Source code for nengo_dl.tensor_node

"""
TensorNodes allow parts of a model to be defined using TensorFlow and smoothly
integrated with the rest of a Nengo model.
"""

from nengo import Node, Connection, Ensemble, builder
from nengo.base import NengoObject
from nengo.builder.operator import Reset
from nengo.exceptions import ValidationError, SimulationError
from nengo.neurons import NeuronType
from nengo.params import Default, IntParam, Parameter
import numpy as np
import tensorflow as tf

from nengo_dl.builder import Builder, OpBuilder, NengoBuilder


def validate_output(output, minibatch_size=None, output_d=None, dtype=None):
    """
    Performs validation on the output of a TensorNode ``tensor_func``.

    Parameters
    ----------
    output : ``tf.Tensor``
        Output from the ``tensor_func``.
    minibatch_size : int
        Expected minibatch size for the simulation.
    output_d
        Expected output dimensionality for the function.
    dtype
        Expected dtype of the function output.
    """

    if not isinstance(output, tf.Tensor):
        raise ValidationError(
            "TensorNode function must return a Tensor (got %s)" % type(output),
            attr="tensor_func",
        )

    shape = output.get_shape()
    if (
        shape.ndims != 2
        or (minibatch_size is not None and shape[0] != minibatch_size)
        or (output_d is not None and shape[1] != output_d)
    ):
        raise ValidationError(
            "TensorNode output should have shape (%s, %s) "
            "(got shape %s)" % (minibatch_size, output_d, output.get_shape()),
            attr="tensor_func",
        )

    if dtype is not None and output.dtype != dtype:
        raise ValidationError(
            "TensorNode output should have dtype %s "
            "(got %s)" % (dtype, output.dtype),
            attr="tensor_func",
        )


class TensorFuncParam(Parameter):
    """Parameter for the ``tensor_func`` parameter of a `.TensorNode`."""

    def __init__(self, name, readonly=False):
        super(TensorFuncParam, self).__init__(name, optional=False, readonly=readonly)

    def coerce(self, node, func):
        """
        Performs validation on the function passed to TensorNode, and sets
        ``size_out`` if necessary.

        Parameters
        ----------
        node : `.TensorNode`
            The node whose ``tensor_func`` parameter is being set.
        func : callable
            The function being assigned to the TensorNode.

        Returns
        -------
        output : callable
            The function after validation is applied.
        """

        output = super(TensorFuncParam, self).coerce(node, func)

        if node.size_out is None:
            if not callable(func):
                raise ValidationError(
                    "TensorNode output must be a function", attr=self.name, obj=node
                )

            with tf.Graph().as_default():
                t, x = tf.constant(0.0), tf.zeros((1, node.size_in))
                args = (t, x) if node.size_in > 0 else (t,)
                try:
                    result = func(*args)
                except Exception as e:
                    raise ValidationError(
                        "Calling TensorNode function with arguments %s "
                        "produced an error:\n%s" % (args, e),
                        attr=self.name,
                        obj=node,
                    )

            validate_output(result)

