def create_cell_mlp(input_nodes):
"""Create a cell with convolution.
Args:
input_nodes (list(Node)): a list of input_nodes for this cell.
Returns:
Cell: the corresponding cell.
"""
cell = Cell(input_nodes)
n1 = ConstantNode(name='N1')
cell.graph.add_edge(input_nodes[0], n1) # fixed input connection
n1.set_op(op=Flatten())
n2 = VariableNode('N2')
n2.add_op(Identity())
n2.add_op(Dense(units=10))
n2.add_op(Dense(units=50))
n2.add_op(Dense(units=100))
n2.add_op(Dense(units=200))
n2.add_op(Dense(units=250))
n2.add_op(Dense(units=500))
n2.add_op(Dense(units=750))
n2.add_op(Dense(units=1000))
n3 = VariableNode('N3')
n3.add_op(Identity())
n3.add_op(Activation(activation='relu'))
n3.add_op(Activation(activation='tanh'))
n3.add_op(Activation(activation='sigmoid'))
n4 = VariableNode('N4')
n4.add_op(Identity())
n4.add_op(Dropout(rate=0.5))
n4.add_op(Dropout(rate=0.4))
n4.add_op(Dropout(rate=0.3))
n4.add_op(Dropout(rate=0.2))
n4.add_op(Dropout(rate=0.1))
n4.add_op(Dropout(rate=0.05))
block = Block()
block.add_node(n1)
block.add_node(n2)
block.add_node(n3)
block.add_node(n4)
block.add_edge(n1, n2)
block.add_edge(n2, n3)
block.add_edge(n3, n4)
cell.add_block(block)
cell.set_outputs()
return cell
def set_outputs(self, node=None):
"""Set output node of the current cell.
node (Node, optional): Defaults to None will create a Concatenation node for the last axis.
"""
if node is None:
stacked_nodes = self.get_blocks_output()
output_node = ConstantNode(name=f'Cell_{self.num}_Output')
output_node.set_op(
Concatenate(self.graph, output_node, stacked_nodes))
else:
output_node = node
self.output = output_node
def create_mlp_block(cell, input_node):
# first node of block
n1 = ConstantNode(op=Dense(1000, tf.nn.relu), name='N1')
cell.graph.add_edge(input_node, n1) # fixed input of current block
# second node of block
n2 = ConstantNode(op=Dense(1000, tf.nn.relu), name='N2')
n3 = ConstantNode(op=Dense(1000, tf.nn.relu), name='N3')
block = Block()
block.add_node(n1)
block.add_node(n2)
block.add_node(n3)
block.add_edge(n1, n2)
block.add_edge(n2, n3)
return block
def create_structure(input_shape=[(2, ), (2, ), (2, )],
output_shape=(1, ),
num_cell=8,
*args,
**kwargs):
network = KerasStructure(input_shape,
output_shape) #, output_op=AddByPadding)
input_nodes = network.input_nodes
# CELL 1
cell1 = create_cell_1(input_nodes)
network.add_cell(cell1)
# CELL Middle
inputs_skipco = [
input_nodes, input_nodes[0], input_nodes[1], input_nodes[2],
cell1.output
]
pred_cell = cell1
n = num_cell
for i in range(n):
cell_i = Cell(input_nodes + [cell1.output])
block1 = create_mlp_block(cell_i, pred_cell.output)
cell_i.add_block(block1)
cnode = VariableNode(name='SkipCo')
nullNode = ConstantNode(op=Tensor([]), name='None')
cnode.add_op(Connect(cell_i.graph, nullNode, cnode)) # SAME
for inpt in inputs_skipco:
cnode.add_op(Connect(cell_i.graph, inpt, cnode))
block2 = Block()
block2.add_node(cnode)
cell_i.add_block(block2)
# set_cell_output_add(cell2)
cell_i.set_outputs()
network.add_cell(cell_i)
# prep. for next iter
inputs_skipco.append(cell_i.output)
pred_cell = cell_i
# CELL LAST
cell_last = Cell([pred_cell.output])
block1 = create_mlp_block(cell_last, pred_cell.output)
cell_last.add_block(block1)
# set_cell_output_add(cell3)
cell_last.set_outputs()
network.add_cell(cell_last)
return network
def set_ops(self, indexes):
"""
Set the operations for each node of each cell of the structure.
Args:
indexes (list): element of list can be float in [0, 1] or int.
output_node (ConstantNode): the output node of the Structure.
