markusai / ra9

JAX-based Spiking Neural Network framework

rA9

---

Spiking Neural Network Library Based on JAX and referencing codes from bintorch

The learning algorithm of this library is spike-based backpropagation proposed from Enabling Spike-Based Backpropagation for Training Deep Neural Network Architectures

Compatitable Operating Systems

Only supports Linux and MacOS, because of the dependency of JAX and they try to fix it. But you can run this library in WSL

Installation

MacOS

CPU

Simple, just type

pip install git+https://github.com/MarkusAI/rA9

GPU

MacOS does not support NVIDIA CUDA.

Linux

CPU

Simple, just type

pip install git+https://github.com/MarkusAI/rA9

GPU

You need to setup JAX before installing the rA9 as GPU-dedicated and install rA9 as following:

pip install git+https://github.com/MarkusAI/rA9

Example

Check LeNet.py in example

ra9's Issues

Need to make Spike stochastic Gradient Descent Optimizer

Same as subject

Engine is the problem

when you see engine and function code,
output didn't work because of:
gamma is not the element of output function
so @jepetolee made not e/T but e @ e (absent means matrix multiplication)

This part should be resolved.

When v_current is set to zero, then it sustains the zero

same as subject

img2col error

Describe the bug
img2col error when index add

To Reproduce
Steps to reproduce the behavior:

1. Just run test.py on example

Expected behavior
work really well in test.py

Performance Enhancement

img2col.py in nn/_functions/
dataloader.py in utils/data/

Because some parts of img2col use numpy not jax.numpy, and dataloader is based on python so it is quite slow

graident didn't work

It didn't work

The Problem of Loss Function

We might don't think about the basic opponents, Function and Variable and Parameter.

Parameter did well on class but function didn't and actually loss is defined as the function.(or inherited)

Please Check this Issue for fixing it.

Loss function Bug

Describe the bug
Loss function don't work when the batch size under the 10

To Reproduce
Steps to reproduce the behavior:

1. Go to examples/cnn_mnist.py
2. Change the batch_size under the 10
3. Execute the cnn_mnist.py
4. See error

Expected behavior
Will raise TypeError

@junhoyeo Please check this issue

Dropout Error

Describe the bug
Dropout Error

To Reproduce
Steps to reproduce the behavior:

1. Run LeNet

Expected behavior
Variable shape

memory leak

memory leak problem occurred

in activation function

Output NaN problem

Describe the bug
Output NaN problem

default to find gamma in Variables

I think you didn't commit your latest backend, so if you read this Issue, give me changed backend codes to usb.

Encoding Problem

Describe the bug
When it goes another data batch, then it fails to generate full spike

The poisson Encoder might be wrecked by me LOL

The design of Poison encoder is wrecked i think.

SpikeLoss grad function undefined

Same as above

The Pencoder code is different

Poisson Encoder is different when comparing the original paper

Convolution layer& Pooling Layer

When you see the paper(Enabling Spike-based Backpropagation for Training Deep Neural Network Architectures), you can see the Convolution & pooling layers are based on LIF Neurons..

So we need to change the convolution and pooling layer into LIF Neurons...

TPU_Cloud_support

Is your feature request related to a problem? Please describe.
TPU Cloud support

Describe the solution you'd like
Need to think

Describe alternatives you've considered
Nothing

Additional context
Nothing

Optimizer didn't work

Optimizer didn't work so I and my partner will fix it

Dropout module

AttributeError: module 'jax.numpy' has no attribute 'random'

TypeError: Shapes must be 1D sequences of concrete values of integer type

Running with #4

CANT FING BUGS!!!!

in rA9/autograd/function.py, the Backward_function()doesn't get grads on output.py

Voltage rest value undefined.

voltage rest value is undefined and i think this is the problem

김 선생님

최대한 논문 보고 backward를 따라 해 보았는데, 받아줘야하는 weight 형식이라던지 업데이트는 bintorch 기준으로만 하실 겁니까?

Using grad to backward

Is your feature request related to a problem? Please describe.
Not just a function to solve, but just use grad function
Describe the solution you'd like
Using jax grad

Describe alternatives you've considered
Nah

Additional context
Nah

TPU_Colab version

Is your feature request related to a problem? Please describe.
TPU_Colab_Version

Describe the solution you'd like
Using JAX example

Describe alternatives you've considered
No

Additional context
Nothing