class IdentityBlock:
def __init__(self, mi, fm_sizes, activation=tf.nn.relu):
# conv1, conv2, conv3
# note: # feature maps shortcut = # feauture maps conv 3
assert (len(fm_sizes) == 3)
# note: kernel size in 2nd conv is always 3
# so we won't bother including it as an arg
self.session = None
self.f = tf.nn.relu
# init main branch
# Conv -> BN -> F() ---> Conv -> BN -> F() ---> Conv -> BN
self.conv1 = ConvLayer(1, mi, fm_sizes[0], 1)
self.bn1 = BatchNormLayer(fm_sizes[0])
self.conv2 = ConvLayer(3, fm_sizes[0], fm_sizes[1], 1, 'SAME')
self.bn2 = BatchNormLayer(fm_sizes[1])
self.conv3 = ConvLayer(1, fm_sizes[1], fm_sizes[2], 1)
self.bn3 = BatchNormLayer(fm_sizes[2])
# in case needed later
self.layers = [
self.conv1,
self.bn1,
self.conv2,
self.bn2,
self.conv3,
self.bn3,
]
# this will not be used when input passed in from
# a previous layer
self.input_ = tf.placeholder(tf.float32, shape=(1, 224, 224, mi))
self.output = self.forward(self.input_)
def forward(self, X):
# main branch
FX = self.conv1.forward(X)
FX = self.bn1.forward(FX)
FX = self.f(FX)
FX = self.conv2.forward(FX)
FX = self.bn2.forward(FX)
FX = self.f(FX)
FX = self.conv3.forward(FX)
FX = self.bn3.forward(FX)
return self.f(FX + X)
def predict(self, X):
assert (self.session is not None)
return self.session.run(self.output, feed_dict={self.input_: X})
def set_session(self, session):
# need to make this a session
# so assignment happens on sublayers too
self.session = session
self.conv1.session = session
self.bn1.session = session
self.conv2.session = session
self.bn2.session = session
self.conv3.session = session
self.bn3.session = session
def copyFromKerasLayers(self, layers):
assert (len(layers) == 10)
# <keras.layers.convolutional.Conv2D at 0x7fa44255ff28>,
# <keras.layers.normalization.BatchNormalization at 0x7fa44250e7b8>,
# <keras.layers.core.Activation at 0x7fa44252d9e8>,
# <keras.layers.convolutional.Conv2D at 0x7fa44253af60>,
# <keras.layers.normalization.BatchNormalization at 0x7fa4424e4f60>,
# <keras.layers.core.Activation at 0x7fa442494828>,
# <keras.layers.convolutional.Conv2D at 0x7fa4424a2da0>,
# <keras.layers.normalization.BatchNormalization at 0x7fa44244eda0>,
# <keras.layers.merge.Add at 0x7fa44245d5c0>,
# <keras.layers.core.Activation at 0x7fa44240aba8>
self.conv1.copyFromKerasLayers(layers[0])
self.bn1.copyFromKerasLayers(layers[1])
self.conv2.copyFromKerasLayers(layers[3])
self.bn2.copyFromKerasLayers(layers[4])
self.conv3.copyFromKerasLayers(layers[6])
self.bn3.copyFromKerasLayers(layers[7])
def get_params(self):
params = []
for layer in self.layers:
params += layer.get_params()
return params
class IdentityBlock(keras.layers.Layer):
def __init__(self, mi, fm_sizes, activation=tf.compat.v1.nn.relu):
super(IdentityBlock, self).__init__()
# conv1, conv2, conv3
# note: # feature maps shortcut = # feauture maps conv 3
assert (len(fm_sizes) == 3)
# note: kernel size in 2nd conv is always 3
# so we won't bother including it as an arg
self.session = None
self.f = tf.compat.v1.nn.relu
# init main branch
# Conv -> BN -> F() ---> Conv -> BN -> F() ---> Conv -> BN
self.conv1 = ConvLayer(1, mi, fm_sizes[0], 1)
self.bn1 = BatchNormLayer(fm_sizes[0])
self.conv2 = ConvLayer(3, fm_sizes[0], fm_sizes[1], 1)
self.bn2 = BatchNormLayer(fm_sizes[1])
self.conv3 = ConvLayer(1, fm_sizes[1], fm_sizes[2], 1)
self.bn3 = BatchNormLayer(fm_sizes[2])
# in case needed later
self.layers = [
self.conv1,
self.bn1,
self.conv2,
self.bn2,
self.conv3,
self.bn3,
]
def call(self, X):
FX = self.conv1(X)
FX = self.bn1(FX)
FX = self.f(FX)
FX = self.conv2(FX)
FX = self.bn2(FX)
FX = self.f(FX)
FX = self.conv3(FX)
FX = self.bn3(FX)
# sum + activation
Y = self.f(tf.math.add(FX, X))
return Y
def copyFromKerasLayers(self, layers):
#assert(len(layers) == 10)
