def mnist_mlp_model(x):
paddings = tf.constant([[0, 0], [0, 1], [0, 1], [0, 0]], name='pad_const')
x = tf.pad(x, paddings)
W_conv1 = tf.compat.v1.get_variable('W_conv1', [5, 5, 1, 5])
y = conv2d_stride_2_valid(x, W_conv1)
W_bc1 = tf.compat.v1.get_variable('W_conv1_bias', [1, 13, 13, 5])
y = y + W_bc1
y = tf.nn.relu(y)
y = max_pool_3x3_same_size(y)
W_conv2 = tf.compat.v1.get_variable('W_conv2', [5, 5, 5, 50])
y = conv2d_stride_2_valid(y, W_conv2)
y = max_pool_3x3_same_size(y)
y = tf.reshape(y, [-1, 5 * 5 * 50])
W_fc1 = tf.compat.v1.get_variable('W_fc1', [5 * 5 * 50, 100])
W_b1 = tf.compat.v1.get_variable('W_fc1_bias', [100])
y = tf.matmul(y, W_fc1)
y = y + W_b1
y = tf.nn.relu(y)
W_fc2 = tf.compat.v1.get_variable('W_fc2', [100, 10])
W_b2 = tf.compat.v1.get_variable('W_fc2_bias', [10])
y = tf.matmul(y, W_fc2)
y = tf.add(y, W_b2, name='output')
return y
def cryptonets_relu_model(x, mode):
if mode not in set(['train', 'test']):
print('mode should be train or test')
raise Exception()
paddings = tf.constant([[0, 0], [0, 1], [0, 1], [0, 0]], name='pad_const')
x = tf.pad(x, paddings)
W_conv1 = get_variable('W_conv1', [5, 5, 1, 5], mode)
y = conv2d_stride_2_valid(x, W_conv1)
W_bc1 = get_variable('W_conv1_bias', [1, 13, 13, 5], mode)
y = y + W_bc1
y = tf.nn.relu(y)
y = avg_pool_3x3_same_size(y)
W_conv2 = get_variable('W_conv2', [5, 5, 5, 50], mode)
y = conv2d_stride_2_valid(y, W_conv2)
y = avg_pool_3x3_same_size(y)
y = tf.reshape(y, [-1, 5 * 5 * 50])
W_fc1 = get_variable('W_fc1', [5 * 5 * 50, 100], mode)
W_b1 = get_variable('W_fc1_bias', [100], mode)
y = tf.matmul(y, W_fc1)
y = y + W_b1
y = tf.nn.relu(y)
W_fc2 = get_variable('W_fc2', [100, 10], mode)
W_b2 = get_variable('W_fc2_bias', [10], mode)
y = tf.matmul(y, W_fc2)
y = y + W_b2
return y
def cryptonets_test_averaged_squashed(x):
"""Constructs test network for Cryptonets using saved weights.
Assumes linear layers have been squashed."""
paddings = [[0, 0], [0, 1], [0, 1], [0, 0]]
x = tf.pad(x, paddings)
W_conv1_Part1 = np.loadtxt('W_conv1_Part1.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1_Part2 = np.loadtxt('W_conv1_Part2.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1_Part3 = np.loadtxt('W_conv1_Part3.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1_Part4 = np.loadtxt('W_conv1_Part4.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1 = (W_conv1_Part1 + W_conv1_Part2 + W_conv1_Part3 + W_conv1_Part4) *0.25
y = conv2d_stride_2_valid(x, W_conv1)
y = tf.square(y)
W_squash_Part1 = np.loadtxt('W_squash_Part1.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part2 = np.loadtxt('W_squash_Part2.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part3 = np.loadtxt('W_squash_Part3.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part4 = np.loadtxt('W_squash_Part4.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash = (W_squash_Part1+W_squash_Part2+W_squash_Part3+W_squash_Part4) *0.25
y = tf.reshape(y, [-1, 5 * 13 * 13])
y = tf.matmul(y, W_squash)
y = tf.square(y)
W_fc2_Part1 = np.loadtxt('W_fc2_Part1.txt',dtype=np.float32).reshape([100, 10])
W_fc2_Part2 = np.loadtxt('W_fc2_Part2.txt',dtype=np.float32).reshape([100, 10])
W_fc2_Part3 = np.loadtxt('W_fc2_Part3.txt',dtype=np.float32).reshape([100, 10])
W_fc2_Part4 = np.loadtxt('W_fc2_Part4.txt',dtype=np.float32).reshape([100, 10])
W_fc2 = (W_fc2_Part1+W_fc2_Part2+W_fc2_Part3+W_fc2_Part4)*0.25
y = tf.matmul(y, W_fc2)
return y
def cryptonets_relu_test_squashed(x):
"""Constructs test network for Cryptonets Relu using saved weights.
