BindsNET学习系列——Network
相关源码：bindsnet/bindsnet/network/network.py
class Network(torch.nn.Module):
# language=rst
"""
Central object of the ``bindsnet`` package. Responsible for the simulation and
interaction of nodes and connections.
**Example:**
.. code-block:: python
import torch
import matplotlib.pyplot as plt
from bindsnet import encoding
from bindsnet.network import Network, nodes, topology, monitors
network = Network(dt=1.0) # Instantiates network.
X = nodes.Input(100) # Input layer.
Y = nodes.LIFNodes(100) # Layer of LIF neurons.
C = topology.Connection(source=X, target=Y, w=torch.rand(X.n, Y.n)) # Connection from X to Y.
# Spike monitor objects.
M1 = monitors.Monitor(obj=X, state_vars=['s'])
M2 = monitors.Monitor(obj=Y, state_vars=['s'])
# Add everything to the network object.
network.add_layer(layer=X, name='X')
network.add_layer(layer=Y, name='Y')
network.add_connection(connection=C, source='X', target='Y')
network.add_monitor(monitor=M1, name='X')
network.add_monitor(monitor=M2, name='Y')
# Create Poisson-distributed spike train inputs.
data = 15 * torch.rand(100) # Generate random Poisson rates for 100 input neurons.
train = encoding.poisson(datum=data, time=5000) # Encode input as 5000ms Poisson spike trains.
# Simulate network on generated spike trains.
inputs = {'X' : train} # Create inputs mapping.
network.run(inputs=inputs, time=5000) # Run network simulation.
# Plot spikes of input and output layers.
spikes = {'X' : M1.get('s'), 'Y' : M2.get('s')}
fig, axes = plt.subplots(2, 1, figsize=(12, 7))
for i, layer in enumerate(spikes):
axes[i].matshow(spikes[layer], cmap='binary')
axes[i].set_title('%s spikes' % layer)
axes[i].set_xlabel('Time'); axes[i].set_ylabel('Index of neuron')
axes[i].set_xticks(()); axes[i].set_yticks(())
axes[i].set_aspect('auto')
plt.tight_layout(); plt.show()
"""
def __init__(
self,
dt: float = 1.0,
batch_size: int = 1,
learning: bool = True,
reward_fn: Optional[Type[AbstractReward]] = None,
) -> None:
# language=rst
"""
Initializes network object.
:param dt: Simulation timestep.
:param batch_size: Mini-batch size.
:param learning: Whether to allow connection updates. True by default.
:param reward_fn: Optional class allowing for modification of reward in case of
reward-modulated learning.
"""
super().__init__()
self.dt = dt
self.batch_size = batch_size
self.layers = {} # 关于节点层的字典
self.connections = {} # 关于连接的字典
self.monitors = {}
self.train(learning)
if reward_fn is not None:
self.reward_fn = reward_fn()
else:
self.reward_fn = None
def add_layer(self, layer: Nodes, name: str) -> None:
# language=rst
"""
Adds a layer of nodes to the network.
:param layer: A subclass of the ``Nodes`` object.
:param name: Logical name of layer.
"""
self.layers[name] = layer
self.add_module(name, layer)
layer.train(self.learning)
layer.compute_decays(self.dt)
layer.set_batch_size(self.batch_size)
def add_connection(
self, connection: AbstractConnection, source: str, target: str
) -> None:
# language=rst
"""
Adds a connection between layers of nodes to the network.
:param connection: An instance of class ``Connection``.
:param source: Logical name of the connection's source layer.
:param target: Logical name of the connection's target layer.
"""
self.connections[(source, target)] = connection
self.add_module(source + "_to_" + target, connection)
connection.dt = self.dt
connection.train(self.learning)
def add_monitor(self, monitor: AbstractMonitor, name: str) -> None:
# language=rst
"""
Adds a monitor on a network object to the network.
:param monitor: An instance of class ``Monitor``.
:param name: Logical name of monitor object.
"""
self.monitors[name] = monitor
monitor.network = self
monitor.dt = self.dt
def save(self, file_name: str) -> None:
# language=rst
"""
Serializes the network object to disk.
:param file_name: Path to store serialized network object on disk.
