id_to_word: A dictionary to map a word index to a word
    """
    text = text.lower()
    text = text.replace('.', ' .')
    words = text.split(' ')
 
    word_to_id = {}
    id_to_word = {}
    for word in words:
        if word not in word_to_id:
            new_id = len(word_to_id)
            word_to_id[word] = new_id
            id_to_word[new_id] = word
 
    sequence= np.array([word_to_id[w] for w in words])
 
    return sequence, word_to_id, id_to_word
 
 
def create_cooccurrence_matrix(sequence, vocabrary_size, context_size=3):
    """
    Args:
        sequence: word index sequence of the original corpus text
        vocabrary_size: number of words in the vocabrary (same with co-occurrence vector size)
        context_size: context (N-gram size N) within which to check co-occurrences.         
    """
    n = sequence_size = len(sequence)
    co_ccurrence_matrix = np.zeros((vocabrary_size, vocabrary_size), dtype=np.int32)
 
    stride = int((context_size - 1)/2 )
    assert(n > stride), "sequence_size {} is less than/equal to stride {}".format(
        n, stride
    )
 
    for position in range(0, n):       
        co_ccurrence_matrix[
            sequence[position],                                                # position  to the word
            sequence[max(0, position-stride) : min((position+stride),n-1) +1]  # positions to co-occurrence words
        ] += 1
 
    np.fill_diagonal(co_ccurrence_matrix, 0)
    return co_ccurrence_matrix
 
 
corpus= "To be, or not to be, that is the question"
 
sequence, word_to_id, id_to_word = preprocess(corpus)
vocabrary_size = max(word_to_id.values()) + 1
create_cooccurrence_matrix(sequence, vocabrary_size , 3)
---
[[0 2 0 1 0 0 0 0]
 [2 0 1 0 1 0 0 0]
 [0 1 0 1 0 0 0 0]
 [1 0 1 0 0 0 0 0]
 [0 1 0 0 0 1 0 0]
 [0 0 0 0 1 0 1 0]
 [0 0 0 0 0 1 0 1]
 [0 0 0 0 0 0 1 0]]

Profiling

Used ptb.train.txt from enter link description here .

Timer unit: 1e-06 s

Total time: 23.0015 s
File: <ipython-input-8-27f5e530d4ff>
Function: create_cooccurrence_matrix at line 1

Line #      Hits         Time  Per Hit   % Time  Line Contents
==============================================================
     1                                           def create_cooccurrence_matrix(sequence, vocabrary_size, context_size=3):
     2                                               """
     3                                               Args: 
     4                                                   sequence: word index sequence of the original corpus text
     5                                                   vocabrary_size: number of words in the vocabrary (same with co-occurrence vector size)
     6                                                   context_size: context (N-gram size N) within to check co-occurrences.
     7                                               Returns:
     8                                                   co_occurrence matrix
     9                                               """
    10         1          4.0      4.0      0.0      n = sequence_size = len(sequence)
    11         1         98.0     98.0      0.0      co_occurrence_matrix = np.zeros((vocabrary_size, vocabrary_size), dtype=np.int32)
    12                                           
    13         1          5.0      5.0      0.0      stride = int((context_size - 1)/2 )
    14         1          1.0      1.0      0.0      assert(n > stride), "sequence_size {} is less than/equal to stride {}".format(
    15                                                   n, stride
    16                                               )
    17                                           
    18                                               """
    19                                               # Handle position=slice(0 : (stride-1) +1),       co-occurrences=slice(max(0, position-stride): min((position+stride),n-1) +1)
    20                                               # Handle position=slice((n-1-stride) : (n-1) +1), co-occurrences=slice(max(0, position-stride): min((position+stride),n-1) +1)
    21                                               indices = [*range(0, (stride-1) +1), *range((n-1)-stride +1, (n-1) +1)]
    22                                               #print(indices)
    23                                               
    24                                               for position in indices:
    25                                                   debug(sequence, position, stride, False)
    26                                                   co_occurrence_matrix[
    27                                                       sequence[position],                                             # position to the word
    28                                                       sequence[max(0, position-stride) : min((position+stride),n-1) +1]  # indices to co-occurance words 
    29                                                   ] += 1
    30                                           
    31                                               
    32                                               # Handle position=slice(stride, ((sequence_size-1) - stride) +1)
    33                                               for position in range(stride, (sequence_size-1) - stride + 1):        
    34                                                   co_occurrence_matrix[
    35                                                       sequence[position],                                 # position to the word
    36                                                       sequence[(position-stride) : (position + stride + 1)]  # indices to co-occurance words 
    37                                                   ] += 1
    38                                               """        
    39                                               
    40    929590    1175326.0      1.3      5.1      for position in range(0, n):        
    41   2788767   15304643.0      5.5     66.5          co_occurrence_matrix[
    42   1859178    2176964.0      1.2      9.5              sequence[position],                                                # position  to the word
    43    929589    3280181.0      3.5     14.3              sequence[max(0, position-stride) : min((position+stride),n-1) +1]  # positions to co-occurance words 
    44    929589    1062613.0      1.1      4.6          ] += 1
    45                                           
    46         1       1698.0   1698.0      0.0      np.fill_diagonal(co_occurrence_matrix, 0)
    47                                               
    48         1          2.0      2.0      0.0      return co_occurrence_matrix

