Skip to content

Version 2.4.6
Compare
Choose a tag to compare
@chaoming0625 chaoming0625 released this 11 Nov 02:43
· 152 commits to master since this release
V2.4.6
92cdc40
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Learn about vigilant mode.

This release contains more than 130 commit updates, and has provided several new features.

New Features

1. surrogate gradient functions are more transparent.

New instances can be used to compute the surrogate gradients. For example:

import brainpy.math as bm
fun = bm.surrogate.Sigmoid()

# forward function
spk = fun(membrane_potential)

# backward function
dV = fun.surrogate_grad(1., membrane_potential)

# surrogate forward function
surro_spk = fun.surrogate_fun(membrane_potential)

2. Add brainpy.math.eval_shape for evaluating the all dynamical variables used in the target function.

This function is similar to jax.eval_shape which has no FLOPs, while it can extract all variables used in the target function. For example:

net = ...  # any dynamical system
inputs = ...  # inputs to the dynamical system
variables, outputs= bm.eval_shape(net, inputs)  
# "variables" are all variables used in the target "net"

In future, this function will be used everywhere to transform all jax transformations into brainpy's oo transformations.

3. Generalize tools and interfaces for state managements.

For a single object:

  • The .reset_state() defines the state resetting of all local variables in this node.
  • The .load_state() defines the state loading from external disks (typically, a dict is passed into this .load_state() function).
  • The .save_state() defines the state saving to external disks (typically, the .save_state() function generates a dict containing all variable values).

Here is an example to define a full class of brainpy.DynamicalSystem.

import brainpy as bp

class YouDynSys(bp.DynamicalSystem):
   def __init__(self, ):  # define parameters
      self.par1 = ....
      self.num = ...

  def reset_state(self, batch_or_mode=None):  # define variables
     self.a = bp.init.variable_(bm.zeros, (self.num,), batch_or_mode)

  def load_state(self, state_dict):  # load states from an external dict
     self.a.value = bm.as_jax(state_dict['a'])

  def save_state(self):  # save states as an external dict
     return {'a': self.a.value}

For a complex network model, brainpy provide unified state managment interface for initializing, saving, and loading states.

  • The brainpy.reset_state() defines the state resetting of all variables in this node and its children nodes.
  • The brainpy.load_state() defines the state loading from external disks of all variables in the node and its children.
  • The brainpy.save_state() defines the state saving to external disks of all variables in the node and its children.
  • The brainpy.clear_input() defines the clearing of all input variables in the node and its children.

4. Unified brain simulation and brain-inspired computing interface through automatic membrane scaling.

The same model used in brain simulation can be easily transformed into the one used for brain-inspired computing for training. For example,

class EINet(bp.DynSysGroup):
  def __init__(self):
    super().__init__()
    self.N = bp.dyn.LifRefLTC(4000, V_rest=-60., V_th=-50., V_reset=-60., tau=20., tau_ref=5.,
                              V_initializer=bp.init.Normal(-55., 2.))
    self.delay = bp.VarDelay(self.N.spike, entries={'I': None})
    self.E = bp.dyn.ProjAlignPost1(
      comm=bp.dnn.EventCSRLinear(bp.conn.FixedProb(0.02, pre=3200, post=4000), weight=bp.init.Normal(0.6, 0.01)),
      syn=bp.dyn.Expon(size=4000, tau=5.),
      out=bp.dyn.COBA(E=0.),
      post=self.N
    )
    self.I = bp.dyn.ProjAlignPost1(
      comm=bp.dnn.EventCSRLinear(bp.conn.FixedProb(0.02, pre=800, post=4000), weight=bp.init.Normal(6.7, 0.01)),
      syn=bp.dyn.Expon(size=4000, tau=10.),
      out=bp.dyn.COBA(E=-80.),
      post=self.N
    )

  def update(self, input):
    spk = self.delay.at('I')
    self.E(spk[:3200])
    self.I(spk[3200:])
    self.delay(self.N(input))
    return self.N.spike.value


# used for brain simulation
with bm.environment(mode=bm.nonbatching_mode):
  net = EINet()


# used for brain-inspired computing
# define the `membrane_scaling` parameter
with bm.environment(mode=bm.TrainingMode(128), membrane_scaling=bm.Scaling.transform([-60., -50.])):
  net = EINet()

5. New apis for operator customization on CPU and GPU devices through brainpy.math.XLACustomOp.

Starting from this release, brainpy introduces Taichi for operator customization. Now, users can write CPU and GPU operators through numba and taichi syntax on CPU device, and taichi syntax on GPu device. Particularly, to define an operator, user can use:

import numba as nb
import taichi as ti
import numpy as np
import jax
import brainpy.math as bm


@nb.njit
def numba_cpu_fun(a, b, out_a, out_b):
  out_a[:] = a
  out_b[:] = b


@ti.kernel
def taichi_gpu_fun(a, b, out_a, out_b):
  for i in range(a.size):
    out_a[i] = a[i]
  for i in range(b.size):
    out_b[i] = b[i]


prim = bm.XLACustomOp(cpu_kernel=numba_cpu_fun, gpu_kernel=taichi_gpu_fun)
a2, b2 = prim(np.random.random(1000), np.random.random(1000),
              outs=[jax.ShapeDtypeStruct(1000, dtype=np.float32),
                    jax.ShapeDtypeStruct(1000, dtype=np.float32)])

6. Generalized STDP models which are compatible with diverse synapse models.

See https://github.com/brainpy/BrainPy/blob/master/brainpy/_src/dyn/projections/tests/test_STDP.py

What's Changed

New Contributors

Full Changelog: V2.4.5...V2.4.6

Contributors

chaoming0625, HoshinoKoji, and 2 other contributors
Assets 2
Loading