三维矩阵相乘Rosetta不支持吗

[xiaoshui240]

我尝试运行三维tensor相乘，在tensorflow中，只需要第一个维度相同，得到结果是后面两个维度相乘。但Rosetta似乎不支持？ 报错： ValueError: Shape must be rank 2 but is rank 3 for 'RttMatmul' (op: 'RttMatmul') with input shapes: [2,2,3], [2,3,2]. 代码如下：

#!/usr/bin/env python3

# Import rosetta package
import latticex.rosetta as rtt
import tensorflow as tf

# You can activate a backend protocol, here use SecureNN
rtt.activate("SecureNN")

# Get private data from Alice (input x), Bob (input y)
x = tf.Variable(rtt.private_input(0, [[[ 1, 2, 3],[ 4, 5, 6]],[[ 7, 8, 9],[10, 11, 12]]]))
y = tf.Variable(rtt.private_input(1, [[[13, 14],[15, 16],[17, 18]],[[19, 20],[21, 22],[23, 24]]]))

# Define matmul operation
res = tf.matmul(x, y)

# Start execution
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    res = sess.run(res)

    # Get the result of Rosetta matmul
    print('matmul:', sess.run(rtt.SecureReveal(res)))

rtt.deactivate()

[YuanFengChen]

我尝试运行三维tensor相乘，在tensorflow中，只需要第一个维度相同，得到结果是后面两个维度相乘。但Rosetta似乎不支持？ 报错： ValueError: Shape must be rank 2 but is rank 3 for 'RttMatmul' (op: 'RttMatmul') with input shapes: [2,2,3], [2,3,2]. 代码如下：

#!/usr/bin/env python3

# Import rosetta package
import latticex.rosetta as rtt
import tensorflow as tf

# You can activate a backend protocol, here use SecureNN
rtt.activate("SecureNN")

# Get private data from Alice (input x), Bob (input y)
x = tf.Variable(rtt.private_input(0, [[[ 1, 2, 3],[ 4, 5, 6]],[[ 7, 8, 9],[10, 11, 12]]]))
y = tf.Variable(rtt.private_input(1, [[[13, 14],[15, 16],[17, 18]],[[19, 20],[21, 22],[23, 24]]]))

# Define matmul operation
res = tf.matmul(x, y)

# Start execution
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    res = sess.run(res)

    # Get the result of Rosetta matmul
    print('matmul:', sess.run(rtt.SecureReveal(res)))

rtt.deactivate()

嗯，目前暂只支持二维，未来再开放多维功能。

[xiaoshui240]

我尝试运行三维tensor相乘，在tensorflow中，只需要第一个维度相同，得到结果是后面两个维度相乘。但Rosetta似乎不支持？ 报错： ValueError: Shape must be rank 2 but is rank 3 for 'RttMatmul' (op: 'RttMatmul') with input shapes: [2,2,3], [2,3,2]. 代码如下：

#!/usr/bin/env python3

# Import rosetta package
import latticex.rosetta as rtt
import tensorflow as tf

# You can activate a backend protocol, here use SecureNN
rtt.activate("SecureNN")

# Get private data from Alice (input x), Bob (input y)
x = tf.Variable(rtt.private_input(0, [[[ 1, 2, 3],[ 4, 5, 6]],[[ 7, 8, 9],[10, 11, 12]]]))
y = tf.Variable(rtt.private_input(1, [[[13, 14],[15, 16],[17, 18]],[[19, 20],[21, 22],[23, 24]]]))

# Define matmul operation
res = tf.matmul(x, y)

# Start execution
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    res = sess.run(res)

    # Get the result of Rosetta matmul
    print('matmul:', sess.run(rtt.SecureReveal(res)))

rtt.deactivate()

嗯，目前暂只支持二维，未来再开放多维功能。

好吧，但要实现多维的话，我可以做哪些方面的努力吗

[joyoFeng]

我尝试运行三维tensor相乘，在tensorflow中，只需要第一个维度相同，得到结果是后面两个维度相乘。但Rosetta似乎不支持？ 报错： ValueError: Shape must be rank 2 but is rank 3 for 'RttMatmul' (op: 'RttMatmul') with input shapes: [2,2,3], [2,3,2]. 代码如下：

#!/usr/bin/env python3

# Import rosetta package
import latticex.rosetta as rtt
import tensorflow as tf

# You can activate a backend protocol, here use SecureNN
rtt.activate("SecureNN")

# Get private data from Alice (input x), Bob (input y)
x = tf.Variable(rtt.private_input(0, [[[ 1, 2, 3],[ 4, 5, 6]],[[ 7, 8, 9],[10, 11, 12]]]))
y = tf.Variable(rtt.private_input(1, [[[13, 14],[15, 16],[17, 18]],[[19, 20],[21, 22],[23, 24]]]))

# Define matmul operation
res = tf.matmul(x, y)

# Start execution
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    res = sess.run(res)

    # Get the result of Rosetta matmul
    print('matmul:', sess.run(rtt.SecureReveal(res)))

rtt.deactivate()

嗯，目前暂只支持二维，未来再开放多维功能。

好吧，但要实现多维的话，我可以做哪些方面的努力吗

主要是需要处理tensor shape，转换为了二维矩阵的方式处理。未来不久将考虑添加这些支持更新