**Example:**
.. code-block:: python
import torch
import matplotlib.pyplot as plt
from pathlib import Path
from bindsnet.network import *
from bindsnet.network import topology
# Build simple network.
network = Network(dt=1.0)
X = nodes.Input(100) # Input layer.
Y = nodes.LIFNodes(100) # Layer of LIF neurons.
C = topology.Connection(source=X, target=Y, w=torch.rand(X.n, Y.n)) # Connection from X to Y.
# Add everything to the network object.
network.add_layer(layer=X, name='X')
network.add_layer(layer=Y, name='Y')
network.add_connection(connection=C, source='X', target='Y')
# Save the network to disk.
network.save(str(Path.home()) + '/network.pt')
"""
torch.save(self, open(file_name, "wb"))
def clone(self) -> "Network":
# language=rst
"""
Returns a cloned network object.
:return: A copy of this network.
"""
virtual_file = tempfile.SpooledTemporaryFile()
torch.save(self, virtual_file)
virtual_file.seek(0)
return torch.load(virtual_file)
def _get_inputs(self, layers: Iterable = None) -> Dict[str, torch.Tensor]:
# language=rst
"""
Fetches outputs from network layers to use as input to downstream layers.
:param layers: Layers to update inputs for. Defaults to all network layers.
:return: Inputs to all layers for the current iteration.
"""
inputs = {}
if layers is None:
layers = self.layers
# Loop over network connections.
for c in self.connections:
if c[1] in layers:
# Fetch source and target populations.
source = self.connections[c].source
target = self.connections[c].target
if not c[1] in inputs:
if isinstance(target, CSRMNodes):
inputs[c[1]] = torch.zeros(
self.batch_size,
target.res_window_size,
*target.shape,
device=target.s.device
)
else:
inputs[c[1]] = torch.zeros(
self.batch_size, *target.shape, device=target.s.device
)
# Add to input: source's spikes multiplied by connection weights.
if isinstance(target, CSRMNodes):
inputs[c[1]] += self.connections[c].compute_window(source.s)
else:
inputs[c[1]] += self.connections[c].compute(source.s)
return inputs
def run(
self, inputs: Dict[str, torch.Tensor], time: int, one_step=False, **kwargs
) -> None:
# language=rst
"""
Simulate network for given inputs and time.
:param inputs: Dictionary of ``Tensor``s of shape ``[time, *input_shape]`` or
``[time, batch_size, *input_shape]``.
:param time: Simulation time.
:param one_step: Whether to run the network in "feed-forward" mode, where inputs
propagate all the way through the network in a single simulation time step.
Layers are updated in the order they are added to the network.
Keyword arguments:
:param Dict[str, torch.Tensor] clamp: Mapping of layer names to boolean masks if
neurons should be clamped to spiking. The ``Tensor``s have shape
``[n_neurons]`` or ``[time, n_neurons]``.
:param Dict[str, torch.Tensor] unclamp: Mapping of layer names to boolean masks
if neurons should be clamped to not spiking. The ``Tensor``s should have
shape ``[n_neurons]`` or ``[time, n_neurons]``.
:param Dict[str, torch.Tensor] injects_v: Mapping of layer names to boolean
masks if neurons should be added voltage. The ``Tensor``s should have shape
``[n_neurons]`` or ``[time, n_neurons]``.
:param Union[float, torch.Tensor] reward: Scalar value used in reward-modulated
learning.
:param Dict[Tuple[str], torch.Tensor] masks: Mapping of connection names to
boolean masks determining which weights to clamp to zero.
**Example:**
.. code-block:: python
import torch
import matplotlib.pyplot as plt
from bindsnet.network import Network
from bindsnet.network.nodes import Input
from bindsnet.network.monitors import Monitor
# Build simple network.
network = Network()
network.add_layer(Input(500), name='I')
network.add_monitor(Monitor(network.layers['I'], state_vars=['s']), 'I')
# Generate spikes by running Bernoulli trials on Uniform(0, 0.5) samples.
spikes = torch.bernoulli(0.5 * torch.rand(500, 500))
# Run network simulation.
network.run(inputs={'I' : spikes}, time=500)
# Look at input spiking activity.
spikes = network.monitors['I'].get('s')
plt.matshow(spikes, cmap='binary')
plt.xticks(()); plt.yticks(());
plt.xlabel('Time'); plt.ylabel('Neuron index')
plt.title('Input spiking')
plt.show()
"""
# Parse keyword arguments.
clamps = kwargs.get("clamp", {})
unclamps = kwargs.get("unclamp", {})
masks = kwargs.get("masks", {})
injects_v = kwargs.get("injects_v", {})
# Compute reward.
if self.reward_fn is not None:
kwargs["reward"] = self.reward_fn.compute(**kwargs)
# Dynamic setting of batch size.
if inputs != {}:
for key in inputs:
# goal shape is [time, batch, n_0, ...]
if len(inputs[key].size()) == 1:
# current shape is [n_0, ...]
# unsqueeze twice to make [1, 1, n_0, ...]
inputs[key] = inputs[key].unsqueeze(0).unsqueeze(0)
elif len(inputs[key].size()) == 2:
# current shape is [time, n_0, ...]
# unsqueeze dim 1 so that we have
# [time, 1, n_0, ...]
inputs[key] = inputs[key].unsqueeze(1)
for key in inputs:
# batch dimension is 1, grab this and use for batch size
if inputs[key].size(1) != self.batch_size:
self.batch_size = inputs[key].size(1)
for l in self.layers:
self.layers[l].set_batch_size(self.batch_size)
for m in self.monitors:
self.monitors[m].reset_state_variables()
break
# Effective number of timesteps.
timesteps = int(time / self.dt) # 100
# Simulate network activity for `time` timesteps.
for t in range(timesteps):
# Get input to all layers (synchronous mode).
current_inputs = {}
if not one_step:
current_inputs.update(self._get_inputs()) # 获取各层当前值
### breakout.py ###
# current_inputs: 'Hidden Layer'([1, 100]), 'Output Layer'([1, 4])
# self.layers: {'Input Layer': Input(), 'Hidden Layer': IzhikevichNodes(), 'Output Layer': IzhikevichNodes()}
# inputs['Input Layer']([100, 1, 1, 1, 80, 80])
for l in self.layers: # layer_name
# Update each layer of nodes.
if l in inputs:
# 更新输入层当前值
if l in current_inputs:
current_inputs[l] += inputs[l][t]
else:
current_inputs[l] = inputs[l][t]
### breakout.py ###
# current_inputs: 'Hidden Layer'([1, 100]), 'Output Layer'([1, 4]), 'Input Layer'([1, 1, 1, 80, 80])
if one_step: # pass
# Get input to this layer (one-step mode).
current_inputs.update(self._get_inputs(layers=[l]))
if l in current_inputs:
self.layers[l].forward(x=current_inputs[l])
else:
self.layers[l].forward(x=torch.zeros(self.layers[l].s.shape))
# Clamp neurons to spike.
clamp = clamps.get(l, None)
if clamp is not None:
if clamp.ndimension() == 1:
self.layers[l].s[:, clamp] = 1
else:
self.layers[l].s[:, clamp[t]] = 1
# Clamp neurons not to spike.
unclamp = unclamps.get(l, None)
if unclamp is not None:
if unclamp.ndimension() == 1:
self.layers[l].s[:, unclamp] = 0
else:
self.layers[l].s[:, unclamp[t]] = 0
# Inject voltage to neurons.
inject_v = injects_v.get(l, None)
if inject_v is not None:
if inject_v.ndimension() == 1:
self.layers[l].v += inject_v
else:
self.layers[l].v += inject_v[t]
# Run synapse updates.
for c in self.connections:
self.connections[c].update(
mask=masks.get(c, None), learning=self.learning, **kwargs
)
# # Get input to all layers.
# current_inputs.update(self._get_inputs())
# Record state variables of interest.
for m in self.monitors:
self.monitors[m].record()
# Re-normalize connections.
for c in self.connections:
self.connections[c].normalize()
def reset_state_variables(self) -> None:
# language=rst
"""
Reset state variables of objects in network.
"""
for layer in self.layers:
self.layers[layer].reset_state_variables()
for connection in self.connections:
self.connections[connection].reset_state_variables()
for monitor in self.monitors:
self.monitors[monitor].reset_state_variables()
def train(self, mode: bool = True) -> "torch.nn.Module":
# language=rst
"""
Sets the node in training mode.
:param mode: Turn training on or off.
:return: ``self`` as specified in ``torch.nn.Module``.
"""
self.learning = mode
return super().train(mode)