Assumes linear layers have been squashed."""
paddings = [[0, 0], [0, 1], [0, 1], [0, 0]]
x = tf.pad(x, paddings)
W_conv1 = get_variable('W_conv1', [5, 5, 1, 5], 'test')
y = conv2d_stride_2_valid(x, W_conv1)
W_bc1 = get_variable('W_conv1_bias', [1, 13, 13, 5], 'test')
y = tf.nn.relu(y)
W_squash = get_variable('W_squash', [5 * 13 * 13, 100], 'test')
y = tf.reshape(y, [-1, 5 * 13 * 13])
y = tf.matmul(y, W_squash)
W_b1 = get_variable('W_fc1_bias', [100], 'test')
y = y + W_b1
y = tf.nn.relu(y)
W_fc2 = get_variable('W_fc2', [100, 10], 'test')
y = tf.matmul(y, W_fc2)
W_b2 = get_variable('W_fc2_bias', [10], 'test')
y = y + W_b2
return y
def cryptonets_HE_avg(x):
paddings = [[0, 0], [0, 1], [0, 1], [0, 0]]
x = tf.pad(x, paddings)
W_conv1 = np.float32(np.load("conv1_avg_dec.npy",allow_pickle=True))
y = conv2d_stride_2_valid(x, W_conv1)
y = tf.square(y)
W_squash_Part1 = np.loadtxt('W_squash_repeatedSampling_Part1.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part2 = np.loadtxt('W_squash_repeatedSampling_Part2.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part3 = np.loadtxt('W_squash_repeatedSampling_Part3.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part4 = np.loadtxt('W_squash_repeatedSampling_Part4.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash = (W_squash_Part1+W_squash_Part2+W_squash_Part3+W_squash_Part4) *0.25
y = tf.reshape(y, [-1, 5 * 13 * 13])
y = tf.matmul(y, W_squash)
y = tf.square(y)
W_fc2 = np.float32(np.load("fc2_av_dec.npy",allow_pickle=True))
y = tf.matmul(y, W_fc2)
return y
def squash_layers():
print("Squashing layers")
tf.compat.v1.reset_default_graph()
# Input from h_conv1 squaring
x = tf.compat.v1.placeholder(tf.float32, [None, 13, 13, 5])
# Pooling layer
h_pool1 = avg_pool_3x3_same_size(x) # To N x 13 x 13 x 5
# Second convolution
W_conv2 = np.loadtxt('W_conv2.txt',
dtype=np.float32).reshape([5, 5, 5, 50])
h_conv2 = conv2d_stride_2_valid(h_pool1, W_conv2)
# Second pooling layer.
h_pool2 = avg_pool_3x3_same_size(h_conv2)
# Fully connected layer 1
# Input: N x 5 x 5 x 50
# Output: N x 100
W_fc1 = np.loadtxt('W_fc1.txt',
dtype=np.float32).reshape([5 * 5 * 50, 100])
h_pool2_flat = tf.reshape(h_pool2, [-1, 5 * 5 * 50])
pre_square = tf.matmul(h_pool2_flat, W_fc1)
with tf.compat.v1.Session() as sess:
x_in = np.eye(13 * 13 * 5)
x_in = x_in.reshape([13 * 13 * 5, 13, 13, 5])
W = (sess.run([pre_square], feed_dict={x: x_in}))[0]
squashed_file_name = "W_squash.txt"
np.savetxt(squashed_file_name, W)
print("Saved to", squashed_file_name)
# Sanity check
x_in = np.random.rand(100, 13, 13, 5)
network_out = (sess.run([pre_square], feed_dict={x: x_in}))[0]
linear_out = x_in.reshape(100, 13 * 13 * 5).dot(W)
assert (np.max(np.abs(linear_out - network_out)) < 1e-5)
print("Squashed layers")
def cryptonets_test_squashed_mode(x,mode):
"""Constructs test network for Cryptonets using saved weights.