Answer 1

EDIT: You could do this using inbuilt sklearn functions extremely easily, but seeing the history of your questions, I believe you are looking for a pure NumPy vectorized implementation.

---

IIUC, you want to create a co-occurrence matrix based on the context window around a word. So, if there are 12 words in a vocabulary, 100 sentences, and say a context size of 2, then you want to look at the rolling windows of 5 (2 left, 1 center, 2 right) size in each of the sentences and iteratively (or vectorized) add the context words to get a (12, 12) matrix which tells you how many times a word occurs in the context window of another word .

Vectorized implementation

You can do this in a completely vectorized manner as such (explanation in last section) -

#Definitions
sentences, vocab, length, context_size = 100, 12, 15, 2

#Create dummy corpus (label encoded)
window = context_size*2+1
corpus = np.random.randint(0, vocab, (sentences, length))  #(100, 15)

#Create rolling window view of the sequences
shape = corpus.shape[0], corpus.shape[1]-window+1, window  #(100, 11, 5) 
stride = corpus.strides[0], corpus.strides[1], corpus.strides[1]  #(120, 8, 8)
rolling_window = np.lib.stride_tricks.as_strided(corpus, shape=shape, strides=stride)  #(100, 11, 5)

#Creating co-occurence matrix based on context window
center_idx = context_size
#position = rolling_window[:,:,context_size]  #(100, 11)
context = np.delete(rolling_window, center_idx, -1)  #(100, 11, 4)
context_multihot = np.sum(np.eye(vocab)[context], axis=-2)  #(100, 11, 12)
cooccurence = np.tensordot(context_multihot.transpose(0,2,1), context_multihot, axes=([0,2],[0,1]))  #(12, 12)
np.fill_diagonal(cooccurence,0)  #(12, 12)
print(cooccurence)

[[  0.  94. 100. 114.  91.  92.  90. 128. 100. 114.  91.  84.]
 [ 94.   0.  78.  96.  90.  65.  76.  68.  76. 108.  58.  68.]
 [100.  78.   0. 125. 107.  93.  83.  84.  73.  84.  97. 110.]
 [114.  96. 125.   0.  84.  97.  76. 110.  80.  94. 117.  97.]
 [ 91.  90. 107.  84.   0.  84.  87. 103.  60. 127. 123.  97.]
 [ 92.  65.  93.  97.  84.   0.  67.  87.  72.  87.  74.  92.]
 [ 90.  76.  83.  76.  87.  67.   0.  83.  73. 118.  81. 108.]
 [128.  68.  84. 110. 103.  87.  83.   0.  72. 100. 115.  69.]
 [100.  76.  73.  80.  60.  72.  73.  72.   0.  83.  81. 100.]
 [114. 108.  84.  94. 127.  87. 118. 100.  83.   0. 109. 110.]
 [ 91.  58.  97. 117. 123.  74.  81. 115.  81. 109.   0. 104.]
 [ 84.  68. 110.  97.  97.  92. 108.  69. 100. 110. 104.   0.]]

