def residual_block_3d(input,
block_function,
filters,
repetitions,
kernel_regularizer,
is_first_layer=False,
scope=''):
for i in range(repetitions):
with tf.name_scope(scope + '_' + str(i)):
strides = (1, 1, 1)
if i == 0 and not is_first_layer:
strides = (1, 2, 2)
input = block_function(input,
filters=filters,
strides=strides,
kernel_regularizer=kernel_regularizer)
_ = print_layer_details(scope + '_' + str(i), input.get_shape())
return input
def inception_v4(x_):
with tf.name_scope("Stem"):
x_max = tf.layers.max_pooling3d(x_,
pool_size=[1, 2, 2],
strides=(1, 2, 2))
x_ = stem(x_max, "Stem")
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
with tf.name_scope("Inception_A"):
shortcut = x_
for i in range(1):
x_ = inception_a(x_, "Inception_A" + str(i))
# map identity with [1,1,1] kernel to match dimensions
shortcut = tf.layers.conv3d(
shortcut,
filters=x_.get_shape()[4],
kernel_size=[1, 1, 1],
strides=(1, 1, 1),
padding='Valid',
kernel_regularizer=tf.keras.regularizers.l2(1e-4),
activation=tf.nn.relu)
# [shortcut matrix, residual matrix]
x_ = tf.keras.layers.add([shortcut, x_])
x_ = tf.nn.relu(x_)
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
with tf.name_scope("Reduction_A"):
x_ = reduction_a(x_, "Reduction_A")
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
with tf.name_scope("Inception_B"):
shortcut = x_
for i in range(1):
x_ = inception_b(x_, "Inception_B")
shortcut = tf.layers.conv3d(
shortcut,
filters=x_.get_shape()[4],
kernel_size=[1, 1, 1],
strides=(1, 1, 1),
padding='Valid',
kernel_regularizer=tf.keras.regularizers.l2(1e-4),
activation=tf.nn.relu)
x_ = tf.keras.layers.add([shortcut, x_])
x_ = tf.nn.relu(x_)
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
with tf.name_scope("Reduction_B"):
x_ = reduction_b(x_, "Reduction_B")
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
with tf.name_scope("Inception_C"):
shortcut = x_
for i in range(1):
x_ = inception_c(x_, "Inception_C")
shortcut = tf.layers.conv3d(
shortcut,
filters=x_.get_shape()[4],
kernel_size=[1, 1, 1],
strides=(1, 1, 1),
padding='Valid',
kernel_regularizer=tf.keras.regularizers.l2(1e-4),
activation=tf.nn.relu)
x_ = tf.keras.layers.add([shortcut, x_])
x_ = tf.nn.relu(x_)
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
# with tf.name_scope("Reduction_Conv"):
# x_ = tf.layers.conv3d(x_, filters=32, kernel_size=[1,1,1], strides=(1,1,1), padding='Valid', kernel_regularizer=tf.keras.regularizers.l2(1e-4))
# print_layer_details(tf.contrib.framework.get_name_scope(), x_.get_shape())
with tf.name_scope("Final_Layer"):
width = int(int(x_.get_shape()[2]) / 2)
height = int(int(x_.get_shape()[3]) / 2)
x_avg = tf.layers.average_pooling3d(x_,
pool_size=[1, width, height],
strides=(1, width, height))
x_flat = tf.layers.flatten(inputs=x_avg)
x_dense = tf.layers.dense(
inputs=x_flat,
units=64,
activation=tf.nn.relu,
kernel_regularizer=tf.keras.regularizers.l2(1e-4))
y_ = tf.layers.dense(inputs=x_dense,
units=3,
kernel_regularizer=tf.keras.regularizers.l2(1e-4))
print_layer_details(tf.contrib.framework.get_name_scope(),
y_.get_shape())
return y_
def inception_v4_se(x_):
#reduction ratio for se_layers
rr = 4
with tf.name_scope("Stem"):
x_ = tf.layers.max_pooling3d(x_,
pool_size=[1, 2, 2],
strides=(1, 2, 2))
x_ = stem(x_, "Stem")
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
with tf.name_scope("Inception_A"):
for i in range(1):
x_ = inception_a(x_, "Inception_A")
channel = int(np.shape(x_)[-1])
x_ = se_layer(x_, out_dim=channel, ratio=rr, scope="SE_A" + str(i))
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
with tf.name_scope("Reduction_A"):
x_ = reduction_a(x_, "Reduction_A")
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
with tf.name_scope("Inception_B"):
for i in range(1):
x_ = inception_b(x_, "Inception_B")
channel = int(np.shape(x_)[-1])
x_ = se_layer(x_, out_dim=channel, ratio=rr, scope="SE_B" + str(i))
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
with tf.name_scope("Reduction_B"):
x_ = reduction_b(x_, "Reduction_B")
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
