比较与tf.image.ssim的功能差异

tf.image.ssim

tf.image.ssim(
    img1,
    img2,
    max_val,
    filter_size=11,
    filter_sigma=1.5,
    k1=0.01,
    k2=0.03
) -> Tensor

更多内容详见tf.image.ssim。

mindspore.nn.SSIM

class mindspore.nn.SSIM(
    max_val=1.0,
    filter_size=11,
    filter_sigma=1.5,
    k1=0.01,
    k2=0.03
)(img1, img2) -> Tensor

更多内容详见mindspore.nn.SSIM。

差异对比

TensorFlow：在TensorFlow中，算子是函数式的，直接接收输入的两张图片并返回结果。max_val参数不可缺省，接受图片的格式为”NHWC”。

MindSpore：在MindSpore中，算子需要先实例化，然后接收输入返回结果。接受图片的格式为”NCHW”。

分类	子类	TensorFlow	MindSpore	差异
参数	参数1	img1	img1	MindSpore在实例化的函数中接收该输入，功能上一致
	参数2	img2	img2	MindSpore在实例化的函数中接收该输入，功能上一致
	参数3	max_val	max_val	TensorFlow中此参数是必须的，MindSpore中此参数可缺省，默认值为1.0
	参数4	filter_size	filter_size	-
	参数5	filter_sigma	filter_sigma	-
	参数6	k1	k1	-
	参数7	k2	k2	-

代码示例1

TensorFlow中参数max_val是必须的，MindSpore中可以缺省，默认值为1.0。

# TensorFlow
import numpy as np
import tensorflow as tf

img1 = tf.ones([1, 16, 16, 3])
img2 = tf.ones([1, 16, 16, 3])
output = tf.image.ssim(img1, img2, max_val=1.0)
print(output.numpy())
# [1.]

# MindSpore
import numpy as np
import mindspore.nn as nn
from mindspore import Tensor

net = nn.SSIM()
img1 = Tensor(np.ones([1, 3, 16, 16]).astype(np.float32))
img2 = Tensor(np.ones([1, 3, 16, 16]).astype(np.float32))
output = net(img1, img2)
print(output)
# [1.]

代码示例2

TensorFlow中接受图片输入的格式为“NHWC”，MindSpore中为”NCHW”。下面的例子用同一个随机种子生成的4D张量，在MindSpore中经(0, 3, 1, 2)的轴变换后可以得到和TensorFlow相同的结果。

# TensorFlow
import numpy as np
import tensorflow as tf

np.random.seed(10)
img1 = np.random.randint(0, 2, (2, 5, 5, 5)).astype(np.float32)
img2 = np.random.randint(0, 2, (2, 5, 5, 5)).astype(np.float32)
output = tf.image.ssim(img1, img2, max_val=1, filter_size=3)
print(output.numpy())
# [-0.00746754 -0.09539266]

# MindSpore
import numpy as np
import mindspore
from mindspore import nn, ops, Tensor

np.random.seed(10)
img1 = np.random.randint(0, 2, (2, 5, 5, 5)).astype(np.float32)
img2 = np.random.randint(0, 2, (2, 5, 5, 5)).astype(np.float32)
img1_t = Tensor(img1)
img2_t = Tensor(img2)
net = nn.SSIM(filter_size=3)
transpose = ops.Transpose()
trans_term = (0, 3, 1, 2)
img1_trans = transpose(img1_t, trans_term)
img2_trans = transpose(img2_t, trans_term)
output_m = net(img1_trans, img2_trans)
print(output_m)
# [-0.00746753 -0.09539266]