def MLP3_test(x, opt, select, labels_id, dropout_rate, perturbation_params,
perturbation_type):
# x is an input, opt is the experiment
# select gives selected neurons where 1 indicates being selected, indexed as [layer][node]
# labels_id=categories
# dropout rate=dropout rate
# perturbation_params=array of keep/drop probs indexed [type][layer]
# perturbation_type is an int in range(5) giving perturbation type:
# 0=weight noise, 1=weight ko, 2=act ko, 3=act noise, 4=targeted act ko
parameters = []
num_neurons_before_fc = int(x.get_shape()[1])
flattened = tf.reshape(x, [-1, num_neurons_before_fc])
# fc1
with tf.name_scope('fc1') as scope:
W = tf.Variable(tf.truncated_normal(
[num_neurons_before_fc,
int(32 * opt.dnn.neuron_multiplier[0])],
dtype=tf.float32,
stddev=1e-3),
name='weights')
b = tf.Variable(0.1 *
tf.ones([int(32 * opt.dnn.neuron_multiplier[0])]),
name='bias')
# weight perturbation
if perturbation_type == 0:
W = pt.weight_knockout(W, perturbation_params[1][0])
if perturbation_type == 1:
W = pt.weight_noise(W, perturbation_params[0][0])
parameters += [W, b]
fc1_predrop = tf.nn.relu(
tf.nn.bias_add(tf.matmul(flattened, W), b, name=scope))
# if perturbation_type == 2:
# fc3_predrop = pt.activation_knockout(fc3_predrop, perturbation_params[2][2])
if perturbation_type == 3:
fc1_predrop = pt.activation_noise(fc1_predrop,
perturbation_params[3][0],
opt.hyper.batch_size)
elif perturbation_type in [2, 4]:
ss = tf.reshape(tf.tile(select[0], [opt.hyper.batch_size]),
[-1, int(fc1_predrop.get_shape()[1])])
fc1_predrop = pt.activation_knockout_mask(
fc1_predrop, perturbation_params[4][0], ss)
dropout1 = tf.nn.dropout(fc1_predrop, dropout_rate)
# fc2
with tf.name_scope('fc2') as scope:
W = tf.Variable(tf.truncated_normal([
int(32 * opt.dnn.neuron_multiplier[0]),
int(32 * opt.dnn.neuron_multiplier[1])
],
dtype=tf.float32,
stddev=1e-3),
name='weights')
b = tf.Variable(0.1 *
tf.ones([int(32 * opt.dnn.neuron_multiplier[1])]),
name='bias')
# weight perturbation
if perturbation_type == 0:
W = pt.weight_knockout(W, perturbation_params[1][1])
if perturbation_type == 1:
W = pt.weight_noise(W, perturbation_params[0][1])
parameters += [W, b]
fc2_predrop = tf.nn.relu(
tf.nn.bias_add(tf.matmul(dropout1, W), b, name=scope))
if perturbation_type == 3:
fc2_predrop = pt.activation_noise(fc2_predrop,
perturbation_params[3][1],
opt.hyper.batch_size)
elif perturbation_type in [2, 4]:
ss = tf.reshape(tf.tile(select[1], [opt.hyper.batch_size]),
[-1, int(fc2_predrop.get_shape()[1])])
fc2_predrop = pt.activation_knockout_mask(
fc2_predrop, perturbation_params[4][1], ss)
dropout2 = tf.nn.dropout(fc2_predrop, dropout_rate)
# fc3
with tf.name_scope('fc2') as scope:
W = tf.Variable(tf.truncated_normal([
int(32 * opt.dnn.neuron_multiplier[1]),
int(32 * opt.dnn.neuron_multiplier[2])
],
dtype=tf.float32,
stddev=1e-3),
name='weights')
b = tf.Variable(0.1 *
tf.ones([int(32 * opt.dnn.neuron_multiplier[2])]),
name='bias')
# weight perturbation
if perturbation_type == 0:
W = pt.weight_knockout(W, perturbation_params[1][2])
if perturbation_type == 1:
W = pt.weight_noise(W, perturbation_params[0][2])
parameters += [W, b]
fc3_predrop = tf.nn.relu(
tf.nn.bias_add(tf.matmul(dropout2, W), b, name=scope))
# if perturbation_type == 2:
# fc3_predrop = pt.activation_knockout(fc3_predrop, perturbation_params[2][2])
if perturbation_type == 3:
fc3_predrop = pt.activation_noise(fc3_predrop,
perturbation_params[3][2],
opt.hyper.batch_size)
elif perturbation_type in [2, 4]:
ss = tf.reshape(tf.tile(select[0], [opt.hyper.batch_size]),
[-1, int(fc3_predrop.get_shape()[1])])
fc3_predrop = pt.activation_knockout_mask(
fc3_predrop, perturbation_params[4][2], ss)
dropout3 = tf.nn.dropout(fc3_predrop, dropout_rate)
# fc8
with tf.name_scope('fc_out') as scope:
W = tf.Variable(tf.truncated_normal(
[int(32 * opt.dnn.neuron_multiplier[2]),
len(labels_id)],
dtype=tf.float32,
stddev=1e-2),
name='weights')
b = tf.Variable(tf.zeros([len(labels_id)]), name='bias')
# weight perturbation
if perturbation_type == 0:
W = pt.weight_knockout(W, perturbation_params[1][3])
if perturbation_type == 1:
W = pt.weight_noise(W, perturbation_params[0][3])
parameters += [W, b]
summ.variable_summaries(W, b, opt)
fc8 = tf.nn.bias_add(tf.matmul(dropout3, W), b, name=scope)
