def build_model(self):
CROP_DIMS = self.config['CROP_DIMS']
C = self.config['NUM_CHANNELS']
LEAK = self.config['LEAK']
LAMBDA = self.config['L2_REG']
INIT = self.config['INIT']
NLAYERS = int(self.config['NLAYERS'] / 2)
NFILTERS_SMALL = self.config['NFILTERS_SMALL']
NFILTERS_LARGE = self.config['NFILTERS_LARGE']
NUM_POINTS = self.config['NUM_CONTOUR_POINTS']
leaky_relu = tf.contrib.keras.layers.LeakyReLU(LEAK)
self.x = tf.placeholder(shape=[None, CROP_DIMS, CROP_DIMS, C],
dtype=tf.float32)
self.y = tf.placeholder(shape=[None, NUM_POINTS], dtype=tf.float32)
self.yclass, self.yhat, _, _ = tf_util.resNet(
self.x,
nlayers_before=NLAYERS,
nlayers_after=NLAYERS,
nfilters=NFILTERS_SMALL,
nfilters_large=NFILTERS_LARGE,
output_filters=NFILTERS_LARGE,
activation=leaky_relu,
init=INIT)
o = leaky_relu(self.yhat)
d = self.config['POOL']
o = tf.nn.pool(o, [d, d], "MAX", "VALID", strides=[d, d])
s = o.get_shape().as_list()
o_vec = tf.reshape(o, shape=[-1, s[1] * s[2] * s[3]])
for i in range(self.config['FC_LAYERS']):
if "HIDDEN_SIZES" in self.config:
h = self.config['HIDDEN_SIZES'][i]
else:
h = self.config['HIDDEN_SIZE']
o_vec = tf_util.fullyConnected(o_vec,
h,
leaky_relu,
std=INIT,
scope='fc_' + str(i))
self.yhat = tf_util.fullyConnected(o_vec,
NUM_POINTS,
tf.sigmoid,
std=INIT,
scope='fc_final')
self.build_loss()
self.saver = tf.train.Saver()
def build_model(self):
CROP_DIMS = self.config['CROP_DIMS']
C = self.config['NUM_CHANNELS']
LEAK = self.config['LEAK']
LAMBDA = self.config['L2_REG']
INIT = self.config['INIT']
NLAYERS = self.config['NLAYERS']
NFILTERS = self.config['NFILTERS']
NUM_POINTS = self.config['NUM_CONTOUR_POINTS']
DIMS = [self.config['CONV_DIMS']]*2
leaky_relu = tf.contrib.keras.layers.LeakyReLU(LEAK)
self.x = tf.placeholder(shape=[None,CROP_DIMS,CROP_DIMS,C],dtype=tf.float32)
self.y = tf.placeholder(shape=[None,NUM_POINTS],dtype=tf.float32)
o = self.x
if "INPUT_POOL" in self.config:
d = self.config['INPUT_POOL']
o = tf.nn.pool(o, [d,d], "MAX", "VALID", strides=[d,d])
for i in range(NLAYERS):
o = tf_util.conv2D(o,dims=DIMS,nfilters=NFILTERS,
init=INIT,
activation=leaky_relu,scope="conv_{}".format(i))
s = o.get_shape().as_list()
o = tf.reshape(o,shape=[-1,s[1]*s[2]*s[3]])
for i in range(self.config['FC_LAYERS']):
if "HIDDEN_SIZES" in self.config:
h = self.config['HIDDEN_SIZES'][i]
else:
h = self.config['HIDDEN_SIZE']
o = tf_util.fullyConnected(o, h,
leaky_relu, std=INIT, scope='fc_'+str(i))
print(o)
if "DROPOUT" in self.config:
o = tf.nn.dropout(o, self.config['DROPOUT'])
self.yhat = tf_util.fullyConnected(o, NUM_POINTS,
tf.sigmoid, std=INIT, scope='fc_final')
self.build_loss()
self.saver = tf.train.Saver()
def build_model(self):
CROP_DIMS = self.config['CROP_DIMS']
C = self.config['NUM_CHANNELS']
LEAK = self.config['LEAK']
NUM_FILTERS = self.config['NUM_FILTERS']
LAMBDA = self.config['L2_REG']
INIT = self.config['INIT']
NUM_POINTS = self.config['NUM_CONTOUR_POINTS']
leaky_relu = tf.contrib.keras.layers.LeakyReLU(LEAK)