"""
cursor = 0
for c in self.struct:
num_nodes = c.num_nodes
c.set_ops(indexes[cursor:cursor+num_nodes])
cursor += num_nodes
self.graph.add_nodes_from(c.graph.nodes())
self.graph.add_edges_from(c.graph.edges())
output_nodes = get_output_nodes(self.graph)
if len(output_nodes) == 1:
node = ConstantNode(op=Identity(), name='Structure_Output')
self.graph.add_node(node)
self.graph.add_edge(output_nodes[0], node)
else:
node = ConstantNode(name='Structure_Output')
node.set_op(self.output_op(self.graph, node, output_nodes))
self.output_node = node
def __init__(self, input_shape, output_shape, output_op=None, *args, **kwargs):
"""A KerasStructure represents a search space of neural networks.
Args:
input_shape (list(tuple(int))): list of shapes of all inputs
output_shape (tuple(int)): shape of output
output_op (Operation): operation which merges outputs of cells
Raises:
RuntimeError: [description]
"""
self.graph = nx.DiGraph()
if type(input_shape) is tuple:
# we have only one input tensor here
op = Tensor(keras.layers.Input(input_shape, name="input_0"))
self.input_nodes = [ConstantNode(op=op, name='Input_0')]
elif type(input_shape) is list and all(map(lambda x: type(x) is tuple, input_shape)):
# we have a list of input tensors here
self.input_nodes = list()
for i in range(len(input_shape)):
op = Tensor(keras.layers.Input(input_shape[i], name=f"input_{i}"))
inode = ConstantNode(op=op, name=f'Input_{i}')
self.input_nodes.append(inode)
else:
raise RuntimeError(f"input_shape must be either of type 'tuple' or 'list(tuple)' but is of type '{type(input_shape)}'!")
self.__output_shape = output_shape
self.output_node = None
self.output_op = Concatenate if output_op is None else output_op
self.struct = []
self.map_sh2int = {}
self._model = None
def create_structure(input_shape=[(2, ), (2, ), (2, )],
output_shape=(1, ),
*args,
**kwargs):
network = KerasStructure(input_shape,
output_shape) #, output_op=AddByPadding)
input_nodes = network.input_nodes
# CELL 1
cell1 = create_cell_1(input_nodes)
network.add_cell(cell1)
# CELL 2
cell2 = Cell(input_nodes + [cell1.output])
block1 = create_mlp_block(cell2, cell1.output)
cell2.add_block(block1)
cnode = VariableNode(name='SkipCo')
nullNode = ConstantNode(op=Tensor([]), name='None')
cnode.add_op(Connect(cell2.graph, nullNode, cnode)) # SAME
cnode.add_op(Connect(cell2.graph, input_nodes[0], cnode))
cnode.add_op(Connect(cell2.graph, input_nodes[1], cnode))
cnode.add_op(Connect(cell2.graph, input_nodes[2], cnode))
cnode.add_op(Connect(cell2.graph, cell1.output, cnode))
cnode.add_op(Connect(cell2.graph, input_nodes, cnode))
cnode.add_op(Connect(cell2.graph, [input_nodes[0], input_nodes[1]], cnode))
cnode.add_op(Connect(cell2.graph, [input_nodes[1], input_nodes[2]], cnode))
cnode.add_op(Connect(cell2.graph, [input_nodes[0], input_nodes[2]], cnode))
block2 = Block()
block2.add_node(cnode)
cell2.add_block(block2)
# set_cell_output_add(cell2)
cell2.set_outputs()
network.add_cell(cell2)
# CELL 3
cell3 = Cell([cell2.output])
block1 = create_mlp_block(cell3, cell2.output)
cell3.add_block(block1)
# set_cell_output_add(cell3)
cell3.set_outputs()
network.add_cell(cell3)
return network
def create_mlp_block(cell, input_node, skipco_inputs):
# first block with one node
block1 = Block()
n1 = VariableNode('N1')
create_mlp_node(n1)
cell.graph.add_edge(input_node, n1) # fixed input of current block
block1.add_node(n1)
# second block for skipconnection
block2 = Block()
nullNode = ConstantNode(op=Tensor([]), name='None')
cnode = VariableNode(name='SkipCo')
cnode.add_op(Connect(cell.graph, nullNode, cnode)) # SAME
for n in skipco_inputs:
cnode.add_op(Connect(cell.graph, n, cnode))
block2.add_node(cnode)
cell.add_block(block1)
cell.add_block(block2)
return n1 # return dense to append it in the skipco_inputs list
def create_mlp_block(cell, input_node):
block = Block()
# first node of block
n1 = VariableNode('N1')
create_mlp_node(n1)
cell.graph.add_edge(input_node, n1) # fixed input of current block
block.add_node(n1)
# second node of block
n2 = VariableNode('N2')
create_mlp_node(n2)
block.add_node(n2)
addNode1 = ConstantNode(name='Merging')
addNode1.set_op(AddByPadding()) # edge created here
block.add_node(addNode1)
block.add_edge(n1, n2)
block.add_edge(n2, addNode1)
block.add_edge(n1, addNode1) # residual connection
n3 = VariableNode('N3')
create_mlp_node(n3)
block.add_node(n3)
block.add_edge(addNode1, n3)
addNode2 = ConstantNode(name='Merging')
addNode2.set_op(AddByPadding()) # edge created here
block.add_node(addNode2)
block.add_edge(n3, addNode2)
block.add_edge(addNode1, addNode2)
cell.add_block(block)
return n1
def create_cell_1(input_nodes):
"""Create a cell with convolution.