# <keras.layers.convolutional.Conv2D at 0x7fa44255ff28>,
# <keras.layers.normalization.BatchNormalization at 0x7fa44250e7b8>,
# <keras.layers.core.Activation at 0x7fa44252d9e8>,
# <keras.layers.convolutional.Conv2D at 0x7fa44253af60>,
# <keras.layers.normalization.BatchNormalization at 0x7fa4424e4f60>,
# <keras.layers.core.Activation at 0x7fa442494828>,
# <keras.layers.convolutional.Conv2D at 0x7fa4424a2da0>,
# <keras.layers.normalization.BatchNormalization at 0x7fa44244eda0>,
# <keras.layers.merge.Add at 0x7fa44245d5c0>,
# <keras.layers.core.Activation at 0x7fa44240aba8>
self.conv1.copyFromKerasLayers(layers[0])
self.bn1.copyFromKerasLayers(layers[1])
self.conv2.copyFromKerasLayers(layers[3])
self.bn2.copyFromKerasLayers(layers[4])
self.conv3.copyFromKerasLayers(layers[6])
self.bn3.copyFromKerasLayers(layers[7])
def get_params(self):
params = []
for layer in self.layers:
params += layer.get_params()
return params
class IdentityBlock:
def __init__(self, mi, fm_sizes, activation=tf.nn.relu):
assert (len(fm_sizes) == 3)
self.session = None
self.f = tf.nn.relu
self.conv1 = ConvLayer(1, mi, fm_sizes[0], 1)
self.bn1 = BatchNormLayer(fm_sizes[0])
self.conv2 = ConvLayer(3, fm_sizes[0], fm_sizes[1], 1, 'SAME')
self.bn2 = BatchNormLayer(fm_sizes[1])
self.conv3 = ConvLayer(1, fm_sizes[1], fm_sizes[2], 1)
self.bn3 = BatchNormLayer(fm_sizes[2])
self.layers = [
self.conv1,
self.bn1,
self.conv2,
self.bn2,
self.conv3,
self.bn3,
]
self.input_ = tf.placeholder(tf.float32, shape=(1, 224, 224, mi))
self.output = self.forward(self.input_)
def forward(self, X):
# main branch
FX = self.conv1.forward(X)
FX = self.bn1.forward(FX)
FX = self.f(FX)
FX = self.conv2.forward(FX)
FX = self.bn2.forward(FX)
FX = self.f(FX)
FX = self.conv3.forward(FX)
FX = self.bn3.forward(FX)
return self.f(FX + X)
def predict(self, X):
assert (self.session is not None)
return self.session.run(self.output, feed_dict={self.input_: X})
def set_session(self, session):
self.session = session
self.conv1.session = session
self.bn1.session = session
self.conv2.session = session
self.bn2.session = session
self.conv3.session = session
self.bn3.session = session
def copyFromKerasLayers(self, layers):
assert (len(layers) == 10)
self.conv1.copyFromKerasLayers(layers[0])
self.bn1.copyFromKerasLayers(layers[1])
self.conv2.copyFromKerasLayers(layers[3])
self.bn2.copyFromKerasLayers(layers[4])
self.conv3.copyFromKerasLayers(layers[6])
self.bn3.copyFromKerasLayers(layers[7])
def get_params(self):
params = []
for layer in self.layers:
params += layer.get_params()
return params
class IdentityBlock:
def __init__(self, c_in, fm_sizes, activation=tf.nn.relu):
# and IdentityBlock consists of 3 ConvLayers:
# conv1, conv2, conv3
assert(len(fm_sizes) == 3)
self.session = None
self.activation = activation
# note: stride is always 1
# init main branch:
# Conv -> BN -> activation ---> Conv -> BN -> activation ---> Conv -> BN
self.conv1 = ConvLayer((1, 1), c_in, fm_sizes[0], stride=1)
self.bn1 = BatchNormLayer(fm_sizes[0])
self.conv2 = ConvLayer((3, 3), fm_sizes[0], fm_sizes[1], stride=1, padding='SAME')
self.bn2 = BatchNormLayer(fm_sizes[1])
self.conv3 = ConvLayer((1, 1), fm_sizes[1], fm_sizes[2], stride=1)
self.bn3 = BatchNormLayer(fm_sizes[2])
# for later use:
self.layers = [
self.conv1, self.bn1,
self.conv2, self.bn2,
self.conv3, self.bn3,
]
# next lines won't be used when using whole ResNet:
self.input_ = tf.placeholder(tf.float32, shape=(1, 224, 224, c_in))
self.output = self.forward(self.input_)
def forward(self, X):
# main branch:
fX = self.conv1.forward(X)
fX = self.bn1.forward(fX)
fX = self.activation(fX)
fX = self.conv2.forward(fX)
fX = self.bn2.forward(fX)
fX = self.activation(fX)
fX = self.conv3.forward(fX)
fX = self.bn3.forward(fX)
# shortcut is just input data X
return self.activation(fX + X)
def predict(self, X):
# we need to run the prediction in a session:
assert(self.session is not None)
return self.session.run(
self.output,
feed_dict={self.input_: X}
)