self.x = tf.placeholder(shape=[None, CROP_DIMS, CROP_DIMS, C],
dtype=tf.float32)
self.y = tf.placeholder(shape=[None, NUM_POINTS], dtype=tf.float32)
self.yclass, self.yhat, _, _ = tf_util.I2INet(self.x,
nfilters=NUM_FILTERS,
activation=leaky_relu,
init=INIT)
o = leaky_relu(self.yhat)
s = o.get_shape().as_list()
o_vec = tf.reshape(o, shape=[-1, s[1] * s[2] * s[3]])
for i in range(self.config['FC_LAYERS'] - 1):
if "HIDDEN_SIZES" in self.config:
h = self.config['HIDDEN_SIZES'][i]
else:
h = self.config['HIDDEN_SIZE']
o_vec = tf_util.fullyConnected(o_vec,
h,
leaky_relu,
std=INIT,
scope='fc_' + str(i))
self.yhat = tf_util.fullyConnected(o_vec,
NUM_POINTS,
tf.identity,
std=INIT,
scope='fc_final')
self.build_loss()
self.saver = tf.train.Saver()
def build_model(self):
CROP_DIMS = self.config['CROP_DIMS']
LEAK = self.config['LEAK']
C = self.config['NUM_CHANNELS']
LAMBDA = self.config['L2_REG']
INIT = self.config['INIT']
NUM_POINTS = self.config['NUM_CONTOUR_POINTS']
leaky_relu = tf.contrib.keras.layers.LeakyReLU(LEAK)
self.x = tf.placeholder(shape=[None, CROP_DIMS, CROP_DIMS, C],
dtype=tf.float32)
self.y = tf.placeholder(shape=[None, NUM_POINTS], dtype=tf.float32)
o = self.x
if "INPUT_POOL" in self.config:
d = self.config['INPUT_POOL']
o = tf.nn.pool(o, [d, d], "MAX", "VALID", strides=[d, d])
s = o.get_shape().as_list()
o_vec = tf.reshape(o, shape=[-1, s[1] * s[2] * s[3]])
for i, h in enumerate(self.config['HIDDEN_SIZES']):
o_vec = tf_util.fullyConnected(o_vec,
h,
leaky_relu,
std=INIT,
scope='fc_' + str(i))
if "DROPOUT" in self.config:
o_vec = tf.nn.dropout(o_vec, self.config['DROPOUT'])
self.yhat = tf_util.fullyConnected(o_vec,
NUM_POINTS,
tf.sigmoid,
std=INIT,
scope='fc_final')
self.build_loss()
self.saver = tf.train.Saver()
def build_model(self):
CROP_DIMS = self.config['CROP_DIMS']
C = self.config['NUM_CHANNELS']
LEAK = self.config['LEAK']
LAMBDA = self.config['L2_REG']
INIT = self.config['INIT']
NLAYERS = int(self.config['NLAYERS']/2)
NFILTERS = self.config['NFILTERS']
NUM_POINTS = self.config['NUM_CONTOUR_POINTS']
DIMS = [self.config['CONV_DIMS']]*2
leaky_relu = tf.contrib.keras.layers.LeakyReLU(LEAK)
self.x = tf.placeholder(shape=[None,CROP_DIMS,CROP_DIMS,C],dtype=tf.float32)
self.y = tf.placeholder(shape=[None,NUM_POINTS],dtype=tf.float32)
self.x_1 = tf.nn.pool(self.x, [2,2], "MAX", "VALID", strides=[2,2])
self.x_2 = tf.nn.pool(self.x_1, [2,2], "MAX", "VALID", strides=[2,2])
o = self.x
for i in range(NLAYERS):
o = tf_util.conv2D(o,dims=DIMS, nfilters=NFILTERS,init=INIT,activation=leaky_relu,
scope="conv_{}".format(i))
o_1 = self.x_1
for i in range(NLAYERS):
o_1 = tf_util.conv2D(o_1,dims=DIMS, nfilters=NFILTERS,init=INIT,activation=leaky_relu, scope="conv_1_{}".format(i))
o_2 = self.x_2
for i in range(NLAYERS):
o_2 = tf_util.conv2D(o_2,dims=DIMS, nfilters=NFILTERS,init=INIT,activation=leaky_relu, scope="conv_2_{}".format(i))
s = o.get_shape().as_list()
s_1 = o_1.get_shape().as_list()
s_2 = o_2.get_shape().as_list()