summ.activation_summaries(fc8, opt)
return fc8, parameters
def Alexnet_test(x, opt, select, labels_id, dropout_rate, perturbation_params,
perturbation_type):
# x is an input, opt is the experiment
# select gives selected neurons where 1 indicates being selected, indexed as [layer][node]
# labels_id=categories
# dropout rate=dropout rate
# perturbation_params=array of keep/drop probs indexed [type][layer]
# perturbation_type is an int in range(5) giving perturbation type:
# 0=weight noise, 1=weight ko, 2=act ko, 3=act noise, 4=targeted act ko
reuse = False
global num_neurons
parameters = []
init_type = tf.glorot_normal_initializer
if opt.init_type == 1:
init_type = tf.glorot_uniform_initializer
elif opt.init_type == 2:
init_type = tf.keras.initializers.he_normal
elif opt.init_type == 3:
init_type = tf.keras.initializers.he_uniform
elif opt.init_type == 4:
init_type = tf.keras.initializers.lecun_normal
elif opt.init_type == 5:
init_type = tf.keras.initializers.lecun_uniform
f_act = tf.nn.relu
if opt.act_function == 1:
f_act = tf.nn.leaky_relu
elif opt.act_function == 2:
f_act = tf.nn.elu
elif opt.act_function == 3:
f_act = tf.nn.selu
elif opt.act_function == 4:
f_act = tf.nn.sigmoid
elif opt.act_function == 5:
f_act = tf.nn.tanh
with tf.variable_scope('conv1', reuse=reuse) as scope:
kernel = tf.get_variable(shape=[
5, 5, 3,
int(num_neurons[0] * opt.dnn.neuron_multiplier[0])
],
initializer=init_type(),
name='weights')
conv = tf.nn.conv2d(x, kernel, [1, 1, 1, 1], padding='SAME')
biases = tf.get_variable(initializer=tf.constant(
0.0, shape=[int(num_neurons[0] * opt.dnn.neuron_multiplier[0])]),
name='biases')
pre_activation = tf.nn.bias_add(conv, biases)
conv1 = f_act(pre_activation, name=scope.name)
# Activation perturbation
# if perturbation_type == 2:
# conv1 = pt.activation_knockout(conv1, perturbation_params[2][0])
if perturbation_type == 3:
conv1 = pt.activation_noise(conv1, perturbation_params[3][0],
opt.hyper.batch_size)
elif perturbation_type in [2, 4]:
ss = tf.reshape(
tf.tile(select[0], [
int(np.prod(conv1.get_shape()[1:3])) * opt.hyper.batch_size
]), [
-1,
int(conv1.get_shape()[1]),
int(conv1.get_shape()[2]),
int(conv1.get_shape()[3])
])
conv1 = pt.activation_knockout_mask(conv1,
perturbation_params[4][0],
ss) # ss is the mask
parameters += [kernel, biases]
summ.variable_summaries(kernel, biases, opt)
summ.activation_summaries(conv1, opt)
# pool1
with tf.variable_scope('pool1') as scope:
pool1 = tf.nn.max_pool(conv1,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool1')
summ.activation_summaries(pool1, opt)
# lrn1
with tf.variable_scope('lrn1') as scope:
lrn1 = tf.nn.local_response_normalization(pool1,
depth_radius=2,
alpha=2e-05,
beta=0.75,
bias=1.)
# conv2
with tf.variable_scope('conv2', reuse=reuse) as scope:
kernel = tf.get_variable(shape=[
5, 5,
int(num_neurons[0] * opt.dnn.neuron_multiplier[0]),
int(num_neurons[1] * opt.dnn.neuron_multiplier[1])
],
initializer=init_type(),
name='weights')
conv = tf.nn.conv2d(lrn1, kernel, [1, 1, 1, 1], padding='SAME')
biases = tf.get_variable(initializer=tf.constant(
0.1, shape=[int(num_neurons[1] * opt.dnn.neuron_multiplier[1])]),
name='biases')
pre_activation = tf.nn.bias_add(conv, biases)
conv2 = f_act(pre_activation, name=scope.name)
# Activation perturbation
# if perturbation_type == 2:
# conv2 = pt.activation_knockout(conv2, perturbation_params[2][1])
if perturbation_type == 3:
conv2 = pt.activation_noise(conv2, perturbation_params[3][1],
opt.hyper.batch_size)
elif perturbation_type in [2, 4]:
ss = tf.reshape(
tf.tile(select[1], [
int(np.prod(conv2.get_shape()[1:3])) * opt.hyper.batch_size
]), [
-1,
int(conv2.get_shape()[1]),
int(conv2.get_shape()[2]),
int(conv2.get_shape()[3])
])
conv2 = pt.activation_knockout_mask(conv2,
perturbation_params[4][1], ss)
parameters += [kernel, biases]
summ.variable_summaries(kernel, biases, opt)
summ.activation_summaries(conv2, opt)
# pool2
with tf.variable_scope('pool2') as scope:
pool2 = tf.nn.max_pool(conv2,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME')
summ.activation_summaries(pool2, opt)
# lrn2
with tf.variable_scope('lrn2') as scope:
lrn2 = tf.nn.local_response_normalization(pool2,
depth_radius=2,
alpha=2e-05,
beta=0.75,
bias=1.)
