What the other nodes' inputs name this one by. Unique in the layer.
OptionallabelWhat the node is called when it is announced. Defaults to id.
OptionalpathThe scopes the node sits in, outermost first, excluding the node
itself. A node at the top level omits it or declares [].
OptionalinputsThe ids of the nodes feeding this one, in the order it takes them.
An id naming no declared node, and a node naming itself, are ignored.
OptionalinputWhat each edge in inputs is called, index for index: the Yes and No
of a flowchart's decision, a state machine's event, a relationship's
verb. null, an empty string or a missing entry leaves that edge
unnamed. An edge declared more than once carries each distinct label it
was given, in declared order.
Announced beside the node's name wherever the edge is heard -- the from and to lists, and the description's inputs and outputs -- because on a diagram the label is often the only thing that says which way a branch goes.
OptionalinputWhich drawn edge each edge in inputs is, index for index, where the
diagram draws one edge that is declared as several:
Edges sharing an id are one edge: the description counts it once, lists
the copies arriving from or leaving for one scope as that scope, and
does not call a line declared each way a cycle. Moves are unchanged, as
every declared edge is still one a reader can follow. null, an empty
string or a missing entry leaves that edge an edge of its own, apart
from any drawn edge between the same two nodes. Ids are compared as
text, so the two directions of a line must carry the same one.
OptionalattributesAnything else the reader should hear about the node, announced as
key, value clauses in declared order and tabulated in the
description: a layer type, an output shape, a parameter count.
One node of a directed graph -- one op or layer of a computation graph.
The edges are declared on the node they arrive at, as
inputs: that is how a TensorFlowGraphDefnames them (NodeDef.input) and how a Keras layer names its inbound layers, so a producer copies a field rather than inverting one. Edge order is node order, then input order -- the order the edge selectors would be in.Scopes are declared as a path, the way TreemapPoint.path declares a hierarchy: acyclic by construction, impossible to orphan, and already what a TensorFlow name scope is (
block1/conv1/Conv2Dis the path['block1', 'conv1']and the nodeConv2D).Example