from __future__ import print_function
from collections import Mapping
import datetime
import logging
import os
import time
import warnings
from nengo.builder import Model
from nengo.exceptions import (ReadonlyError, SimulatorClosed, NengoWarning,
SimulationError)
import numpy as np
import tensorflow as tf
from tensorflow.python.client.timeline import Timeline
from tensorflow.python.ops import gradient_checker
from nengo_dl import utils, DATA_DIR
from nengo_dl.tensor_graph import TensorGraph
from nengo_dl.utils import print_and_flush
logger = logging.getLogger(__name__)
[docs]class Simulator(object):
"""Simulate network using the ``nengo_dl`` backend.
Parameters
----------
network : :class:`~nengo:nengo.Network` or None
a network object to be built and then simulated. If None,
then a built model must be passed to ``model`` instead
dt : float, optional
length of a simulator timestep, in seconds
seed : int, optional
seed for all stochastic operators used in this simulator
model : :class:`~nengo:nengo.builder.Model`, optional
pre-built model object
tensorboard : bool, optional
if True, save network output in the Tensorflow summary format,
which can be loaded into Tensorboard
dtype : ``tf.DType``, optional
floating point precision to use for simulation
step_blocks : int, optional
controls how many simulation steps run each time the graph is
executed (affects memory usage and graph construction time)
device : None or ``"/cpu:0"`` or ``"/gpu:[0-n]"``, optional
device on which to execute computations (if None then uses the
default device as determined by Tensorflow)
unroll_simulation : bool, optional
if True, unroll simulation loop by explicitly building each iteration
(up to ``step_blocks``) into the computation graph. if False, use a
symbolic loop, which is more general and produces a simpler graph, but
is likely to be slower to simulate
minibatch_size : int, optional
the number of simultaneous inputs that will be passed through the
network
"""
# unsupported unit tests
unsupported = [
("nengo/tests/test_simulator.py:test_warn_on_opensim_del",
"nengo_dl raises a different (more visible) warning (see "
"tests/test_nengo_tests.py:test_warn_on_opensim_del"),
("nengo/tests/test_simulator.py:test_signal_init_values",
"different method required to manually step simulator (see "
"tests/test_nengo_tests.py:test_signal_init_values"),
("nengo/tests/test_simulator.py:test_entry_point",
"overridden so we can pass custom test simulators (see "
"tests/test_nengo_tests.py:test_entry_point"),
("nengo/tests/test_node.py:test_args",
"time is passed as np.float32, not a float (see "
"tests/test_nengo_tests.py:test_args"),
("nengo/tests/test_node.py:test_unconnected_node",
"need to set `step_blocks` to ensure node runs the correct number "
"of times (see tests/test_nengo_tests.py:test_unconnected_node"),
]
def __init__(self, network, dt=0.001, seed=None, model=None,
tensorboard=False, dtype=tf.float32, step_blocks=None,
device=None, unroll_simulation=False, minibatch_size=None):
self.closed = None
self.sess = None
self.tensorboard = tensorboard
self.step_blocks = step_blocks
self.minibatch_size = 1 if minibatch_size is None else minibatch_size
if unroll_simulation and step_blocks is None:
raise SimulationError("`step_blocks` must be specified when the "
"simulation is being unrolled")
# TODO: allow the simulator to be called flexibly with/without
# minibatching
# build model (uses default nengo builder)
if model is None:
self.model = Model(dt=float(dt), label="%s, dt=%f" % (network, dt))
else:
if dt != model.dt:
warnings.warn("Model dt (%g) does not match Simulator "
"dt (%g)" % (model.dt, dt), NengoWarning)
self.model = model
if network is not None:
print_and_flush("Building network", end="")
start = time.time()
self.model.build(network, progress_bar=False)
print("\rBuilding completed in %s " %
datetime.timedelta(seconds=int(time.time() - start)))
# set up tensorflow graph plan
self.tensor_graph = TensorGraph(
self.model, self.dt, step_blocks, unroll_simulation, dtype,
self.minibatch_size, device)
self.data = ProbeDict(
self.model.params,
{p: (minibatch_size if self.model.sig[p]["in"].minibatched
else -1) for p in self.model.probes})
if seed is None:
seed = np.random.randint(np.iinfo(np.int32).max)
self.reset(seed=seed)
[docs] def reset(self, seed=None):
"""Resets the simulator to initial conditions.