---

Testing on example given

Let's test this on a single sentence corpus to be or not to be that is the question

sentence = 'to be or not to be that is the question'
corpus = np.array([[0, 1, 2, 3, 0, 1, 4, 5, 6, 7]])

#Definitions
vocab, context_size = 8, 2
window = context_size*2+1

#Create rolling window view of the sequences
shape = corpus.shape[0], corpus.shape[1]-window+1, window
stride = corpus.strides[0], corpus.strides[1], corpus.strides[1]
rolling_window = np.lib.stride_tricks.as_strided(corpus, shape=shape, strides=stride)

#Creating co-occurence matrix based on context window
center_idx = context_size
#position = rolling_window[:,:,context_size]  
context = np.delete(rolling_window, center_idx, -1)  
context_multihot = np.sum(np.eye(vocab)[context], axis=-2)  
cooccurence = np.tensordot(context_multihot.transpose(0,2,1), context_multihot, axes=([0,2],[0,1]))
np.fill_diagonal(cooccurence,0)
print(cooccurence)

[[0. 5. 1. 3. 1. 2. 1. 0.]
 [5. 0. 3. 2. 2. 1. 2. 1.]
 [1. 3. 0. 1. 1. 0. 0. 0.]
 [3. 2. 1. 0. 2. 1. 0. 0.]
 [1. 2. 1. 2. 0. 1. 1. 1.]
 [2. 1. 0. 1. 1. 0. 1. 0.]
 [1. 2. 0. 0. 1. 1. 0. 1.]
 [0. 1. 0. 0. 1. 0. 1. 0.]]

---

Detailed Explanation

Let's start with creating some label encoded dummy data. Here there are 100 sentences, with a vocab of 12 size. The length of each sentence is 15 and the window that I am taking is 5 (2+1+2) -

sentences, vocab, length, context_size = 100, 12, 15, 2
window = context_size*2+1
corpus = np.random.randint(0, vocab, (sentences, length))
corpus[0:2]

#top 2 sentences
array([[ 9,  8,  9,  4,  2, 10,  9,  0,  7,  1, 11,  0,  7,  3,  1],
       [ 7,  9,  4,  0,  1,  9, 10,  7,  4,  2,  2,  3,  5,  8,  8]])

Next, we want to create rolling window views of the window size so that we can then get to our next stages. The shape of this new view would be equal to (sentences, number of windows, window size) and so using stride_tricks we can create a rolling window view of this matrix quite easily.

#Create shape and stride definitions
shape = corpus.shape[0], corpus.shape[1]-window+1, window
stride = corpus.strides[0], corpus.strides[1], corpus.strides[1]
print(shape, stride)

#create view
rolling_window = np.lib.stride_tricks.as_strided(corpus, shape=shape, strides=stride)  #(100, 11, 5)
print('\nView for first sequence ->')
print(rolling_window[0])

(100, 11, 5) (120, 8, 8)

View for first sequence ->
[[ 9  8  9  4  2]
 [ 8  9  4  2 10]
 [ 9  4  2 10  9]
 [ 4  2 10  9  0]
 [ 2 10  9  0  7]
 [10  9  0  7  1]
 [ 9  0  7  1 11]
 [ 0  7  1 11  0]
 [ 7  1 11  0  7]
 [ 1 11  0  7  3]
 [11  0  7  3  1]]

Next let's look at only a single sentence first and get that into a co-occurance matrix. After that we can scale it to a higher dimension matrix.

