Function Differences with tf.keras.metrics.CosineSimilarity

tf.keras.metrics.CosineSimilarity

tf.keras.metrics.CosineSimilarity(
    name='cosine_similarity', dtype=None, axis=-1
)

For more information, see tf.keras.metrics.CosineSimilarity.

mindspore.train.CosineSimilarity

mindspore.train.CosineSimilarity(similarity="cosine", reduction="none", zero_diagonal=True)

For more information, see mindspore.train.CosineSimilarity.

Usage

MindSpore: The input is a matrix, each row of the matrix can be regarded as a sample, and the return value is the similarity matrix. If similarity="cosine", it is cosine similarity calculation logic, same as tf.keras.metrics.CosineSimilarity calculation logic, and if similarity="dot", it is matrix dot product transpose matrix. reduction can be set to none, sum, mean, which correspond to the original result matrix, sum and average calculation respectively.

TensorFlow: The inputs are the predicted and true values, which are computed by cosine similarity = (a . b) / ||a|| ||b|| is computed and the return result is the mean value of cosine similarity for all data streams.

Code Example

import tensorflow as tf
tf.enable_eager_execution()

m = tf.keras.metrics.CosineSimilarity(axis=1)
m.update_state([[1, 3, 4]], [[2, 4, 2]])
print(m.result().numpy())

# output: 0.8807048


from mindspore.train import CosineSimilarity
import numpy as np

input_data = np.array([[1, 3, 4], [2, 4, 2], [0, 1, 0]])
metric = CosineSimilarity()
metric.update(input_data)
print(metric.eval())

# output:
# [[0.         0.88070485 0.58834841]
#  [0.88070485 0.         0.81649658]
#  [0.58834841 0.81649658 0.        ]]