def set_session(self, session):
# we need to set a session on every layer/sub-layer:
self.session = session
self.conv1.session = session
self.bn1.session = session
self.conv2.session = session
self.bn2.session = session
self.conv3.session = session
self.bn3.session = session
def copyFromKerasLayers(self, layers):
assert(len(layers) == 10)
# <keras.layers.convolutional.Conv2D at 0x7fa44255ff28>,
# <keras.layers.normalization.BatchNormalization at 0x7fa44250e7b8>,
# <keras.layers.core.Activation at 0x7fa44252d9e8>,
# <keras.layers.convolutional.Conv2D at 0x7fa44253af60>,
# <keras.layers.normalization.BatchNormalization at 0x7fa4424e4f60>,
# <keras.layers.core.Activation at 0x7fa442494828>,
# <keras.layers.convolutional.Conv2D at 0x7fa4424a2da0>,
# <keras.layers.normalization.BatchNormalization at 0x7fa44244eda0>,
# <keras.layers.merge.Add at 0x7fa44245d5c0>,
# <keras.layers.core.Activation at 0x7fa44240aba8>
self.conv1.copyFromKerasLayers(layers[0])
self.bn1.copyFromKerasLayers(layers[1])
self.conv2.copyFromKerasLayers(layers[3])
self.bn2.copyFromKerasLayers(layers[4])
self.conv3.copyFromKerasLayers(layers[6])
self.bn3.copyFromKerasLayers(layers[7])
def get_params(self):
params = []
for layer in self.layers:
params += layer.get_params()
return params
class IdentityBlock:
def __init__(self, mi, fm_sizes, activation=tf.nn.relu):
# conv1, conv2, conv3
# note: # feature maps shortcut = # feauture maps conv 3
assert(len(fm_sizes) == 3)
# note: kernel size in 2nd conv is always 3
# so we won't bother including it as an arg
self.session = None
self.f = tf.nn.relu
# init main branch
# Conv -> BN -> F() ---> Conv -> BN -> F() ---> Conv -> BN
self.conv1 = ConvLayer(1, mi, fm_sizes[0], 1)
self.bn1 = BatchNormLayer(fm_sizes[0])
self.conv2 = ConvLayer(3, fm_sizes[0], fm_sizes[1], 1, 'SAME')
self.bn2 = BatchNormLayer(fm_sizes[1])
self.conv3 = ConvLayer(1, fm_sizes[1], fm_sizes[2], 1)
self.bn3 = BatchNormLayer(fm_sizes[2])
# in case needed later
self.layers = [
self.conv1, self.bn1,
self.conv2, self.bn2,
self.conv3, self.bn3,
]
# this will not be used when input passed in from
# a previous layer
self.input_ = tf.placeholder(tf.float32, shape=(1, 224, 224, mi))
self.output = self.forward(self.input_)
def forward(self, X):
# main branch
FX = self.conv1.forward(X)
FX = self.bn1.forward(FX)
FX = self.f(FX)
FX = self.conv2.forward(FX)
FX = self.bn2.forward(FX)
FX = self.f(FX)
FX = self.conv3.forward(FX)
FX = self.bn3.forward(FX)
return self.f(FX + X)
def predict(self, X):
assert(self.session is not None)
return self.session.run(
self.output,
feed_dict={self.input_: X}
)
def set_session(self, session):
# need to make this a session
# so assignment happens on sublayers too
self.session = session
self.conv1.session = session
self.bn1.session = session
self.conv2.session = session
self.bn2.session = session
self.conv3.session = session
self.bn3.session = session
def copyFromKerasLayers(self, layers):
assert(len(layers) == 10)
# <keras.layers.convolutional.Conv2D at 0x7fa44255ff28>,
# <keras.layers.normalization.BatchNormalization at 0x7fa44250e7b8>,
# <keras.layers.core.Activation at 0x7fa44252d9e8>,
# <keras.layers.convolutional.Conv2D at 0x7fa44253af60>,
# <keras.layers.normalization.BatchNormalization at 0x7fa4424e4f60>,
# <keras.layers.core.Activation at 0x7fa442494828>,
# <keras.layers.convolutional.Conv2D at 0x7fa4424a2da0>,
# <keras.layers.normalization.BatchNormalization at 0x7fa44244eda0>,
# <keras.layers.merge.Add at 0x7fa44245d5c0>,
# <keras.layers.core.Activation at 0x7fa44240aba8>
self.conv1.copyFromKerasLayers(layers[0])
self.bn1.copyFromKerasLayers(layers[1])
self.conv2.copyFromKerasLayers(layers[3])
self.bn2.copyFromKerasLayers(layers[4])
self.conv3.copyFromKerasLayers(layers[6])
self.bn3.copyFromKerasLayers(layers[7])
def get_params(self):
params = []
for layer in self.layers:
params += layer.get_params()
return params