def eq_cifar_fn(x, output_dim=10, trainable=True):
gconv_indices, gconv_shape_info, w_shape = gconv2d_util(h_input='Z2',
h_output='C4',
in_channels=3,
out_channels=8,
ksize=3)
w = tf.get_variable('w1', shape=w_shape)
conv1 = gconv2d(input=x,
filter=w,
strides=[1, 2, 2, 1],
padding='SAME',
gconv_indices=gconv_indices,
gconv_shape_info=gconv_shape_info)
tf.add_to_collection('conv_output1', conv1)
pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=[2, 2], strides=2)
gconv_indices, gconv_shape_info, w_shape = gconv2d_util(h_input='C4',
h_output='C4',
in_channels=16,
out_channels=32,
ksize=5)
w = tf.get_variable('w2', shape=w_shape)
conv2 = gconv2d(input=conv1,
filter=w,
strides=[1, 2, 2, 1],
padding='SAME',
gconv_indices=gconv_indices,
gconv_shape_info=gconv_shape_info)
pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2], strides=2)
gconv_indices, gconv_shape_info, w_shape = gconv2d_util(h_input='C4',
h_output='C4',
in_channels=8,
out_channels=2,
ksize=5)
w = tf.get_variable('w3', shape=w_shape)
conv3 = gconv2d(input=conv2,
filter=w,
strides=[1, 1, 1, 1],
padding='SAME',
gconv_indices=gconv_indices,
gconv_shape_info=gconv_shape_info)
conv3 = tf.reshape(conv3,
conv3.get_shape().as_list()[:3] + [4] + [out_channels])
conv3 = tf.reduce_mean(conv3, axis=3)
pool3 = tf.layers.max_pooling2d(inputs=conv3, pool_size=[2, 2], strides=2)
pool3_flat = tf.layers.flatten(pool3)
u = pool3_flat
u = tf.layers.dense(inputs=pool3_flat,
units=output_dim,
activation=tf.nn.relu,
trainable=trainable)
tf.add_to_collection('conv_output2', conv2)
return u
def eq_cnn_fn(x, output_dim=10, trainable=True, group='C4', num_filters=2):
nchannels = x.shape[3]
gconv_indices, gconv_shape_info, w_shape = gconv2d_util(
h_input='Z2',
h_output='C4',
in_channels=nchannels,
out_channels=2,
ksize=5)
w = tf.get_variable('w1', shape=w_shape)
conv1 = gconv2d(input=x,
filter=w,
strides=[1, 1, 1, 1],
padding='SAME',
gconv_indices=gconv_indices,
gconv_shape_info=gconv_shape_info)
tf.add_to_collection('conv_output1', conv1)
pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=[2, 2], strides=2)
# pool1 = layers.Dropout(0.25)(pool1)
out_channels = 2
gconv_indices, gconv_shape_info, w_shape = gconv2d_util(
h_input='C4',
h_output='C4',
in_channels=2,
out_channels=out_channels,
ksize=5)
w = tf.get_variable('w2', shape=w_shape)
conv2 = gconv2d(input=conv1,
filter=w,
strides=[1, 1, 1, 1],
padding='SAME',
gconv_indices=gconv_indices,
gconv_shape_info=gconv_shape_info)
conv2 = tf.reshape(conv2,
conv2.get_shape().as_list()[:3] + [4] + [out_channels])
conv2 = tf.reduce_mean(conv2, axis=3)
conv2 = tf.reshape(conv2, conv2.get_shape().as_list()[:3] + [out_channels])
pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2], strides=2)
pool2_flat = tf.layers.flatten(pool2)
u = pool2_flat
print(u.shape)
u = tf.layers.dense(inputs=pool2_flat,
units=output_dim,
activation=tf.nn.relu,
trainable=trainable)
tf.add_to_collection('conv_output2', conv2)
return u
def call(self, inputs, **kwargs):
x = gconv2d(input=inputs,
filter=self.w,
strides=self.strides,
padding=self.padding,
gconv_indices=self.gconv_indices,
gconv_shape_info=self.gconv_shape_info)
return x
def make_graph(h_input, h_output):
gconv_indices, gconv_shape_info, w_shape = gconv2d_util(
h_input=h_input, h_output=h_output, in_channels=1, out_channels=1, ksize=3)
nti = gconv_shape_info[-2]
x = tf.placeholder(tf.float32, [None, 5, 5, 1 * nti])
w = tf.Variable(tf.truncated_normal(w_shape, stddev=1.))