Assumes linear layers have been squashed."""
paddings = [[0, 0], [0, 1], [0, 1], [0, 0]]
x = tf.pad(x, paddings)
if mode==1:
W_conv1_Part1 = np.loadtxt('W_conv1_InOrderSampling_Part1.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1_Part2 = np.loadtxt('W_conv1_InOrderSampling_Part2.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1_Part3 = np.loadtxt('W_conv1_InOrderSampling_Part3.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1_Part4 = np.loadtxt('W_conv1_InOrderSampling_Part4.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1 = (W_conv1_Part1 + W_conv1_Part2 + W_conv1_Part3 + W_conv1_Part4) *0.25
W_squash_Part1 = np.loadtxt('W_squash_InOrderSampling_Part1.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part2 = np.loadtxt('W_squash_InOrderSampling_Part2.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part3 = np.loadtxt('W_squash_InOrderSampling_Part3.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part4 = np.loadtxt('W_squash_InOrderSampling_Part4.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash = (W_squash_Part1+W_squash_Part2+W_squash_Part3+W_squash_Part4) *0.25
W_fc2_Part1 = np.loadtxt('W_fc2_InOrderSampling_Part1.txt',dtype=np.float32).reshape([100, 10])
W_fc2_Part2 = np.loadtxt('W_fc2_InOrderSampling_Part2.txt',dtype=np.float32).reshape([100, 10])
W_fc2_Part3 = np.loadtxt('W_fc2_InOrderSampling_Part3.txt',dtype=np.float32).reshape([100, 10])
W_fc2_Part4 = np.loadtxt('W_fc2_InOrderSampling_Part4.txt',dtype=np.float32).reshape([100, 10])
W_fc2 = (W_fc2_Part1+W_fc2_Part2+W_fc2_Part3+W_fc2_Part4)*0.25
elif mode==2:
W_conv1_Part1 = np.loadtxt('W_conv1_nonRepeatedSampling_Part1.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1_Part2 = np.loadtxt('W_conv1_nonRepeatedSampling_Part2.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1_Part3 = np.loadtxt('W_conv1_nonRepeatedSampling_Part3.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1_Part4 = np.loadtxt('W_conv1_nonRepeatedSampling_Part4.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1 = (W_conv1_Part1 + W_conv1_Part2 + W_conv1_Part3 + W_conv1_Part4) *0.25
W_squash_Part1 = np.loadtxt('W_squash_nonRepeatedSampling_Part1.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part2 = np.loadtxt('W_squash_nonRepeatedSampling_Part2.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part3 = np.loadtxt('W_squash_nonRepeatedSampling_Part3.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part4 = np.loadtxt('W_squash_nonRepeatedSampling_Part4.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash = (W_squash_Part1+W_squash_Part2+W_squash_Part3+W_squash_Part4) *0.25