o_vec = tf.reshape(o,shape=[-1,s[1]*s[2]*s[3]])
o_vec_1 = tf.reshape(o_1,shape=[-1,s_1[1]*s_1[2]*s_1[3]])
o_vec_2 = tf.reshape(o_2,shape=[-1,s_2[1]*s_2[2]*s_2[3]])
o = tf.concat([o_vec, o_vec_1, o_vec_2], axis=1)
for i in range(self.config['FC_LAYERS']-1):
if "HIDDEN_SIZES" in self.config:
h = self.config['HIDDEN_SIZES'][i]
else:
h = self.config['HIDDEN_SIZE']
o = tf_util.fullyConnected(o, h,
leaky_relu, std=INIT, scope='fc_'+str(i))
self.yhat = tf_util.fullyConnected(o, NUM_POINTS,
tf.identity, std=INIT, scope='fc_final')
self.build_loss()
self.saver = tf.train.Saver()
def build_model(self):
CROP_DIMS = self.config['CROP_DIMS']
C = self.config['NUM_CHANNELS']
LEAK = self.config['LEAK']
LAMBDA = self.config['L2_REG']
INIT = self.config['INIT']
NLAYERS = int(self.config['NLAYERS']/2)
NFILTERS_SMALL = self.config['NFILTERS_SMALL']
NFILTERS_LARGE = self.config['NFILTERS_LARGE']
NUM_POINTS = self.config['NUM_CONTOUR_POINTS']
leaky_relu = tf.contrib.keras.layers.LeakyReLU(LEAK)
self.x = tf.placeholder(shape=[None,CROP_DIMS,CROP_DIMS,C],dtype=tf.float32)
if self.config['MULTI_TYPE'] == "POOL":
self.x_1 = tf.nn.pool(self.x, [2,2], "MAX", "VALID", strides=[2,2])
self.x_2 = tf.nn.pool(self.x_1, [2,2], "MAX", "VALID", strides=[2,2])
elif self.config['MULTI_TYPE'] == "CROP":
self.x_1 = tf.image.central_crop(self.x, central_fraction=0.5)
self.x_2 = tf.image.central_crop(self.x_1, central_fraction=0.5)
else:
raise RuntimeError("Unrecognized multi type")
self.y = tf.placeholder(shape=[None,NUM_POINTS],dtype=tf.float32)
self.yclass,self.yhat,_,_ = tf_util.resNet(self.x,
nlayers_before=NLAYERS, nlayers_after=NLAYERS,
nfilters=NFILTERS_SMALL, nfilters_large=NFILTERS_LARGE,
output_filters=NFILTERS_LARGE, activation=leaky_relu, init=INIT)
self.yclass_1,self.yhat_1,_,_ = tf_util.resNet(self.x_1,
nlayers_before=NLAYERS, nlayers_after=NLAYERS,
nfilters=NFILTERS_SMALL, nfilters_large=NFILTERS_LARGE,
output_filters=NFILTERS_LARGE, activation=leaky_relu, init=INIT,
scope="resnet_1")
self.yclass_2,self.yhat_2,_,_ = tf_util.resNet(self.x_2,
nlayers_before=NLAYERS, nlayers_after=NLAYERS,
nfilters=NFILTERS_SMALL, nfilters_large=NFILTERS_LARGE,
output_filters=NFILTERS_LARGE, activation=leaky_relu, init=INIT,
scope="resnet_2")
o = leaky_relu(self.yhat)
o_1 = leaky_relu(self.yhat_1)
o_2 = leaky_relu(self.yhat_2)
s = o.get_shape().as_list()
s_1 = o_1.get_shape().as_list()
s_2 = o_2.get_shape().as_list()
o_vec = tf.reshape(o,shape=[-1,s[1]*s[2]*s[3]])
o_vec_1 = tf.reshape(o_1,shape=[-1,s_1[1]*s_1[2]*s_1[3]])
o_vec_2 = tf.reshape(o_2,shape=[-1,s_2[1]*s_2[2]*s_2[3]])
o = tf.concat([o_vec, o_vec_1, o_vec_2], axis=1)
print(o)
for i in range(self.config['FC_LAYERS']-1):
if "HIDDEN_SIZES" in self.config:
h = self.config['HIDDEN_SIZES'][i]
else:
h = self.config['HIDDEN_SIZE']
o = tf_util.fullyConnected(o, h,
leaky_relu, std=INIT, scope='fc_'+str(i))
self.yhat = tf_util.fullyConnected(o, NUM_POINTS,
tf.identity, std=INIT, scope='fc_final')
self.build_loss()
self.saver = tf.train.Saver()