y = gconv2d(input=x, filter=w, strides=[1, 1, 1, 1], padding='SAME',
gconv_indices=gconv_indices, gconv_shape_info=gconv_shape_info)
return x, y
def call(self, x):
act = gconv2d(input=x,
filter=self.W,
strides=self.stridestf,
padding=self.paddingtf,
gconv_indices=self.gconv_indices,
gconv_shape_info=self.gconv_shape_info)
if self.use_bias:
bias_replicated = tf.tile(self.b, [1, self.gconv_shape_info[1]])
bias_reshaped = tf.reshape(
bias_replicated,
[1, 1, 1, self.num_groups * self.gconv_shape_info[1]])
#act_reshaped = tf.reshape()
#act = tf.nn.bias_add(act,bias_replicated)
act = act + bias_reshaped
return act
def make_graph(h_input, h_output):
gconv_indices, gconv_shape_info, w_shape = gconv2d_util(h_input=h_input,
h_output=h_output,
in_channels=1,
out_channels=1,
ksize=3)
nti = gconv_shape_info[-2]
x = tf.placeholder(tf.float32, [None, 5, 5, 1 * nti])
w = tf.Variable(tf.truncated_normal(w_shape, stddev=1.))
y = gconv2d(input=x,
filter=w,
strides=[1, 1, 1, 1],
padding='SAME',
gconv_indices=gconv_indices,
gconv_shape_info=gconv_shape_info)
return x, y
def gconv_bn_act(x,
gconv_type,
C_in,
C_out,
ksize=3,
padding='VALID',
bn=True,
is_train=True,
act=tf.nn.relu,
name='gconv_bn_act',
reuse=False):
if gconv_type == 'Z2_to_P4':
h_in = 'Z2'
h_out = 'C4'
elif gconv_type == 'Z2_to_P4M':
h_in = 'Z2'
h_out = 'D4'
elif gconv_type == 'P4_to_P4':
h_in = 'C4'
h_out = 'C4'
elif gconv_type == 'P4M_to_P4M':
h_in = 'D4'
h_out = 'D4'
else:
raise NotImplemented('Unsupported gconv_type: {}!'.format(gconv_type))
with tf.variable_scope(name, reuse=reuse) as vs:
gconv_indices, gconv_shape_info, w_shape = gconv2d_util(
h_input=h_in,
h_output=h_out,
in_channels=C_in,
out_channels=C_out,
ksize=ksize)
# w = tf.Variable(tf.truncated_normal(w_shape, stddev=1.), name='kernel')
w = tf.get_variable("kernel",
shape=w_shape,
initializer=tf.contrib.layers.xavier_initializer()
) # initialization is very important!!!
x = gconv2d(input=x,
filter=w,
strides=[1, 1, 1, 1],
padding=padding,
gconv_indices=gconv_indices,
gconv_shape_info=gconv_shape_info,
use_cudnn_on_gpu=True)
if bn:
x = tf.layers.batch_normalization(x, training=is_train)
return act(x)
def _gconv(self, name, x, filter_size, in_filters, out_filters, strides):
"""G - Convolution."""
gconv_indices, gconv_shape_info, w_shape = gconv2d_util(
h_input='D4',
h_output='D4',
in_channels=in_filters,
out_channels=out_filters,
ksize=3)
w = tf.Variable(tf.truncated_normal(w_shape, stddev=1.))
x = gconv2d(input=x,
filter=w,
strides=strides,
padding='SAME',
gconv_indices=gconv_indices,
gconv_shape_info=gconv_shape_info,
name=name)
return x
def g_conv_no_decorator(self,
input,
input_type,
output_type,
kernel_size,
c_o,
s_h,
s_w,
name,
padding=DEFAULT_PADDING,
group=1):
# Verify that the padding is acceptable
self.validate_padding(padding)
c_i = np.int32(input.get_shape()[-1])
if input_type == 'Z2':
c_i = c_i
elif input_type == 'C4':
c_i = c_i // 4
else:
c_i = c_i // 8
# output = GConv2D(c_o, (kernel_size, kernel_size), kernel_initializer='he_normal', padding=padding, strides=(s_h, s_w), use_bias=False, h_input=input_type, h_output=output_type, name=name)(input)
with tf.variable_scope(name) as scope:
gconv_indices, gconv_shape_info, w_shape = gconv2d_util(
h_input=input_type,
h_output=output_type,
in_channels=c_i,
out_channels=c_o,
ksize=kernel_size)
# w = tf.Variable(tf.truncated_normal(w_shape, stddev=1.))
w = self.make_var(name='weight', shape=w_shape)
output = gconv2d(input=input,
filter=w,
strides=[1, 1, 1, 1],
padding='SAME',
gconv_indices=gconv_indices,
gconv_shape_info=gconv_shape_info)
return output
def _build_model(self,
config,
filters,
num_ids,
differentiable=False,
adversarial_ce=False,
nat_ce=False,
pad_mode='CONSTANT',
pad_size=32):
"""Build the core model within the graph."""