Parameters
----------
seed : int, optional
if not None, overwrite the default simulator seed with this value
(note: this becomes the new default simulator seed)
"""
if self.closed:
raise SimulatorClosed("Cannot reset closed Simulator.")
# close old session
if self.sess is not None:
self.close()
if seed is not None:
self.seed = seed
self.rng = np.random.RandomState(self.seed)
# TODO: why is setting the tensorflow seed necessary to make
# gradient descent training deterministic?
tf.set_random_seed(self.seed)
# (re)build graph
print_and_flush("Constructing graph", end="")
start = time.time()
self.tensor_graph.build(self.rng)
print("\rConstruction completed in %s " %
datetime.timedelta(seconds=int(time.time() - start)))
# output graph description to tensorboard summary
if self.tensorboard:
directory = "%s/%s" % (DATA_DIR, self.model.toplevel.label)
if os.path.isdir(directory):
run_number = max(
[int(x[4:]) for x in os.listdir(directory)
if x.startswith("run")]) + 1
else:
run_number = 0
self.summary = tf.summary.FileWriter(
"%s/run_%d" % (directory, run_number),
graph=self.tensor_graph.graph)
# start session
# note: we need to allow soft placement when using tf.while_loop,
# because tensorflow pins loop variables to the CPU
config = tf.ConfigProto(
allow_soft_placement=not self.tensor_graph.unroll_simulation,
log_device_placement=False,
)
self.sess = tf.InteractiveSession(graph=self.tensor_graph.graph,
config=config)
self.closed = False
# initialize variables
self.soft_reset(include_trainable=True, include_probes=True)
self.n_steps = 0
self.time = 0.0
self.final_bases = [
x[0] for x in self.tensor_graph.base_arrays_init.values()]
[docs] def soft_reset(self, include_trainable=False, include_probes=False):
"""Resets the internal state of the simulation, but doesn't
rebuild the graph.
Parameters
----------
include_trainable : bool, optional
if True, also reset any training that has been performed on
network parameters (e.g., connection weights)
include_probes : bool, optional
if True, also clear probe data
"""
init_ops = [self.tensor_graph.local_init_op,
self.tensor_graph.global_init_op]
if include_trainable:
init_ops.append(self.tensor_graph.trainable_init_op)
self.sess.run(init_ops)
if include_probes:
for p in self.model.probes:
self.model.params[p] = []
[docs] def step(self, **kwargs):
"""Run the simulation for one time step.
Parameters
----------
kwargs : dict
see :meth:`._run_steps`
"""
self.run_steps(1, **kwargs)
[docs] def run(self, time_in_seconds, **kwargs):
"""Simulate for the given length of time.
Parameters
----------
time_in_seconds : float
amount of time to run the simulation for
kwargs : dict
see :meth:`._run_steps`
"""
steps = int(np.round(float(time_in_seconds) / self.dt))
self.run_steps(steps, **kwargs)
[docs] def run_steps(self, n_steps, **kwargs):
"""Simulate for the given number of steps.
Parameters
----------
n_steps : int
the number of simulation steps to be executed
kwargs : dict
see :meth:`._run_steps`
Notes
-----
If ``step_blocks`` is specified, and ``n_steps > step_blocks``, this
will repeatedly execute ``step_blocks`` timesteps until the the number
of steps executed is >= ``n_steps``.
"""
if self.closed:
raise SimulatorClosed("Simulator cannot run because it is closed.")
if self.step_blocks is not None and n_steps % self.step_blocks != 0:
warnings.warn(
"Number of steps (%d) is not an even multiple of `step_blocks`"
" (%d). Simulation will run for %d steps, which may have "
"unintended side effects." %
(n_steps, self.step_blocks,
self.step_blocks * (n_steps // self.step_blocks + 1)))
print_and_flush("Simulation started", end="")
start = time.time()
if self.step_blocks is None:
probe_data = self._run_steps(n_steps, **kwargs)
self._update_probe_data(probe_data, self.n_steps - n_steps,
n_steps)
else:
# break the run up into `step_blocks` sized chunks
remaining_steps = n_steps
while remaining_steps > 0:
probe_data = self._run_steps(self.step_blocks, **kwargs)
remaining_steps -= self.step_blocks
self._update_probe_data(
probe_data, self.n_steps - self.step_blocks,
self.step_blocks + min(remaining_steps, 0))
# update n_steps/time
self.n_steps += remaining_steps
self.time = self.n_steps * self.dt
print("\rSimulation completed in %s" %
datetime.timedelta(seconds=int(time.time() - start)))
[docs] def _run_steps(self, n_steps, profile=False, input_feeds=None):
"""Execute ``step_blocks`` sized segments of the simulation.
