def _forward(self):
"""
_forward() uses defined convolutional neural networks with initial input
:return:
"""
if self.task_parameter is None:
self.task_parameter = self.create_initial_parameter(
primary_outerparameter=self.outer_param_dict)
for i in range(len(self.dim_hidden)):
if self.use_t:
self + network_utils.conv_block_t(
self,
self.task_parameter["conv" + str(i)],
self.task_parameter["bias" + str(i)],
self.model_param_dict["conv" + str(i) + "_z"],
)
elif self.use_warp:
self + network_utils.conv_block_warp(
self,
self.task_parameter["conv" + str(i)],
self.task_parameter["bias" + str(i)],
self.model_param_dict["conv" + str(i) + "_z"],
self.model_param_dict["bias" + str(i) + "_z"],
)
else:
self + network_utils.conv_block(
self,
self.task_parameter["conv" + str(i)],
self.task_parameter["bias" + str(i)],
)
if self.max_pool:
self + tf.reshape(
self.out,
[-1,
np.prod([int(dim) for dim in self.out.get_shape()[1:]])])
self + tf.add(
tf.matmul(
self.out,
self.task_parameter["w" + str(len(self.dim_hidden))]),
self.task_parameter["bias" + str(len(self.dim_hidden))],
)
else:
self + tf.add(
tf.matmul(
tf.reduce_mean(self.out, [1, 2]),
self.task_parameter["w" + str(len(self.dim_hidden))],
),
self.task_parameter["bias" + str(len(self.dim_hidden))],
)
if self.use_t:
self + tf.matmul(
self.out,
self.model_param_dict["w" + str(len(self.dim_hidden)) + "_z"])
def _forward(self):
"""
for i in range(4):
self.conv_layer(filters=self.dim_hidden[i],stride=self.stride, max_pool=self.max_pool)
flattened_shape = reduce(lambda a, v: a * v, self.layers[-1].get_shape().as_list()[1:])
self + tf.reshape(self.out, shape=(-1, flattened_shape), name='representation')
"""
for i in range(len(self.dim_hidden)):
if self.use_T:
self + network_utils.conv_block_t(
self,
self.outer_param_dict["conv" + str(i)],
self.outer_param_dict["bias" + str(i)],
self.model_param_dict["conv" + str(i) + "_z"],
)
elif self.use_Warp:
self + network_utils.conv_block_warp(
self,
self.outer_param_dict["conv" + str(i)],
self.outer_param_dict["bias" + str(i)],
self.model_param_dict["conv" + str(i) + "_z"],
self.model_param_dict["bias" + str(i) + "_z"],
)
else:
self + network_utils.conv_block(
self,
self.outer_param_dict["conv" + str(i)],
self.outer_param_dict["bias" + str(i)],
)
if self.flatten:
flattened_shape = reduce(lambda a, v: a * v,
self.layers[-1].get_shape().as_list()[1:])
self + tf.reshape(
self.out, shape=(-1, flattened_shape), name="representation")
else:
if self.max_pool:
self + tf.reshape(
self.out,
[
-1,
np.prod([int(dim) for dim in self.out.get_shape()[1:]])
],
)
else:
self + tf.reduce_mean(self.out, [1, 2])
def _forward(self):
"""
_forward() uses defined convolutional neural networks with initial input
:return:
"""
for i in range(len(self.dim_hidden)):
if self.use_t:
self + network_utils.conv_block_t(
self,
self.outer_param_dict["conv" + str(i)],
self.outer_param_dict["bias" + str(i)],
self.model_param_dict["conv" + str(i) + "_z"],
)
elif self.use_warp:
self + network_utils.conv_block_warp(
self,
self.outer_param_dict["conv" + str(i)],
self.outer_param_dict["bias" + str(i)],
self.model_param_dict["conv" + str(i) + "_z"],
self.model_param_dict["bias" + str(i) + "_z"],
)
else:
self + network_utils.conv_block(
self,
self.outer_param_dict["conv" + str(i)],
self.outer_param_dict["bias" + str(i)],
)
if self.flatten:
flattened_shape = reduce(lambda a, v: a * v,
self.layers[-1].get_shape().as_list()[1:])
self + tf.reshape(
self.out, shape=(-1, flattened_shape), name="representation")
else:
if self.max_pool:
self + tf.reshape(
self.out,
[
-1,
np.prod([int(dim) for dim in self.out.get_shape()[1:]])
],
)
else:
self + tf.reduce_mean(self.out, [1, 2])