def test_stimulate_forward(self):
"""Check neuronal firing algorithm forward stimulation of graph."""
graph = self.graph
data = self.data
f = Firing(graph=graph)
output = f.stimulate(neurons=["x", "y"], signals=[data, 1])
self.assertNotEqual(output, {})
def infer(graph, inputs):
"""Use neural network graph to infer some outcomes from inputs."""
neurons = list(inputs.keys())
# setting up our graph in both normal and reversed directions for
# forward and backward pass
forward = Firing(graph=graph)
train = list(inputs.values())
# external counter as I want to rework this in future to work
# with generators + its more efficient to use itertools than to
# iterate over the training set manually using a counter + lookup
# on each iteration which would start from head of list
i = 0
activations = {}
with tqdm(total=len(inputs[neurons[0]]), desc="Infer") as pbar:
for signals in itertools.zip_longest(*train):
# forward pass over all avaliable nodes on graph
out = forward.stimulate(
neurons=neurons,
signals=list(signals),
receptor="forward")
pbar.update(1)
i += 1
# remapping outputs into a single dictionary again just like inputs
for key, value in out.items():
if activations.get(key) is None:
activations[key] = []
activations[key].append(value)
return activations
def test_propogate_signal(self):
"""Check that signal propogation occurs properly for all edges.
This is standard iterable return NOT YIELD, applied to each edge.
"""
some_signal = np.array([1, 2, 3])
graph = nx.MultiDiGraph()
graph.add_node("x", node=IO()) # from this node
graph.add_node("a", node=IO())
graph.add_node("y", node=IO())
graph.add_edge("x", "y") # no signals yet
graph.add_edge("x", "y")
graph.add_edge("x", "y")
graph.add_edge("a", "y") # should never recieve a signal
f = Firing()
f._propogate_signal(graph=graph,
node_name="x",
signal_name="fwd",
signal=some_signal)
# check signal has been applied to each individually
for edge in graph.edges("x", data=True):
np.testing.assert_array_almost_equal(edge[2]["fwd"],
some_signal,
decimal=1,
verbose=True)
# check not applied to seperate node
for edge in graph.edges("a", data=True):
self.assertEqual(edge[2].get("fwd"), None)
def test_get_signal_one(self):
"""Check get single signal is working as expected.
This includes single edges which should not be in a meta container.
"""
some_signal = np.array([1, 2, 3])
graph = nx.MultiDiGraph()
graph.add_node("x", node=IO())
graph.add_node("y", node=IO()) # from this node
graph.add_edge("x", "y", fwd=some_signal)
f = Firing()
signal = f._get_signal(graph=graph, node_name="y", signal_name="fwd")
# should return just the input without meta container
np.testing.assert_array_almost_equal(signal,
some_signal,
decimal=1,
verbose=True)
def test_use_signal(self):
"""Check that func is modifying the state of the node properly."""
some_signal = np.array([1, 0.5, 0])
some_signal = np.broadcast_to(some_signal, shape=(4, 3))
graph = nx.MultiDiGraph()
graph.add_node("y", node=RELU())
f = Firing()
activation = f._use_signal(graph=graph,
node_name="y",
receptor_name="forward",
signal=some_signal)
activation_truth = RELU().forward(some_signal)
np.testing.assert_array_almost_equal(activation,
activation_truth,
decimal=1,
verbose=True)
# quick check to ensure RELU actually did see the input
self.assertEqual(len(graph.nodes(data=True)["y"]["node"].inputs), 1)
def test_harvest(self):
"""Check that probes are harvesting signals."""
some_signal = np.array([1, 2, 3])
graph = nx.MultiDiGraph()
graph.add_node("x", node=IO())
graph.add_node("a", node=IO())
graph.add_node("y", node=IO()) # from this node
graph.add_edge("x", "y", forward=some_signal)
graph.add_edge("x", "y", forward=some_signal)
graph.add_edge("x", "y", forward=some_signal)
graph.add_edge("a", "y", forward=some_signal)
f = Firing(graph=graph)
crop = f.harvest(["y", "y"])
truth = np.broadcast_to(some_signal, shape=(4, 3))
for (_, signal) in crop:
np.testing.assert_array_almost_equal(signal,
truth,
decimal=1,
verbose=True)
def test_get_signal_none(self):
"""Check get non-existant signal is working as expected.
If the node does not have all of its inputs then it is not ready.
So getsignal should return None as the signal is incomplete.
"""
# create a graph with 3 parallel edges from one node and one other
some_signal = np.array([1, 2, 3])
graph = nx.MultiDiGraph()
graph.add_node("x", node=IO())
graph.add_node("a", node=IO())
graph.add_node("y", node=IO()) # from this node
graph.add_edge("x", "y", fwd=some_signal)
graph.add_edge("x", "y") # some missing signal
graph.add_edge("x", "y", fwd=some_signal)
graph.add_edge("a", "y", fwd=some_signal)
f = Firing()
signal = f._get_signal(graph=graph, node_name="y", signal_name="fwd")
self.assertEqual(signal, None)
def test_propogate_signal_yield(self):
"""Check that propogate signal works properly with yield.
If the signal is of type generator/ yield it should map each output
in order with the graph edges. Thus the edges can be different to one
another.
