有关泊松编码

[201528014227051]

感谢作者公开的框架。
关于泊松编码，框架中给的代码是简单的对比：
out_spike = torch.rand_like(x).le(x)
这个和bindsnet当中对于泊松编码的实现过程不一样，bindsnet代码如下:
代码来自：https://gitee.com/CtrlPlayer/notebook/blob/master/torch_notebook/distribution_poisson.ipynb
def poisson(datum: torch.Tensor, time: int, dt: float = 1.0, **kwargs) -> torch.Tensor:
# language=rst
"""
Generates Poisson-distributed spike trains based on input intensity. Inputs must be
non-negative, and give the firing rate in Hz. Inter-spike intervals (ISIs) for
non-negative data incremented by one to avoid zero intervals while maintaining ISI
distributions.
:param datum: Tensor of shape [n_1, ..., n_k].
:param time: Length of Poisson spike train per input variable.
:param dt: Simulation time step.
:return: Tensor of shape [time, n_1, ..., n_k] of Poisson-distributed spikes.
"""
assert (datum >= 0).all(), "Inputs must be non-negative"

# Get shape and size of data.
shape, size = datum.shape, datum.numel()
datum = datum.flatten()
time = int(time / dt)

# Compute firing rates in seconds as function of data intensity,
# accounting for simulation time step.
rate = torch.zeros(size)
rate[datum != 0] = 1 / datum[datum != 0] * (1000 / dt)

# Create Poisson distribution and sample inter-spike intervals
# (incrementing by 1 to avoid zero intervals).
dist = torch.distributions.Poisson(rate=rate)
intervals = dist.sample(sample_shape=torch.Size([time + 1]))
intervals[:, datum != 0] += (intervals[:, datum != 0] == 0).float()

#-------------------------------------------------------------------
# 与上一部分代码的循环作用等效
# Calculate spike times by cumulatively summing over time dimension.
times = torch.cumsum(intervals, dim=0).long()
times[times >= time + 1] = 0

# Create tensor of spikes.
spikes = torch.zeros(time + 1, size).byte()
spikes[times, torch.arange(size)] = 1
spikes = spikes[1:]
#-------------------------------------------------------------------

return spikes.view(time, *shape)

请问两者之前的区别是什么？期盼回复

[Yanqi-Chen]

本框架的模拟方式是用二项分布逼近Poisson分布（发放率意义上），讨论见 #68 ，其速度更快但更不精确（相比于真正意义上的Poisson采样）。

def poisson(datum: torch.Tensor, time: int, dt: float = 1.0, **kwargs) -> torch.Tensor:
  # language=rst
  """
  Generates Poisson-distributed spike trains based on input intensity. Inputs must be
  non-negative, and give the firing rate in Hz. Inter-spike intervals (ISIs) for
  non-negative data incremented by one to avoid zero intervals while maintaining ISI
  distributions.
  :param datum: Tensor of shape [n_1, ..., n_k].
  :param time: Length of Poisson spike train per input variable.
  :param dt: Simulation time step.
  :return: Tensor of shape [time, n_1, ..., n_k] of Poisson-distributed spikes.
  """
  # Get shape and size of data.
  shape, size = datum.shape, datum.numel()
  datum = datum.flatten()
  time = int(time / dt)
  
  # Compute firing rates in seconds as function of data intensity,
  # accounting for simulation time step.
  rate = torch.zeros(size)
  rate[datum != 0] = 1 / datum[datum != 0] * (1000 / dt)
  
  # Create Poisson distribution and sample inter-spike intervals
  # (incrementing by 1 to avoid zero intervals).
  dist = torch.distributions.Poisson(rate=rate)
  intervals = dist.sample(sample_shape=torch.Size([time + 1]))
  intervals[:, datum != 0] += (intervals[:, datum != 0] == 0).float()
  
  #-------------------------------------------------------------------
  # 与上一部分代码的循环作用等效
  # Calculate spike times by cumulatively summing over time dimension.
  times = torch.cumsum(intervals, dim=0).long()
  times[times >= time + 1] = 0
  
  # Create tensor of spikes.
  spikes = torch.zeros(time + 1, size).byte()
  spikes[times, torch.arange(size)] = 1
  spikes = spikes[1:]
  #-------------------------------------------------------------------
  
  return spikes.view(time, *shape)

这一段代码采用的是从真实Poisson分布中采样获得脉冲发放间隔，将时间截断到仿真步长后获得脉冲序列，因此相比用二项分布近似更精确，但也更慢。

[201528014227051]

明白了，感谢回复 @Yanqi-Chen