Parameters
----------
n_steps : int
the number of simulation steps to be executed
profile : bool, optional
if True, collect TensorFlow profiling information while the
simulation is running (this will slow down the simulation)
input_feeds : dict of {:class:`~nengo:nengo.Node`: \
:class:`~numpy:numpy.ndarray`}
override the values of input Nodes with the given data. arrays
should have shape ``(sim.minibatch_size, n_steps, node.size_out)``.
Notes
-----
- This function should not be called directly; run the simulator
through :meth:`.Simulator.step`, :meth:`.Simulator.run_steps`, or
:meth:`.Simulator.run`.
- The ``input_feeds`` argument allows the user to pass several
simultaneous input sequences through the model. That is, instead of
running the model ``n`` times with 1 input at a time, the model
can be run once with ``n`` inputs at a time. Only the values of
input nodes (nodes with no incoming Connections) can be overwritten
in this way.
"""
if profile:
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
else:
run_options = None
run_metadata = None
# execute the simulation loop
try:
steps_run, probe_data = self.sess.run(
[self.tensor_graph.steps_run, self.tensor_graph.probe_arrays],
feed_dict=self._fill_feed(n_steps, input_feeds,
start=self.n_steps),
options=run_options, run_metadata=run_metadata)
except (tf.errors.InternalError, tf.errors.UnknownError) as e:
if e.op.type == "PyFunc":
raise SimulationError(
"Function '%s' caused an error "
"(see error log above)" % e.op.name)
else:
raise e
# update n_steps
assert steps_run == n_steps
self.n_steps += steps_run
self.time = self.n_steps * self.dt
if profile:
timeline = Timeline(run_metadata.step_stats)
with open("nengo_dl_profile.json", "w") as f:
f.write(timeline.generate_chrome_trace_format())
return probe_data
[docs] def train(self, inputs, targets, optimizer, n_epochs=1, objective="mse",
shuffle=True):
"""Optimize the trainable parameters of the network using the given
optimization method, minimizing the objective value over the given
inputs and targets.
Parameters
----------
inputs : dict of {:class:`~nengo:nengo.Node`: \
:class:`~numpy:numpy.ndarray`}
input values for Nodes in the network; arrays should have shape
``(batch_size, sim.step_blocks, node.size_out)``
targets : dict of {:class:`~nengo:nengo.Probe`: \
:class:`~numpy:numpy.ndarray`}
desired output value at Probes, corresponding to each value in
``inputs``; arrays should have shape
``(batch_size, sim.step_blocks, probe.size_in)``
optimizer : ``tf.train.Optimizer``
Tensorflow optimizer, e.g.
``tf.train.GradientDescentOptimizer(learning_rate=0.1)``
n_epochs : int, optional
run training for the given number of epochs (complete passes
through ``inputs``)
objective : ``"mse"`` or callable, optional
the objective to be minimized. passing ``"mse"`` will train with
mean squared error. a custom function
``f(output, target) -> loss`` can be passed that consumes the
actual output and target output for a probe in ``targets``
and returns a ``tf.Tensor`` representing the scalar loss value for
that Probe (loss will be averaged across Probes).
shuffle : bool, optional
if True, randomize the data into different minibatches each epoch
Notes
-----
- Deep learning methods require the network to be differentiable, which
means that trying to train a network with non-differentiable elements
will result in an error. Examples of common non-differentiable
elements include :class:`~nengo:nengo.LIF`,
:class:`~nengo:nengo.Direct`, or processes/neurons that don't have a
custom TensorFlow implementation (see
:class:`.processes.SimProcessBuilder`/
:class:`.neurons.SimNeuronsBuilder`)
- Most TensorFlow optimizers do not have GPU support for networks with
sparse reads, which are a common element in Nengo models. If your
network contains sparse reads then training will have to be
executed on the CPU (by creating the simulator via
``nengo_dl.Simulator(..., device="/cpu:0")``), or is limited to
optimizers with GPU support (currently this is only
``tf.train.GradientDescentOptimizer``). Follow `this issue
<https://github.com/tensorflow/tensorflow/issues/2314>`_ for updates
on Tensorflow GPU support.