"""
def yielder():
i = 0 # np.array([0]) <- would return by reference
while True:
yield i
i += 1
some_signal = yielder() # as generator
graph = nx.MultiDiGraph()
graph.add_node("x", node=IO()) # from this node
graph.add_node("a", node=IO())
graph.add_node("y", node=IO())
graph.add_edge("x", "y") # no signals yet
graph.add_edge("x", "y")
graph.add_edge("x", "y")
graph.add_edge("x", "DrWho?") # checking works with non yield too
graph.add_edge("a", "y") # should never recieve a signal
f = Firing()
f._propogate_signal(graph=graph,
node_name="x",
signal_name="fwd",
signal=some_signal)
# check signal has been applied to each individually
truth = yielder()
for (_, _, edge) in graph.edges("x", data=True):
np.testing.assert_array_almost_equal(edge["fwd"],
next(truth),
decimal=1,
verbose=True)
# check not applied to seperate node
for edge in graph.edges("a", data=True):
self.assertEqual(edge[2].get("fwd"), None)
def test_get_signal_many(self):
"""Check get multi signal is working as expected.
This includes considering parallel edges, and standard multi edges.
"""
# create a graph with 3 parallel edges from one node and one other
some_signal = np.array([1, 2, 3])
graph = nx.MultiDiGraph()
graph.add_node("x", node=IO())
graph.add_node("a", node=IO())
graph.add_node("y", node=IO()) # from this node
graph.add_edge("x", "y", fwd=some_signal)
graph.add_edge("x", "y", fwd=some_signal)
graph.add_edge("x", "y", fwd=some_signal)
graph.add_edge("a", "y", fwd=some_signal)
f = Firing()
signal = f._get_signal(graph=graph, node_name="y", signal_name="fwd")
truth = np.broadcast_to(some_signal, shape=(4, 3))
# should return all inputs in meta container of shape (4,3) here
np.testing.assert_array_almost_equal(signal,
truth,
decimal=1,
verbose=True)
def test_propogate_none(self):
"""Check that signal propogation does not occur when signal=None."""
some_signal = None
graph = nx.MultiDiGraph()
graph.add_node("x", node=IO()) # from this node
graph.add_node("a", node=IO())
graph.add_node("y", node=IO())
graph.add_edge("x", "y") # no signals yet
graph.add_edge("x", "y")
graph.add_edge("x", "y")
graph.add_edge("a", "y") # should never recieve a signal
f = Firing()
f._propogate_signal(graph=graph,
node_name="x",
signal_name="fwd",
signal=some_signal)
# check signal has not been applied to each edge
for edge in graph.edges("x", data=True):
with self.assertRaises(KeyError):
edge[2]["fwd"]
# check not applied to seperate node
for edge in graph.edges("a", data=True):
self.assertEqual(edge[2].get("fwd"), None)
def test_stimulate_backward(self):
"""Check neuronal firing algorithm forward stimulation of graph."""
graph = self.graph
data = self.data
forward = Firing(graph=graph)
output = forward.stimulate(neurons=["x", "y"], signals=[data, 1])
self.assertNotEqual(output, {})
backward = Firing(graph=graph.reverse(copy=False))
grads = backward.stimulate(neurons=["Loss-CCE", "y_hat"],
signals=[output["Loss-CCE"], 0],
receptor="backward")
self.assertFalse(grads.get("x") is None, "There is no input grad x!")
def test_single_train(self):
graph = self.graph
data = self.data
forward = Firing(graph=graph)
output = forward.stimulate(neurons=["x", "y"], signals=[data, 1])
self.assertNotEqual(output, {})
backward = Firing(graph=graph.reverse(copy=False))
grads = backward.stimulate(neurons=["Loss-CCE", "y_hat"],
signals=[output["Loss-CCE"], 0],
receptor="backward")
for node_meta in graph.nodes(data=True):
node_name = node_meta[0]
node = node_meta[1]["node"]
node.updates()
def train(graph, inputs, batch_size, debug=False):
"""Train neural network graph through backpropagation."""
neurons = list(inputs.keys())
# setting up our graph in both normal and reversed directions for
# forward and backward pass
forward = Firing(graph=graph)
backward = Firing(graph=graph.reverse(copy=False)) # we want them linked
train = list(inputs.values())
# external counter as I want to rework this in future to work
# with generators + its more efficient to use itertools than to
# iterate over the training set manually using a counter + lookup
# on each iteration which would start from head of list
i = 0
with tqdm(total=len(inputs[neurons[0]]), desc="Learn") as pbar:
for signals in itertools.zip_longest(*train):
# forward pass over all avaliable nodes on graph
out = forward.stimulate(
neurons=neurons,
signals=list(signals),
receptor="forward")
# backward pass using output from forward pass to select
# the nodes they came from to pass them back in but as
# losses (or ignored if not a loss)
backward.stimulate(
neurons=list(out.keys()),
signals=list(out.values()),
receptor="backward")
# if we happen to be at the end of a batch update using avg of
# our calculated gradients in all backward passes
# (internal state of the graph nodes so no need to do it
# ourselves)
if i % batch_size == 0:
for node_meta in graph.nodes(data=True):
node = node_meta[1]["node"]
node.updates()
# iterate counter to keep track of batch sizes
pbar.update(1)
i += 1
return out