with tf.variable_scope('input'):
self.group = tf.placeholder(tf.int32, [None], name="group")
self.num_ids = num_ids
self.x_input = tf.placeholder(tf.float32, shape=[None, 32, 32, 3])
self.y_input = tf.placeholder(tf.int64, shape=None)
self.transform = tf.placeholder(tf.float32, shape=[None, 3])
trans_x, trans_y, rot = tf.unstack(self.transform, axis=1)
rot *= np.pi / 180 # convert degrees to radians
self.is_training = tf.placeholder(tf.bool)
x = self.x_input
x = tf.pad(x, [[0, 0], [16, 16], [16, 16], [0, 0]], pad_mode)
if not differentiable:
# For spatial non-PGD attacks: rotate and translate image
ones = tf.ones(shape=tf.shape(trans_x))
zeros = tf.zeros(shape=tf.shape(trans_x))
trans = tf.stack([
ones, zeros, -trans_x, zeros, ones, -trans_y, zeros, zeros
],
axis=1)
x = tf.contrib.image.rotate(x, rot, interpolation='BILINEAR')
x = tf.contrib.image.transform(x,
trans,
interpolation='BILINEAR')
else:
# for spatial PGD attacks need to use diffble transformer
theta = tf.stack([
tf.cos(rot), -tf.sin(rot), trans_x / 64,
tf.sin(rot),
tf.cos(rot), trans_y / 64
],
axis=1)
x = transformer(x, theta, (64, 64))
x = tf.image.resize_image_with_crop_or_pad(x, pad_size, pad_size)
# everything below this point is generic (independent of spatial attacks)
self.x_image = x
x = tf.map_fn(lambda img: tf.image.per_image_standardization(img),
x)
gconv_indices, gconv_shape_info, w_shape = gconv2d_util(
h_input='Z2',
h_output='D4',
in_channels=3,
out_channels=filters[0],
ksize=3)
w = tf.Variable(tf.truncated_normal(w_shape, stddev=1.))
x = gconv2d(input=x,
filter=w,
strides=[1, 1, 1, 1],
padding='SAME',
gconv_indices=gconv_indices,
gconv_shape_info=gconv_shape_info)
# x = self._conv('init_conv', x, 3, 3, 16, self._stride_arr(1))
# "filters": [16, 16, 32, 64],
strides = [1, 2, 2]
activate_before_residual = [True, False, False]
res_func = self._residual
with tf.variable_scope('unit_1_0'):
x = res_func(x, filters[0], filters[1],
self._stride_arr(strides[0]),
activate_before_residual[0])
for i in range(1, config.resnet_depth_n):
with tf.variable_scope('unit_1_%d' % i):
x = res_func(x, filters[1], filters[1], self._stride_arr(1),
False)
with tf.variable_scope('unit_2_0'):
x = res_func(x, filters[1], filters[2],
self._stride_arr(strides[1]),
activate_before_residual[1])
for i in range(1, config.resnet_depth_n):
with tf.variable_scope('unit_2_%d' % i):
x = res_func(x, filters[2], filters[2], self._stride_arr(1),
False)
with tf.variable_scope('unit_3_0'):
x = res_func(x, filters[2], filters[3],
self._stride_arr(strides[2]),
activate_before_residual[2])
for i in range(1, config.resnet_depth_n):
with tf.variable_scope('unit_3_%d' % i):
x = res_func(x, filters[3], filters[3], self._stride_arr(1),
False)
with tf.variable_scope('unit_last'):
x = self._batch_norm('final_bn', x)
x = self._relu(x, 0.1)
x = self._global_avg_pool(x)
# uncomment to add and extra fc layer
#with tf.variable_scope('unit_fc'):
# self.pre_softmax = self._fully_connected(x, 1024)
# x = self._relu(x, 0.1)
with tf.variable_scope('logit'):
self.pre_softmax = self._fully_connected(x, config.n_classes)
self.predictions = tf.argmax(self.pre_softmax, 1)
self.correct_prediction = tf.equal(self.predictions, self.y_input)
self.num_correct = tf.reduce_sum(
tf.cast(self.correct_prediction, tf.int64))
self.accuracy = tf.reduce_mean(
tf.cast(self.correct_prediction, tf.float32))
with tf.variable_scope('costs'):
if adversarial_ce:
indices_adv = tf.cast(
tf.range(self.num_ids,
tf.shape(self.pre_softmax)[0]), tf.int32)
adversarial_ex_presoft = tf.gather(self.pre_softmax,
indices_adv)
adversarial_ex_y = tf.gather(self.y_input, indices_adv)
self.y_xent_for_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=adversarial_ex_presoft, labels=adversarial_ex_y)
self.y_xent = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=self.pre_softmax, labels=self.y_input)
elif nat_ce:
print("******************pure_nat_loss/n")
indices_nat = tf.cast(
tf.range(tf.shape(self.pre_softmax)[0] - self.num_ids),
tf.int32)
nat_ex_presoft = tf.gather(self.pre_softmax, indices_nat)
nat_ex_y = tf.gather(self.y_input, indices_nat)
self.y_xent_for_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=nat_ex_presoft, labels=nat_ex_y)
self.y_xent = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=self.pre_softmax, labels=self.y_input)
else:
self.y_xent_for_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=self.pre_softmax, labels=self.y_input)
self.y_xent = self.y_xent_for_loss
self.xent = tf.reduce_sum(self.y_xent, name='y_xent')
self.mean_xent = tf.reduce_mean(self.y_xent_for_loss)
self.weight_decay_loss = self._decay()
#TRADES penalty
# self.core_loss = self._CoRe()
self.core_loss2 = self._CoRe_2tensors()