"""
if self.closed:
raise SimulatorClosed("Simulator cannot be trained because it is "
"closed.")
if self.step_blocks is None:
raise SimulationError(
"Simulator `step_blocks` must be set for training "
"(`Simulator(..., step_blocks=n)`)")
for n, x in inputs.items():
if x.shape[1] != self.step_blocks:
raise SimulationError(
"Length of input sequence (%s) does not match "
"`step_blocks` (%s)" % (x.shape[1], self.step_blocks))
if x.shape[2] != n.size_out:
raise SimulationError(
"Dimensionality of input sequence (%d) does not match "
"node.size_out (%d)" % (x.shape[2], n.size_out))
for p, x in targets.items():
if x.shape[1] != self.step_blocks:
raise SimulationError(
"Length of target sequence (%s) does not match "
"`step_blocks` (%s)" % (x.shape[1], self.step_blocks))
if x.shape[2] != p.size_in:
raise SimulationError(
"Dimensionality of target sequence (%d) does not match "
"probe.size_in (%d)" % (x.shape[2], p.size_in))
# check for non-differentiable elements in graph
# utils.find_non_differentiable(
# [self.tensor_graph.invariant_ph[n] for n in inputs],
# [self.tensor_graph.probe_arrays[self.model.probes.index(p)]
# for p in targets])
# build optimizer op
opt_op, opt_slots_init = self.tensor_graph.build_optimizer(
optimizer, tuple(targets.keys()), objective)
# initialize any variables that were created by the optimizer
self.sess.run(opt_slots_init)
progress = utils.ProgressBar(n_epochs, "Training")
for n in range(n_epochs):
for inp, tar in utils.minibatch_generator(
inputs, targets, self.minibatch_size, rng=self.rng,
shuffle=shuffle):
# TODO: set up queue to feed in data more efficiently
self.soft_reset()
try:
self.sess.run([opt_op], feed_dict=self._fill_feed(
self.step_blocks, inp, tar))
except tf.errors.InvalidArgumentError:
raise SimulationError(
"TensorFlow does not yet support this optimizer on "
"the GPU; try `Simulator(..., device='/cpu:0')`")
progress.step()
self.soft_reset()
[docs] def loss(self, inputs, targets, objective):
"""Compute the loss value for the given objective and inputs/targets.
Parameters
----------
inputs : dict of {:class:`~nengo:nengo.Node`: \
:class:`~numpy:numpy.ndarray`}
input values for Nodes in the network; arrays should have shape
``(batch_size, sim.step_blocks, node.size_out)``
targets : dict of {:class:`~nengo:nengo.Probe`: \
:class:`~numpy:numpy.ndarray`}
desired output value at Probes, corresponding to each value in
``inputs``; arrays should have shape
``(batch_size, sim.step_blocks, probe.size_in)``
objective : ``"mse"`` or callable
the objective used to compute loss. passing ``"mse"`` will use
mean squared error. a custom function
``f(output, target) -> loss`` can be passed that consumes the
actual output and target output for a probe in ``targets``
and returns a ``tf.Tensor`` representing the scalar loss value for
that Probe (loss will be averaged across Probes)
Notes
-----
Calling this function will reset all values in the network, so it
should not be intermixed with calls to :meth:`.Simulator.run`.
"""
if self.closed:
raise SimulatorClosed("Loss cannot be computed after simulator is "
"closed.")
for n, x in inputs.items():
if x.shape[1] != self.step_blocks:
raise SimulationError(
"Length of input sequence (%s) does not match "
"`step_blocks` (%s)" % (x.shape[1], self.step_blocks))
if x.shape[2] != n.size_out:
raise SimulationError(
"Dimensionality of input sequence (%d) does not match "
"node.size_out (%d)" % (x.shape[2], n.size_out))
for p, x in targets.items():
if x.shape[1] != self.step_blocks:
raise SimulationError(
"Length of target sequence (%s) does not match "
"`step_blocks` (%s)" % (x.shape[1], self.step_blocks))
if x.shape[2] != p.size_in:
raise SimulationError(
"Dimensionality of target sequence (%d) does not match "
"probe.size_in (%d)" % (x.shape[2], p.size_in))
# get loss op
loss = self.tensor_graph.build_loss(objective, tuple(targets.keys()))
# compute loss on data
loss_val = 0
for i, (inp, tar) in enumerate(utils.minibatch_generator(
inputs, targets, self.minibatch_size, rng=self.rng)):
self.soft_reset()
loss_val += self.sess.run(
loss, feed_dict=self._fill_feed(self.step_blocks, inp, tar))
self.soft_reset()
loss_val /= i + 1
return loss_val
[docs] def _fill_feed(self, n_steps, inputs, targets=None, start=0):
"""Create a feed dictionary containing values for all the placeholder
inputs in the network, which will be passed to ``tf.Session.run``.