W_fc2_Part1 = np.loadtxt('W_fc2_nonRepeatedSampling_Part1.txt',dtype=np.float32).reshape([100, 10])
W_fc2_Part2 = np.loadtxt('W_fc2_nonRepeatedSampling_Part2.txt',dtype=np.float32).reshape([100, 10])
W_fc2_Part3 = np.loadtxt('W_fc2_nonRepeatedSampling_Part3.txt',dtype=np.float32).reshape([100, 10])
W_fc2_Part4 = np.loadtxt('W_fc2_nonRepeatedSampling_Part4.txt',dtype=np.float32).reshape([100, 10])
W_fc2 = (W_fc2_Part1+W_fc2_Part2+W_fc2_Part3+W_fc2_Part4)*0.25
else:
W_conv1_Part1 = np.loadtxt('W_conv1_repeatedSampling_Part1.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1_Part2 = np.loadtxt('W_conv1_repeatedSampling_Part2.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1_Part3 = np.loadtxt('W_conv1_repeatedSampling_Part3.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1_Part4 = np.loadtxt('W_conv1_repeatedSampling_Part4.txt',dtype=np.float32).reshape([5,5,1,5])
W_conv1 = (W_conv1_Part1 + W_conv1_Part2 + W_conv1_Part3 + W_conv1_Part4) *0.25
W_squash_Part1 = np.loadtxt('W_squash_repeatedSampling_Part1.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part2 = np.loadtxt('W_squash_repeatedSampling_Part2.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part3 = np.loadtxt('W_squash_repeatedSampling_Part3.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash_Part4 = np.loadtxt('W_squash_repeatedSampling_Part4.txt',dtype=np.float32).reshape([5 * 13 * 13, 100])
W_squash = (W_squash_Part1+W_squash_Part2+W_squash_Part3+W_squash_Part4) *0.25
W_fc2_Part1 = np.loadtxt('W_fc2_repeatedSampling_Part1.txt',dtype=np.float32).reshape([100, 10])
W_fc2_Part2 = np.loadtxt('W_fc2_repeatedSampling_Part2.txt',dtype=np.float32).reshape([100, 10])
W_fc2_Part3 = np.loadtxt('W_fc2_repeatedSampling_Part3.txt',dtype=np.float32).reshape([100, 10])
W_fc2_Part4 = np.loadtxt('W_fc2_repeatedSampling_Part4.txt',dtype=np.float32).reshape([100, 10])
W_fc2 = (W_fc2_Part1+W_fc2_Part2+W_fc2_Part3+W_fc2_Part4)*0.25
y = conv2d_stride_2_valid(x, W_conv1)
y = tf.square(y)
y = tf.reshape(y, [-1, 5 * 13 * 13])
y = tf.matmul(y, W_squash)
y = tf.square(y)
y = tf.matmul(y, W_fc2)
return y
def cryptonets_model(x, mode):
"""Builds the graph for classifying digits based on Cryptonets
Args:
x: an input tensor with the dimensions (N_examples, 28, 28)
Returns:
A tuple (y, a scalar placeholder). y is a tensor of shape
(N_examples, 10), with values equal to the logits of classifying the
digit into one of 10 classes (the digits 0-9).
"""
if mode not in set(['train', 'test']):
print('mode should be train or test')
raise Exception()
