0%
May 28, 2022

Masks in Numpy and Retrive Corresponding Indexes

python

Contents

Masks

Suppose that

  • target_masks is of shape (N, width, height, 10, 2),
  • where N denotes the number of batches,
  • width, height denote the size of a feature,
  • and for every given n, j, i, k, target_masks[n, j, i, k] is a one-hot vector. Then we can create a mask by
positive_sample = target_masks[..., 1] == 1

This will create a mask of shape (N, width, height, 10) in which all values are boolean, but how do we know all the indexes n, j, i, k that the mask indicates as True?

Sometimes the mask itself is enough to filter out other related np.array's (by applying the mask like other_nparray[positive_sample]), but sometimes the spartial index j, i also provides us the information we need.

For that we will use the following function:

get_indexes_from_mask(mask)

This function is as simple as:

def get_indexes_from_mask(mask):
    return np.array(list(zip(*np.where(mask))))

For example, suppose that each position j, i was assigned a box in boxes with shape (N, height, width, 4) and a score in scores with shape (N, height, width, 1), then we can create a mask by

score_mask = scores[..., 0] > 0.8

Now apart from getting the desired scores by

  • selected_scores=scores[score_mask] (this will be flattened and of shape (n', 1)), we can equivalently:
  • selected_indexes = get_indexes_from_mask(score_mask)
selected_scores = np.array([scores[tuple(index)]
                            for index in selected_indexes])
    Note that
    • numpy array does not accept int-np-array as an index (unlike bool-np-array),
    • We need to convert int-np-array into int-tuple. The conversion is simply by tuple(np-array).

tf.where, tf.gather (to be updated)

Equivalently we can use tf.where and tf.gather.