NetGraph

NetGraph[{layer₁,layer₂,…},{m₁n₁,m₂n₂,…}]

specifies a neural net defined by a graph in which the output of layer m_i is given as input to layer n_i.

NetGraph["name₁"layer₁,"name₂"layer₂,…,{"name_m1""name_n1",…}]

specifies a net with explicitly named layers.

Details

For a net with a single input port, NetGraph[…][data] gives the result of applying the net to data.
For a net with multiple input ports, NetGraph[…][<|port₁->data₁,…|>] provides data to each port.
For a net with a single output port, NetGraph[…][data] gives the output for that port.
For a net with multiple output ports, NetGraph[…][data,oport] gives the output for the output port named oport. NetGraph[…][data] or NetGraph[…][data,All] gives an association of the outputs for all ports.
NetGraph[…][data,NetPortGradient[iport]] gives the gradient of the output with respect to the value of the input port iport.
Normal[NetGraph[…]] will return a list or association of the layers used to construct the graph.
If one or more input or output ports of any layers are left unconnected, these will become ports of the entire NetGraph.
If multiple output ports of layers are left unconnected and share the same name, they will become separate ports of the entire NetGraph with names "Port1", "Port2", etc.
Input or output ports for the entire NetGraph can be created by specifying NetPort["input"]->… or …->NetPort["output"] in the list of connections.
If the n layer has more than one input port or more than one output port, these can be disambiguated by writing NetPort[{n,"port"}] or NetPort[n,port].
If a layer has a port that accepts multiple inputs, such as CatenateLayer or SummationLayer, multiple connections can be made simultaneously by writing {m₁,m₂,…}->n, which is equivalent to …,m₁->n,…,m₂->n,…. The inputs m_i are always passed to n in the order m₁,m₂,….
A linear chain of connections within the graph can be specified as layer₁->layer₂->…->layer_n, which will cause each layer_i to be connected to layer_i+1.
When ambiguous, the tensor shapes of input and output ports of the entire graph can be specified with options of the form "port"shape. Valid shapes include:

	"Real"	a single real number
	"Integer"	a single integer
	n	a vector of length n
	{n₁,n₂,…}	a tensor of dimensions n₁×n₂×…
	"Varying"	a variable-length vector
	{"Varying",n₂,n₃,…}	a variable-length sequence of tensors of dimensions n₂×n₃×…
	NetEncoder[…]	an encoder (for input ports)
	NetDecoder[…]	a decoder (for output ports)
	"type"	NetEncoder["type"] or NetDecoder["type"]
	{n,coder}	an encoder or decoder mapped over a sequence of length n

NetGraph supports the following special layer specifications when giving individual layers:
Ramp,LogisticSigmoid,… ElementwiseLayer[f]

Plus,Times,Divide,… ThreadingLayer[f]

{layer₁,layer₂,…} NetChain[{layer₁,layer₂,…}]
The StandardForm of NetGraph shows the connectivity of layers in the graph and annotates edges with the dimensions of the tensor that the edge represents. Clicking a layer in the graph shows more information about that layer.
The TraditionalForm of NetGraph shows a more publication-appropriate depiction of the graph.
Take[NetGraph[…],{start,end}] returns a subgraph of the given NetGraph that contains only the layers that connect start and end. The following forms can be given for start and end:

	n,"layer"	a specified layer
	NetPort[layer,"port"]	the specified input or output port of a layer
	NetPort["port"]	an input or output port of the entire graph
	All	all of the inputs or outputs of the graph
	{spec₁,spec₂,…}	the union of the spec_i

VertexDelete[NetGraph[…],layer] deletes one or more layers from a NetGraph, returning a new graph. Layers such as ElementwiseLayer[…] that have the same input and output size will be removed such that their output is connected directly to their input.
NetGraph[…][data,opts] specifies that options should be used in applying the net to data. Possible options include:
NetEvaluationMode "Test" what mode to use in performing evaluation

TargetDevice "CPU" the target device on which to perform evaluation
With the setting NetEvaluationMode->"Training", layers such as DropoutLayer will behave as they do for training rather than ordinary evaluation.
NetGraph[…][[spec]] extracts the layer specified by spec from the net.