Parameters
----------
n_steps : int
the number of execution steps
input_feeds : dict of {:class:`~nengo:nengo.Node`: \
:class:`~numpy:numpy.ndarray`}
override the values of input Nodes with the given data. arrays
should have shape ``(sim.minibatch_size, n_steps, node.size_out)``.
targets : dict of {:class:`~nengo:nengo.Probe`: \
:class:`~numpy:numpy.ndarray`}, optional
values for target placeholders (only necessary if loss is being
computed, e.g. when training the network)
start : int, optional
initial value of simulator timestep
Returns
-------
dict of {``tf.Tensor``: :class:`~numpy:numpy.ndarray`}
feed values for placeholder tensors in the network
"""
# fill in loop variables
feed_dict = {
self.tensor_graph.step_var: start,
self.tensor_graph.stop_var: start + n_steps
}
# fill in values for base variables from previous run
# TODO: remove this if we're sure we're not going back to the tensor
# approach
feed_dict.update(
{k: v for k, v in zip(
self.tensor_graph.base_vars,
self.final_bases) if k.op.type == "Placeholder"})
# fill in input values
tmp = self._generate_inputs(inputs, n_steps)
feed_dict.update(tmp)
# fill in target values
if targets is not None:
feed_dict.update(
{self.tensor_graph.target_phs[p]: np.moveaxis(t, 0, -1)
for p, t in targets.items()})
return feed_dict
def _generate_inputs(self, input_feeds, n_steps):
"""Generate inputs for the network (the output values of each Node with
no incoming connections).
Parameters
----------
input_feeds : dict of {:class:`~nengo:nengo.Node`: \
:class:`~numpy:numpy.ndarray`}
override the values of input Nodes with the given data. arrays
should have shape ``(sim.minibatch_size, n_steps, node.size_out)``.
n_steps : int
number of simulation timesteps for which to generate input data
"""
if input_feeds is None:
input_feeds = {}
else:
# validate inputs
for n, v in input_feeds.items():
target_shape = (self.minibatch_size, n_steps, n.size_out)
if v.shape != target_shape:
raise SimulationError(
"Input feed for node %s has wrong shape; expected %s, "
"saw %s" % (n, target_shape, v.shape))
feed_vals = {}
for n in self.tensor_graph.invariant_inputs:
# if the output signal is not in sig map, that means no operators
# use the output of this node. similarly, if node.size_out is 0,
# the node isn't producing any output values.
using_output = (
self.model.sig[n]["out"] in self.tensor_graph.sig_map and
n.size_out > 0)
if using_output:
if n in input_feeds:
# move minibatch dimension to the end
feed_val = np.moveaxis(input_feeds[n], 0, -1)
elif isinstance(n.output, np.ndarray):
feed_val = np.tile(n.output[None, :, None],
(n_steps, 1, self.minibatch_size))
else:
func = self.tensor_graph.invariant_funcs[n]
feed_val = []
for i in range(self.n_steps + 1,
self.n_steps + n_steps + 1):
# note: need to copy the output of func, as func
# may mutate its outputs in-place on subsequent calls
feed_val += [np.array(func(i * self.dt))]
feed_val = np.stack(feed_val, axis=0)
feed_val = np.tile(feed_val[..., None],
(1, 1, self.minibatch_size))
feed_vals[self.tensor_graph.invariant_ph[n]] = feed_val
elif (not isinstance(n.output, np.ndarray) and
n.output in self.tensor_graph.invariant_funcs.values()):
# note: we still call the function even if the output
# is not being used, because it may have side-effects
func = self.tensor_graph.invariant_funcs[n]
for i in range(self.n_steps + 1, self.n_steps + n_steps + 1):
func(i * self.dt)
return feed_vals
def _update_probe_data(self, probe_data, start, n_steps):
"""Updates the stored probe data (since the last reset) with the data
from the latest run.