# fc3 (6 in Alexnet)
with tf.variable_scope('fc3', reuse=reuse) as scope:
dim1 = int(np.prod(lrn2.get_shape()[1:]))
dim2 = int(num_neurons[2] * opt.dnn.neuron_multiplier[2])
flattened = tf.reshape(lrn2, [opt.hyper.batch_size, -1])
weights = tf.get_variable(shape=[dim1, dim2],
initializer=init_type(),
name='weights')
biases = tf.get_variable(initializer=tf.constant(0.1, shape=[dim2]),
name='biases')
# weight perturbation
if perturbation_type == 0:
weights = pt.weight_knockout(weights, perturbation_params[1][2])
if perturbation_type == 1:
weights = pt.weight_noise(weights, perturbation_params[0][2])
parameters += [weights, biases]
summ.variable_summaries(weights, biases, opt)
fc3_predrop = f_act(tf.matmul(flattened, weights) + biases,
name=scope.name)
# activation perturbation
# if perturbation_type == 2:
# fc3_predrop = pt.activation_knockout(fc3_predrop, perturbation_params[2][2])
if perturbation_type == 3:
fc3_predrop = pt.activation_noise(fc3_predrop,
perturbation_params[3][2],
opt.hyper.batch_size)
elif perturbation_type in [2, 4]:
ss = tf.reshape(tf.tile(select[2], [opt.hyper.batch_size]),
[-1, int(fc3_predrop.get_shape()[1])])
fc3_predrop = pt.activation_knockout_mask(
fc3_predrop, perturbation_params[4][2], ss)
fc3 = tf.nn.dropout(fc3_predrop, dropout_rate)
summ.activation_summaries(fc3, opt)
# fc4 (fc7 in Alexnet)
with tf.variable_scope('fc4', reuse=reuse) as scope:
dim3 = int(num_neurons[3] * opt.dnn.neuron_multiplier[3])
weights = tf.get_variable(shape=[dim2, dim3],
initializer=init_type(),
name='weights')
biases = tf.get_variable(initializer=tf.constant(
0.1, shape=[int(num_neurons[3] * opt.dnn.neuron_multiplier[3])]),
name='biases')
# weight perturbation
if perturbation_type == 0:
weights = pt.weight_knockout(weights, perturbation_params[1][3])
if perturbation_type == 1:
weights = pt.weight_noise(weights, perturbation_params[0][3])
parameters += [weights, biases]
summ.variable_summaries(weights, biases, opt)
fc4_predrop = f_act(tf.matmul(fc3, weights) + biases, name=scope.name)
# activation perturbation
# if perturbation_type == 2:
# fc4_predrop = pt.activation_knockout(fc4_predrop, perturbation_params[2][3])
if perturbation_type == 3:
fc4_predrop = pt.activation_noise(fc4_predrop,
perturbation_params[3][3],
opt.hyper.batch_size)
elif perturbation_type in [2, 4]:
ss = tf.reshape(tf.tile(select[3], [opt.hyper.batch_size]),
[-1, int(fc4_predrop.get_shape()[1])])
fc4_predrop = pt.activation_knockout_mask(
fc4_predrop, perturbation_params[4][3], ss)
fc4 = tf.nn.dropout(fc4_predrop, dropout_rate)
summ.activation_summaries(fc4, opt)
# linear softmax (fc8 in Alexnet)
# We don't apply softmax--tf.nn.sparse_softmax_cross_entropy_with_logits accepts the unscaled logits
# and performs the softmax internally for efficiency
with tf.variable_scope('softmax_linear', reuse=reuse) as scope:
weights = tf.get_variable(shape=[dim3, len(labels_id)],
initializer=init_type(),
name='weights')
biases = tf.get_variable(initializer=tf.constant(
0.0, shape=[len(labels_id)]),
name='biases')
# weight perturbation
if perturbation_type == 0:
weights = pt.weight_knockout(weights, perturbation_params[1][4])
if perturbation_type == 1:
weights = pt.weight_noise(weights, perturbation_params[0][4])
parameters += [weights, biases]
summ.variable_summaries(weights, biases, opt)
fc5 = tf.add(tf.matmul(fc4, weights), biases, name=scope.name)
summ.activation_summaries(fc5, opt)
return fc5, parameters