Args:
input_nodes (list(Node)): a list of input_nodes for this cell.
Returns:
Cell: the corresponding cell.
"""
cell = Cell(input_nodes)
input_dose1 = input_nodes[0]
input_rnaseq = input_nodes[1]
input_drug1descriptor = input_nodes[2]
input_drug1fingerprints = input_nodes[3]
def create_block_3_nodes(input_node):
# first node of block
n1 = VariableNode('N1')
create_mlp_node(n1)
cell.graph.add_edge(input_node, n1) # fixed input of current block
# second node of block
n2 = VariableNode('N2')
create_mlp_node(n2)
# third node of the block
n3 = VariableNode('N3')
create_mlp_node(n3)
block = Block()
block.add_node(n1)
block.add_node(n2)
block.add_node(n3)
block.add_edge(n1, n2)
block.add_edge(n2, n3)
return block, (n1, n2, n3)
# BLOCK FOR: dose1
n = ConstantNode(op=Identity(), name='N1', )
cell.graph.add_edge(input_dose1, n)
block0 = Block()
block0.add_node(n)
cell.add_block(block0)
# BLOCK FOR: rnaseq
block3, _ = create_block_3_nodes(input_rnaseq)
cell.add_block(block3)
# BLOCK FOR: drug1.descriptor
block4, _ = create_block_3_nodes(input_drug1descriptor)
cell.add_block(block4)
# BLOCK FOR: drug1.fingerprints
block5, _ = create_block_3_nodes(input_drug1fingerprints)
cell.add_block(block5)
# set_cell_output_add(cell)
cell.set_outputs()
return cell
def set_cell_output_add(cell):
addNode = ConstantNode(name='Merging')
addNode.set_op(AddByPadding(cell.graph, addNode, cell.get_blocks_output()))
cell.set_outputs(node=addNode)
def create_cell_conv(input_nodes):
"""Create a cell with convolution.
Args:
input_nodes (list(Node)): a list of input_nodes for this cell.
Returns:
Cell: the corresponding cell.
"""
cell = Cell(input_nodes)
n1 = ConstantNode(op=Conv1D(filter_size=20, num_filters=128), name='N1')
cell.graph.add_edge(input_nodes[0], n1) # fixed input connection
n2 = ConstantNode(op=Activation(activation='relu'), name='N2')
n3 = ConstantNode(op=MaxPooling1D(pool_size=1, padding='same'), name='N3')
n4 = ConstantNode(op=Conv1D(filter_size=10, num_filters=128),name='N4')
n5 = ConstantNode(op=Activation(activation='relu'), name='N5')
n6 = ConstantNode(op=MaxPooling1D(pool_size=10, padding='same'), name='N6')
n7 = ConstantNode(op=Flatten(), name='N7')
n8 = ConstantNode(op=Dense(units=200), name='N8')
n9 = ConstantNode(op=Activation(activation='relu'), name='N9')
n10 = ConstantNode(op=Dropout(rate=0.1), name='N10')
n11 = ConstantNode(op=Dense(units=20), name='N11')
n12 = ConstantNode(op=Activation(activation='relu'), name='N12')
n13 = ConstantNode(op=Dropout(rate=0.1), name='N13')
block = Block()
block.add_node(n1)
block.add_node(n2)
block.add_node(n3)
block.add_node(n4)
block.add_node(n5)
block.add_node(n6)
block.add_node(n7)
block.add_node(n8)
block.add_node(n9)
block.add_node(n10)
block.add_node(n11)
block.add_node(n12)
block.add_node(n13)
block.add_edge(n1, n2)
block.add_edge(n2, n3)
block.add_edge(n3, n4)
block.add_edge(n4, n5)
block.add_edge(n5, n6)
block.add_edge(n6, n7)
block.add_edge(n7, n8)
block.add_edge(n8, n9)
block.add_edge(n9, n10)
block.add_edge(n10, n11)
block.add_edge(n11, n12)
block.add_edge(n12, n13)
cell.add_block(block)
cell.set_outputs()
return cell