Downsamples the probe data returned from tensorflow (from every
simulation timestep) according to probe `sample_every` and the number
of steps run.
Parameters
----------
probe_data : list of `np.ndarray`
probe data from every timestep
start : int
the simulation timestep at which probe data starts
n_steps : int
the number of timesteps over which we want to collect data
"""
# first, remove any extra timesteps (due to `step_blocks` mismatch)
probe_data = [p[:n_steps] for p in probe_data]
for i, p in enumerate(self.model.probes):
if p.sample_every is not None:
# downsample probe according to `sample_every`
period = p.sample_every / self.dt
steps = np.arange(start, start + n_steps)
probe_data[i] = probe_data[i][(steps + 1) % period < 1]
# update stored probe data
self.model.params[p] += [probe_data[i]]
[docs] def save_params(self, path):
"""Save trainable network parameters to the given ``path``.
Parameters
----------
path : str
filepath of parameter output file
"""
if self.closed:
raise SimulationError("Simulation has been closed, cannot save "
"parameters")
path = tf.train.Saver().save(self.sess, path)
logger.info("Model parameters saved to %s", path)
[docs] def load_params(self, path):
"""Load trainable network parameters from the given ``path``.
Parameters
----------
path : str
filepath of parameter input file
"""
if self.closed:
raise SimulationError("Simulation has been closed, cannot load "
"parameters")
tf.train.Saver().restore(self.sess, path)
[docs] def print_params(self, msg=None):
"""Print current values of trainable network parameters.
Parameters
----------
msg : str, optional
title for print output, useful to differentiate multiple print
calls
"""
if self.closed:
raise SimulationError("Simulation has been closed, cannot print "
"parameters")
param_sigs = {k: v for k, v in self.tensor_graph.sig_map.items()
if k.trainable}
params = {v.key: self.tensor_graph.signals.bases[v.key]
for v in param_sigs.values()}
param_vals = self.sess.run(params)
print("%s:" % "Parameters" if msg is None else msg)
for sig, tens in param_sigs.items():
print("-" * 10)
print(sig)
print(param_vals[tens.key][tens.indices])
[docs] def close(self):
"""Close the simulation, freeing resources.
Notes
-----
The simulation cannot be restarted after it is closed. This is not a
technical limitation, just a design decision made for all Nengo
simulators.
"""
if not self.closed:
self.sess.close()
self.closed = True
self.sess = None
# note: we use getattr in case it crashes before the summary
# object is created
if getattr(self, "summary", None) is not None:
self.summary.close()
def __enter__(self):
return self
def __exit__(self, exc_type, exc_value, traceback):
self.close()
@property
def dt(self):
"""(float) The time step of the simulator."""
return self.model.dt
@dt.setter
def dt(self, dummy):
raise ReadonlyError(attr='dt', obj=self)
def __del__(self):
"""Raise a RuntimeWarning if the Simulator is deallocated while open.
"""
if self.closed is not None and not self.closed:
warnings.warn(
"Simulator with model=%s was deallocated while open. "
"Simulators should be closed manually to ensure resources "
"are properly freed." % self.model, RuntimeWarning)
self.close()
[docs] def trange(self, dt=None):
"""Create a vector of times matching probed data.
Note that the range does not start at 0 as one might expect, but at
the first timestep (i.e., ``dt``).
Parameters
----------
dt : float, optional
the sampling period of the probe to create a range for;
if None, the simulator's ``dt`` will be used.
"""
dt = self.dt if dt is None else dt
n_steps = int(self.n_steps * (self.dt / dt))
return dt * np.arange(1, n_steps + 1)
[docs] def check_gradients(self, outputs=None, atol=1e-5, rtol=1e-3):
"""Perform gradient checks for the network (used to verify that the
analytic gradients are correct).
Raises a simulation error if the difference between analytic and
numeric gradient is greater than ``atol + rtol * numeric_grad``
(elementwise).
Parameters
----------
outputs : ``tf.Tensor`` or list of ``tf.Tensor``
compute gradients wrt this output (if None, computes wrt each
output probe)
atol : float, optional
absolute error tolerance
rtol : float, optional
relative (to numeric grad) error tolerance
Notes
-----
Calling this function will reset all values in the network, so it
should not be intermixed with calls to :meth:`.Simulator.run`.