with tf.name_scope("Inception_C"):
for i in range(1):
x_ = inception_c(x_, "Inception_C")
channel = int(np.shape(x_)[-1])
x_ = se_layer(x_, out_dim=channel, ratio=rr, scope="SE_C" + str(i))
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
with tf.name_scope("Final_Layer"):
width = int(int(x_.get_shape()[2]) / 2)
height = int(int(x_.get_shape()[3]) / 2)
x_avg = tf.layers.average_pooling3d(x_,
pool_size=[1, width, height],
strides=(1, width, height))
x_flat = tf.layers.flatten(inputs=x_avg)
x_dense = tf.layers.dense(
inputs=x_flat,
units=64,
kernel_regularizer=tf.keras.regularizers.l2(1e-4))
print_layer_details(tf.contrib.framework.get_name_scope(),
x_dense.get_shape())
y_ = tf.layers.dense(inputs=x_dense,
units=3,
kernel_regularizer=tf.keras.regularizers.l2(1e-4))
return y_
def all_conv(x_):
CONV = 6
num_classes = 3
# # # Summary
# Total Number of Vars = 1.370.817
# # # Layer A
# input 8 x 84 x 84 x 1
# output 8 x 84 x 84 x 96
# number of variables = 84192
### Conv1 = 1 x 3 x 3 x 96 + 96 = 960
### Conv2 = 96 x 3 x 3 x 96 + 96 = 83040
### Batchnorm = Filter * 2 = 192
with tf.name_scope("Layer_A"):
x_ = Conv(x_,
filters=96,
kernel_size=[1, 3, 3],
strides=(1, 1, 1),
padding='same')
x_ = Conv(x_,
filters=96,
kernel_size=[1, 3, 3],
strides=(1, 1, 1),
padding='same')
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
# # # Layer B
# input 8 x 84 x 84 x 96
# output 8 x 41 x 41 x 96
# number of variables = 83232
### Conv1 = 96 x 3 x 3 x 96 + 96 = 83040
### Batchnorm = Filter * 2 = 192
with tf.name_scope("Layer_B"):
x_ = Conv(x_, filters=96, kernel_size=[1, 3, 3], strides=(1, 2, 2))
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
# # # Layer C_1 & C_2
# input 8 x 41 x 41 x 96
# output 8 x 41 x 41 x 192
# number of variables = 498.816
### Conv1 = 96 x 3 x 3 x 192 + 192 = 166080
### Conv2 = 192 x 3 x 3 x 192 + 192 = 331968
### Batchnorm = Filter * 4 = 768
for n in range(2):
with tf.name_scope("Layer_C_" + str(n)):
x_ = Conv(x_,
filters=192,
kernel_size=[1, 3, 3],
strides=(1, 1, 1),
padding='same')
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
# # # Layer D
# input 8 x 41 x 41 x 192
# output 8 x 20 x 20 x 192
# number of variables = 332352
### Conv1 = 192 x 3 x 3 x 192 + 192 = 331968
### Batchnorm = Filter * 2 = 384
with tf.name_scope("Layer_D"):
x_ = Conv(x_, filters=192, kernel_size=[1, 3, 3], strides=(1, 2, 2))
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
# # # Layer E
# input 8 x 20 x 20 x 192
# output 8 x 19 x 19 x 192
# number of variables = 332352
### Conv1 = 192 x 3 x 3 x 192 + 192 = 331968
### Batchnorm = Filter * 2 = 384
with tf.name_scope("Layer_E"):
x_ = Conv(x_,
filters=192,
kernel_size=[1, 3, 3],
strides=(1, 1, 1),
padding='same')
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
# # # Layer F
# input 8 x 19 x 19 x 192
# output 8 x 19 x 19 x 192
# number of variables = 37440
### Conv1 = 192 x 1 x 1 x 192 + 192 = 37056
### Batchnorm = Filter * 2 = 384
with tf.name_scope("Layer_F"):
x_ = Conv(x_, filters=192, kernel_size=[1, 1, 1], strides=(1, 1, 1))
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
# # # Dim Red Layer
# input 8 x 19 x 19 x 192
# output 8 x 19 x 19 x 10
# number of variables = 1950
### Conv1 = 192 x 1 x 1 x 10 + 10 = 1930
### Batchnorm = Filter * 2 = 20
with tf.name_scope("Dimensional_Red_Layer"):
x_ = Conv(x_,
filters=10,
kernel_size=[1, 1, 1],
strides=(1, 1, 1),
padding='same')
print_layer_details(tf.contrib.framework.get_name_scope(),
x_.get_shape())
# # # Final Layer
# output (?, 3)
### Dense Input = 8 * 2 * 1 * 10 = 160 Nodes
### FCL = 160 * 3 + 3 = 483
with tf.name_scope("Final_Layer"):
width = int(int(x_.get_shape()[2]) / 2)
height = int(int(x_.get_shape()[3]) / 2)
x_avg = tf.layers.average_pooling3d(x_,
pool_size=[1, width, height],
strides=(1, width, height))
x_ = tf.layers.flatten(inputs=x_avg)
y_ = tf.layers.dense(inputs=x_,
units=num_classes,
kernel_regularizer=tf.keras.regularizers.l2(1e-4))
print_layer_details(tf.contrib.framework.get_name_scope(),
y_.get_shape())
return y_