For a SINGLE SENTENCE, we can do the following steps -

1. Get the position word (center word)
2. Get the context words by deleting the center word column
3. Create a one-hot matrix using np.eye(vocab) and filtering for the context labels
4. Sum this over last axis to get multi-hot matrix for each window
5. Take a dot product to get a (word, word) co-occurence matrix from the multi-hot context vectors for each window.
6. Fill diagonal with 0 to ignore occurrence of the same word with itself

position = rolling_window[0][:,2]
context = np.delete(rolling_window[0], 2, 1)
context_multihot = np.sum(np.eye(vocab)[context], axis=1)
cooccurence = context_multihot.T@context_multihot
np.fill_diagonal(cooccurence,0)
print(cooccurence)

[[0. 3. 2. 1. 1. 0. 0. 5. 0. 2. 1. 4.]
 [3. 0. 0. 2. 0. 0. 0. 4. 0. 2. 1. 3.]
 [2. 0. 0. 0. 2. 0. 0. 1. 2. 3. 2. 0.]
 [1. 2. 0. 0. 0. 0. 0. 1. 0. 0. 0. 2.]
 [1. 0. 2. 0. 0. 0. 0. 0. 1. 4. 1. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [5. 4. 1. 1. 0. 0. 0. 0. 0. 1. 2. 2.]
 [0. 0. 2. 0. 1. 0. 0. 0. 0. 2. 1. 0.]
 [2. 2. 3. 0. 4. 0. 0. 1. 2. 0. 4. 1.]
 [1. 1. 2. 0. 1. 0. 0. 2. 1. 4. 0. 0.]
 [4. 3. 0. 2. 0. 0. 0. 2. 0. 1. 0. 0.]]

We have now been able to do the whole thing for 1 sentence. Now we just have to scale to 100 sentences without for loops. For this only a few things need to change.

1. Create dynamic indexes for position and context words (previously hard-coded)
2. Handle axis since now we are dealing with 3D tensor instead of 2D
3. Take transpose of context_multihot over last 2 axis before dot product
4. Change the np.dot to np.tensordot so that we can reduce specified axis. In this case, we have to perform (100, 12, 11) @ (100, 11, 12) -> (12, 12) . So select axis accordingly.

#Creating co-occurence matrix based on context window
center_idx = context_size
#position = rolling_window[:,:,context_size]  #(100, 11)
context = np.delete(rolling_window, center_idx, -1)  #(100, 11, 4)
context_multihot = np.sum(np.eye(vocab)[context], axis=-2)  #(100, 11, 12)
cooccurence = np.tensordot(context_multihot.transpose(0,2,1), context_multihot, axes=([0,2],[0,1]))  #(12, 12)
np.fill_diagonal(cooccurence,0)  #(12, 12)
print(cooccurence)

[[  0.  94. 100. 114.  91.  92.  90. 128. 100. 114.  91.  84.]
 [ 94.   0.  78.  96.  90.  65.  76.  68.  76. 108.  58.  68.]
 [100.  78.   0. 125. 107.  93.  83.  84.  73.  84.  97. 110.]
 [114.  96. 125.   0.  84.  97.  76. 110.  80.  94. 117.  97.]
 [ 91.  90. 107.  84.   0.  84.  87. 103.  60. 127. 123.  97.]
 [ 92.  65.  93.  97.  84.   0.  67.  87.  72.  87.  74.  92.]
 [ 90.  76.  83.  76.  87.  67.   0.  83.  73. 118.  81. 108.]
 [128.  68.  84. 110. 103.  87.  83.   0.  72. 100. 115.  69.]
 [100.  76.  73.  80.  60.  72.  73.  72.   0.  83.  81. 100.]
 [114. 108.  84.  94. 127.  87. 118. 100.  83.   0. 109. 110.]
 [ 91.  58.  97. 117. 123.  74.  81. 115.  81. 109.   0. 104.]
 [ 84.  68. 110.  97.  97.  92. 108.  69. 100. 110. 104.   0.]]