# Reshape to use within a conv neural net.
# Last dimension is for "features" - there is only one here, since images
# are grayscale -- it would be 3 for an RGB image, 4 for RGBA, etc.
# CryptoNets's output of the first conv layer has feature map size 13 x 13,
# therefore, we manually add paddings.
with tf.name_scope('reshape'):
print('padding')
paddings = [[0, 0], [0, 1], [0, 1], [0, 0]]
x = tf.pad(x, paddings)
print('padded')
# First conv layer
# Input: N x 28 x 28 x 1
# Filter: 5 x 5 x 1 x 5
# Output: N x 13 x 13 x 5
with tf.name_scope('conv1'):
W_conv1 = get_variable("W_conv1", [5, 5, 1, 5], mode)
h_conv1 = tf.square(conv2d_stride_2_valid(x, W_conv1))
# Pooling layer
# Input: N x 13 x 13 x 5
# Output: N x 13 x 13 x 5
with tf.name_scope('pool1'):
h_pool1 = avg_pool_3x3_same_size(h_conv1)
# Second convolution
# Input: N x 13 x 13 x 5
# Filter: 5 x 5 x 5 x 50
# Output: N x 5 x 5 x 50
with tf.name_scope('conv2'):
W_conv2 = get_variable("W_conv2", [5, 5, 5, 50], mode)
h_conv2 = conv2d_stride_2_valid(h_pool1, W_conv2)
# Second pooling layer
# Input: N x 5 x 5 x 50
# Output: N x 5 x 5 x 50
with tf.name_scope('pool2'):
h_pool2 = avg_pool_3x3_same_size(h_conv2)
# Fully connected layer 1
# Input: N x 5 x 5 x 50
# Input flattened: N x 1250
# Weight: 1250 x 100
# Output: N x 100
with tf.name_scope('fc1'):
h_pool2_flat = tf.reshape(h_pool2, [-1, 5 * 5 * 50])
W_fc1 = get_variable("W_fc1", [5 * 5 * 50, 100], mode)
h_fc1 = tf.square(tf.matmul(h_pool2_flat, W_fc1))
# Map the 100 features to 10 classes, one for each digit
# Input: N x 100
# Weight: 100 x 10
# Output: N x 10
with tf.name_scope('fc2'):
W_fc2 = get_variable("W_fc2", [100, 10], mode)
y_conv = tf.matmul(h_fc1, W_fc2)
return y_conv
def cryptonets_train(x):
"""Builds the graph for classifying digits based on Cryptonets
Args:
x: an input tensor with the dimensions (N_examples, 784), where 784 is
the number of pixels in a standard MNIST image.
Returns:
A tuple (y, a scalar placeholder). y is a tensor of shape
(N_examples, 10), with values equal to the logits of classifying the
digit into one of 10 classes (the digits 0-9).
"""
# Reshape to use within a conv neural net.
# Last dimension is for "features" - there is only one here, since images
# are grayscale -- it would be 3 for an RGB image, 4 for RGBA, etc.
with tf.name_scope('reshape'):
paddings = [[0, 0], [0, 1], [0, 1], [0, 0]]
x = tf.pad(x, paddings)
# First conv layer
# CryptoNets's output of the first conv layer has feature map size 13 x 13,
# therefore, we manually add paddings.
# Input: N x 28 x 28 x 1
# Filter: 5 x 5 x 1 x 5
# Output: N x 12 x 12 x 5
# Output after padding: N x 13 x 13 x 5
with tf.name_scope('conv1'):
W_conv1 = tf.get_variable("W_conv1", [5, 5, 1, 5])
h_conv1_no_pad = tf.square(conv2d_stride_2_valid(x, W_conv1))
paddings = tf.constant([[0, 0], [0, 1], [0, 1], [0, 0]],
name='pad_const')
h_conv1 = tf.pad(h_conv1_no_pad, paddings)
# Pooling layer
# Input: N x 13 x 13 x 5
# Output: N x 13 x 13 x 5
with tf.name_scope('pool1'):
h_pool1 = avg_pool_3x3_same_size(h_conv1)
# Second convolution
# Input: N x 13 x 13 x 5
# Filter: 5 x 5 x 5 x 50
# Output: N x 5 x 5 x 50
with tf.name_scope('conv2'):
W_conv2 = tf.get_variable("W_conv2", [5, 5, 5, 50])
h_conv2 = conv2d_stride_2_valid(h_pool1, W_conv2)
# Second pooling layer
# Input: N x 5 x 5 x 50
# Output: N x 5 x 5 x 50
with tf.name_scope('pool2'):
h_pool2 = avg_pool_3x3_same_size(h_conv2)
# Fully connected layer 1
# Input: N x 5 x 5 x 50
# Input flattened: N x 1250
# Weight: 1250 x 100
# Output: N x 100
with tf.name_scope('fc1'):
h_pool2_flat = tf.reshape(h_pool2, [-1, 5 * 5 * 50])
W_fc1 = tf.get_variable("W_fc1", [5 * 5 * 50, 100])
h_fc1 = tf.square(tf.matmul(h_pool2_flat, W_fc1))
# Map the 100 features to 10 classes, one for each digit
# Input: N x 100
# Weight: 100 x 10
# Output: N x 10
with tf.name_scope('fc2'):
W_fc2 = tf.get_variable("W_fc2", [100, 10])
y_conv = tf.matmul(h_fc1, W_fc2)
return y_conv