sciml.interpreter¶

The interpreter of the forward pass of a network.

The forward pass of the NN YAML is the list of the nodes of a torch.fx graph: placeholder (an input), call_module (a layer), call_function and call_method (a function), and output. evaluate walks the list once with a backend, so the forward pass in numpy and the compilation into expressions are the same interpreter and the same layers.

The NN YAML writes a reference to a node as the name of the node, so a reference and a string literal look the same. The arguments of a layer and of an output are references only. The arguments of a function are references where the implementation takes arrays, i.e. where its parameter is annotated with numpy.ndarray (or a type which holds it, e.g. a sequence of arrays), and literals everywhere else, so gelu(x, "tanh") stays the literal "tanh" in a network with a node named tanh.

evaluate ¶

evaluate(model, parameters, inputs, backend, on_node=None)

Evaluate the forward pass of a network.

Parameters:

Name	Type	Description	Default
`model`	`NNModel`	the architecture of the network.	required
`parameters`	`Mapping[str, Mapping[str, ndarray]]`	the arrays of the layers, layer id -> array name -> array, in the PyTorch layout.	required
`inputs`	`Sequence[ArrayLike]`	the inputs, one per `placeholder` node in the order of the nodes.	required
`backend`	`Backend`	the backend the layers and functions are evaluated with.	required
`on_node`	`NodeHook \| None`	called after every layer and function with the node and its value, the next nodes see what it returns. The compilation uses it to replace the expressions of a node by symbols.	`None`

Returns:

Type	Description
`ndarray`	The outputs of the network, arrays which share no memory with the
`...`	inputs.

Raises:

Type	Description
`UnsupportedLayerError`	if a layer or function is not implemented, not available in the backend or called with arguments its implementation does not take.
`ValueError`	if the number of inputs is not the number of placeholders, if a node has an unknown opcode, if an argument does not name a node which was evaluated or if a layer or function cannot be evaluated on its input; the message names the network and the node.