如何在张量流中对张量的列执行相似度函数

Question

I have a tensor like this:

tf_a1 =      [[-0.65 0.   0.   0.   0.42  0.   0.   0.51 0.   0.34 0.]
              [0.   -0.51 0.   0.  -0.52  0.   0.   0.   0.53 0.42 0.]
              [0.    0.32 0.  -0.50 0.34  0.   0.   0.39 0.32 0.52 0.]
              [0.    0.23 0.37 0.   0.    0.37 0.37 0.   0.47 0.39 0.3 ]]

I want to apply cosine similarity over each column of this tensor. So, I want to find the similarity of the first column versus rest of the columns. Again, second column against rest of the columns and so on.

I have done this using the for loop as such:

def cosine_score(x):
    for i, arr in enumerate(x):
        if i == 0 :
            first = cosine_similarity(x[i,].reshape(1, -1), x)
        else:
            second = cosine_similarity(x[i,].reshape(1, -1), x)
            final = tf.concat((first, second), axis=0)
            first = final
    return final
sim_topics = cosine_score(tf_a1)

Now, When I want to include this in my model, I can not use foor loop as it is. seems I have to use tf.map_fn to go over it.

I also have done like this:

def cosine_score(x):
    def cos_similarity(col):
        for i, arr in enumerate(col):
            if i == 0:
                first = cosine_similarity(col[i, ].reshape(1, -1), col)
            else:
                second = cosine_similarity(col[i, ].reshape(1, -1), col)
                final = tf.concat((first, second), axis=0)
                first = final
        return final
    sim = tf.map_fn(cos_similarity, x, dtype=tf.float32)
    return sim

But here I need to remove the for loop . My problem is that if I remove for loop and access each column seperately, how can I access the rest of the columns to compare and apply cosine similarity .

Please let me know if its not clear.

Answer 1

Cosine similarity is nothing more than an L2 normalized dot product. So, in Tensorflow this should do the trick for you:

# Normalize the columns of the tensor
normalized_tensor = tf.math.l2_normalize(tf_a1, axis=0)
# Get the dot product between the columns
scores = tf.matmul(normalized_tensor, normalized_tensor, transpose_a=True)

The tensor scores contains the cosine similarity between the columns of tf_a1 . Moreover, below is a Numpy equivalent implementation:

# Normalize the columns of the tensor
normalized_tensor = tf_a1 / np.linalg.norm(tf_a1, axis=0)
# Get the dot product between the columns
scores = np.dot(normalized_tensor.T, normalized_tensor)

Finally, if you want to keep only one of the triangles (for example the upper triangle), and set the main diagonal to 0 , you can do the following in Tensorflow :

zero_diag = tf.linalg.set_diag(scores, tf.zeros(tf.shape(scores)[0]))
triangular = tf.matrix_band_part(zero_diag, 0, -1)