Networks

mujou.Network

Bases: Component

A composite of nodes and other networks, declared as attributes.

Assign nodes and networks as attributes and connect their ports with Port.connect. The children of a network are the nodes and networks among its attributes, in the order of assignment: mujou.run reads them when it starts, as nn.Module finds its submodules, so nodes in a list, a tuple or a dict are not children (the run raises TypeError for them). A port assigned to a network is only a reference to it: expose the ports the host uses that way (by convention prev and next, e.g. self.prev = self.l1.prev). A network is only the structure of a run: its nodes still run independently and asynchronously.

mujou.run(net) runs the network; the Run it returns is how the host talks to the running nodes. A node of the network is at the path of its attributes from the root ("body.l1"), which names it in the state tree and in errors.

Example::

class Stack(Network):
    def __init__(self) -> None:
        super().__init__()
        self.l1 = Hidden(784, 256)
        self.l2 = Hidden(256, 256)
        self.l1.next.connect(self.l2.prev)
        self.prev = self.l1.prev  # expose the previous side
        self.next = self.l2.next  # expose the next side


class Classifier(Network):
    def __init__(self) -> None:
        super().__init__()
        self.body = Stack()
        self.head = Head()  # a node with prev, next and teacher ports
        self.body.next.connect(self.head.prev)  # l2.next <-> head.prev
        self.prev = self.body.prev  # the input
        self.teacher = self.head.teacher  # the target, a port of head


net = Classifier()
with mujou.run(net, inputs={net.prev: x0, net.teacher: y0}) as run:
    run.write({net.prev: x, net.teacher: y})  # switch together
    run.timer.sleep(0.1)  # present the input for 0.1 s of the run's time
    prediction = run.read(net.head.next)
    state = run.state_dict()

mujou.Group

Group(**children: Node | Network)

Bases: Network

A network of the given nodes and networks side by side, each at its keyword, unconnected unless they were connected before.

The node l1 of Group(enc=a) is at "enc.l1".

Example::

enc, dec = Encoder(), Decoder()
enc.next.connect(dec.prev)
with mujou.run(Group(enc=enc, dec=dec)) as run:
    state = run.state_dict()  # nodes "enc.l1", ..., "dec.l1", ...