"""
delta = 1e-3
feed = self._fill_feed(
self.step_blocks, {n: np.zeros((self.minibatch_size,
self.step_blocks, n.size_out))
for n in self.tensor_graph.invariant_inputs},
{p: np.zeros((self.minibatch_size, self.step_blocks, p.size_in))
for p in self.tensor_graph.target_phs})
if outputs is None:
# note: the x + 0 is necessary because `gradient_checker`
# doesn't work properly if the output variable is a tensorarray
outputs = [x + 0 for x in self.tensor_graph.probe_arrays]
elif isinstance(outputs, tf.Tensor):
outputs = [outputs]
# check gradient wrt inp
for node, inp in self.tensor_graph.invariant_ph.items():
inp_shape = inp.get_shape().as_list()
inp_tens = self.tensor_graph.invariant_ph[node]
feed[inp_tens] = np.ascontiguousarray(feed[inp_tens])
inp_val = np.ravel(feed[inp_tens])
for out in outputs:
out_shape = out.get_shape().as_list()
# we need to compute the numeric jacobian manually, to
# correctly handle variables (tensorflow doesn't expect
# state ops in `compute_gradient`, because it doesn't define
# gradients for them)
numeric = np.zeros((np.prod(inp_shape, dtype=np.int32),
np.prod(out_shape, dtype=np.int32)))
for i in range(numeric.shape[0]):
self.soft_reset()
inp_val[i] = delta
plus = self.sess.run(out, feed_dict=feed)
self.soft_reset()
inp_val[i] = -delta
minus = self.sess.run(out, feed_dict=feed)
numeric[i] = np.ravel((plus - minus) / (2 * delta))
inp_val[i] = 0
self.soft_reset()
dx, dy = gradient_checker._compute_dx_and_dy(
inp, out, out_shape)
analytic = gradient_checker._compute_theoretical_jacobian(
inp, inp_shape, np.zeros(inp_shape), dy, out_shape, dx,
extra_feed_dict=feed)
if np.any(np.isnan(analytic)) or np.any(np.isnan(numeric)):
raise SimulationError("NaNs detected in gradient")
fail = abs(analytic - numeric) >= atol + rtol * abs(numeric)
if np.any(fail):
raise SimulationError(
"Gradient check failed for input %s and output %s\n"
"numeric values:\n%s\n"
"analytic values:\n%s\n" % (node, out, numeric[fail],
analytic[fail]))
self.soft_reset()
logger.info("Gradient check passed")
class ProbeDict(Mapping):
"""Map from :class:`~nengo:nengo.Probe` -> :class:`~numpy:numpy.ndarray`,
used to access output of the model after simulation.
This is more like a view on the dict that the simulator manipulates.
However, for speed reasons, the simulator uses Python lists,
and we want to return NumPy arrays. Additionally, this mapping
is readonly, which is more appropriate for its purpose.
Parameters
----------
raw : dict of {:class:`~nengo:nengo.Probe`: \
list of :class:`~numpy:numpy.ndarray`}
the raw probe output from the simulator (a list of arrays containing
the output from each ``step_blocks`` execution segment)
minibatches : dict of {:class:`~nengo:nengo.Probe`: int or None}
the minibatch size for each probe in the dictionary (or -1 if the
probed signal does not have a minibatch dimension)
Notes
-----
ProbeDict should never be created/accessed directly by the user, but rather
via ``sim.data`` (which is an instance of ProbeDict).
"""
def __init__(self, raw, minibatches):
self.raw = raw
self.minibatches = minibatches
self._cache = {}
def __getitem__(self, key):
cache_miss = (key not in self._cache or
len(self._cache[key]) != len(self.raw[key]))
if cache_miss:
rval = self.raw[key]
if isinstance(rval, list):
# combine data from _run_steps iterations
rval = np.concatenate(rval, axis=0)
if self.minibatches[key] != -1:
if self.minibatches[key] is None:
# get rid of batch dimension
rval = rval[..., 0]
else:
# move batch dimension to front
rval = np.moveaxis(rval, -1, 0)
rval.setflags(write=False)
self._cache[key] = rval
return self._cache[key]
def __iter__(self):
return iter(self.raw)
def __len__(self):
return len(self.raw)
def __repr__(self):
return repr(self.raw)
def __str__(self):
return str(self.raw)