            node.size_out = result.get_shape()[1].value

        return output


[docs]class TensorNode(Node): """ Inserts TensorFlow code into a Nengo model. Parameters ---------- tensor_func : callable A function that maps node inputs to outputs size_in : int (Default: 0) The number of elements in the input vector size_out : int (Default: None) The number of elements in the output vector (if None, value will be inferred by calling ``tensor_func``) label : str (Default: None) A name for the node, used for debugging and visualization """ tensor_func = TensorFuncParam("tensor_func") size_in = IntParam("size_in", default=0, low=0, optional=True) size_out = IntParam("size_out", default=None, low=1, optional=True) def __init__(self, tensor_func, size_in=Default, size_out=Default, label=Default): # pylint: disable=non-parent-init-called,super-init-not-called # note: we bypass the Node constructor, because we don't want to # perform validation on `output` NengoObject.__init__(self, label=label, seed=None) self.size_in = size_in self.size_out = size_out self.tensor_func = tensor_func @property def output(self): """ Ensure that nothing tries to evaluate the `output` attribute (indicating that something is trying to simulate this as a regular `nengo.Node` rather than a TensorNode. """ def output_func(*_): raise SimulationError( "Cannot call TensorNode output function (this probably means " "you are trying to use a TensorNode inside a Simulator other " "than NengoDL)" ) return output_func
[docs]@NengoBuilder.register(TensorNode) def build_tensor_node(model, node): """This is the Nengo build function, so that Nengo knows what to do with TensorNodes.""" # input signal if node.size_in > 0: sig_in = builder.Signal(np.zeros(node.size_in), name="%s.in" % node) model.add_op(Reset(sig_in)) else: sig_in = None sig_out = builder.Signal(np.zeros(node.size_out), name="%s.out" % node) model.sig[node]["in"] = sig_in model.sig[node]["out"] = sig_out model.params[node] = None model.operators.append(SimTensorNode(node.tensor_func, model.time, sig_in, sig_out))
[docs]class SimTensorNode(builder.Operator): # pylint: disable=abstract-method """Operator for TensorNodes (constructed by `.build_tensor_node`). Parameters ---------- func : callable The TensorNode function (``tensor_func``) time : `~nengo.builder.Signal` Signal representing the current simulation time input : `~nengo.builder.Signal` or None Input Signal for the TensorNode (or None if size_in==0) output : `~nengo.builder.Signal` Output Signal for the TensorNode tag : str A label associated with the operator, for debugging Notes ----- 1. sets ``[output]`` 2. incs ``[]`` 3. reads ``[time] if input is None else [time, input]`` 4. updates ``[]`` """ def __init__(self, func, time, input, output, tag=None): super(SimTensorNode, self).__init__(tag=tag) self.func = func self.input = input self.output = output self.sets = [output] self.incs = [] self.reads = [time] if input is None else [time, input] self.updates = []
[docs]@Builder.register(SimTensorNode) class SimTensorNodeBuilder(OpBuilder): """Builds a `~.tensor_node.SimTensorNode` operator into a NengoDL model.""" def __init__(self, ops, signals, config): super(SimTensorNodeBuilder, self).__init__(ops, signals, config) # SimTensorNodes should never be merged assert len(ops) == 1 op = ops[0] if op.input is None: self.src_data = None else: self.src_data = signals[op.input] assert self.src_data.ndim == 1 self.dst_data = signals[op.output] self.func = op.func if hasattr(self.func, "pre_build"): self.func.pre_build( None if self.src_data is None else ((signals.minibatch_size,) + self.src_data.shape), (signals.minibatch_size,) + self.dst_data.shape, )
[docs] def build_step(self, signals): if self.src_data is None: output = self.func(signals.time) else: input = signals.gather(self.src_data) # move minibatch dimension to front input = tf.transpose(a=input, perm=(1, 0)) output = self.func(signals.time, input) validate_output( output, minibatch_size=signals.minibatch_size, output_d=self.dst_data.shape[0], dtype=signals.dtype, ) # move minibatch dimension back to end output = tf.transpose(a=output, perm=(1, 0)) signals.scatter(self.dst_data, output)
[docs] def build_post(self, ops, signals, sess, rng): if hasattr(self.func, "post_build"): self.func.post_build(sess, rng)
def reshaped(shape_in): """A decorator to reshape the inputs to a function into non-vector shapes. The output of the function will be flatten back into (batched) vectors. Parameters ---------- shape_in : tuple of int The desired shape for inputs to the function (not including the first dimension, which corresponds to the batch axis) Returns ------- reshaper : callable The decorated function """ def reshape_dec(func): def reshaped_func(t, x): batch_size = x.get_shape()[0] x = tf.reshape(x, (batch_size,) + shape_in) x = func(t, x) x = tf.reshape(x, (batch_size, -1)) return x return reshaped_func return reshape_dec
[docs]def tensor_layer( input, layer_func, shape_in=None, synapse=None, transform=1, return_conn=False, # TODO: remove once there is a black release with this # bugfix https://github.com/psf/black/pull/763 # fmt: off **layer_args # fmt: on ): """A utility function to construct TensorNodes that apply some function to their input (analogous to the ``tf.layers`` syntax). Parameters ---------- input : ``NengoObject`` Object providing input to the layer layer_func : callable or `~nengo.neurons.NeuronType` A function that takes the value from ``input`` (represented as a ``tf.Tensor``) and maps it to some output value, or a Nengo neuron type, defining a nonlinearity that will be applied to ``input``. shape_in : tuple of int If not None, reshape the input to the given shape synapse : float or `~nengo.synapses.Synapse` Synapse to apply on connection from ``input`` to this layer transform : `~numpy.ndarray` Transform matrix to apply on connection from ``input`` to this layer return_conn : bool If True, also return the connection linking this layer to ``input`` layer_args : dict These arguments will be passed to ``layer_func`` if it is callable, or `~nengo.Ensemble` if ``layer_func`` is a `~nengo.neurons.NeuronType` Returns ------- node : `.TensorNode` or `~nengo.ensemble.Neurons` A TensorNode that implements the given layer function (if ``layer_func`` was a callable), or a Neuron object with the given neuron type, connected to ``input`` conn : `~nengo.Connection` If ``return_conn`` is True, also returns the connection object linking ``input`` and ``node``. """ if isinstance(transform, np.ndarray) and transform.ndim == 2: size_in = transform.shape[0] elif shape_in is not None: size_in = np.prod(shape_in) else: size_in = input.size_out if isinstance(layer_func, NeuronType): node = Ensemble(size_in, 1, neuron_type=layer_func, **layer_args).neurons else: # add (ignored) time input and pass kwargs def node_func(_, x): return layer_func(x, **layer_args) # reshape input if necessary if shape_in is not None: node_func = reshaped(shape_in)(node_func) node = TensorNode(node_func, size_in=size_in) conn = Connection(input, node, synapse=synapse, transform=transform) return (node, conn) if return_conn else node