pytorch-sbs/README.md
2022-05-01 17:06:56 +02:00

6.5 KiB
Raw Blame History

pytorch-sbs

SbS Extension for PyTorch

Based on these scientific papers

Back-Propagation Learning in Deep Spike-By-Spike Networks
David Rotermund and Klaus R. Pawelzik
Front. Comput. Neurosci., https://doi.org/10.3389/fncom.2019.00055
https://www.frontiersin.org/articles/10.3389/fncom.2019.00055/full

Efficient Computation Based on Stochastic Spikes
Udo Ernst, David Rotermund, and Klaus Pawelzik
Neural Computation (2007) 19 (5): 13131343. https://doi.org/10.1162/neco.2007.19.5.1313
https://direct.mit.edu/neco/article-abstract/19/5/1313/7183/Efficient-Computation-Based-on-Stochastic-Spikes

Python

It was programmed with 3.10.4. And I used some 3.10 Python expression. Thus you might get problems with older Python versions.

C++

It works without compiling the C++ modules. However it is 10x slower.
You need to modify the Makefile in the C++ directory to your Python installation.
In addition yoir Python installation needs the PyBind11 package installed. You might want to perform a
pip install pybind11
The Makefile uses clang as a compiler. If you want something else then you need to change the Makefile. The SbS.py autodetectes if the required C++ .so modules are in the same directory as the SbS.py file.

SbS layer class

Variables

epsilon_xy  
epsilon_0  
epsilon_t  
weights  
kernel_size  
stride  
dilation  
padding  
output_size  
number_of_spikes  
number_of_cpu_processes  
number_of_neurons  
number_of_input_neurons  
h_initial  
alpha_number_of_iterations  

Constructor

def __init__(  
    self,  
    number_of_input_neurons: int,  
    number_of_neurons: int,  
    input_size: list[int],  
    forward_kernel_size: list[int],  
    number_of_spikes: int,  
    epsilon_t: torch.Tensor,  
    epsilon_xy_intitial: float = 0.1,  
    epsilon_0: float = 1.0,  
    weight_noise_amplitude: float = 0.01,  
    is_pooling_layer: bool = False,  
    strides: list[int] = [1, 1],  
    dilation: list[int] = [0, 0],  
    padding: list[int] = [0, 0],  
    alpha_number_of_iterations: int = 0,  
    number_of_cpu_processes: int = 1,  
) -> None:

Methods

def initialize_weights(  
    self,  
    is_pooling_layer: bool = False,  
    noise_amplitude: float = 0.01,  
) -> None:  

For the generation of the initital weights. Switches between normal initial random weights and pooling weights.


def initialize_epsilon_xy(  
    self,  
    eps_xy_intitial: float,  
) -> None:  

Creates initial epsilon xy matrices.


def set_h_init_to_uniform(self) -> None:  

def backup_epsilon_xy(self) -> None:  
def restore_epsilon_xy(self) -> None:  
def backup_weights(self) -> None:  
def restore_weights(self) -> None:  

def threshold_epsilon_xy(self, threshold: float) -> None:  
def threshold_weights(self, threshold: float) -> None:  

def mean_epsilon_xy(self) -> None:  

def norm_weights(self) -> None:

Parameters in JSON file

data_mode: str = field(default="")  

data_path: str = field(default="./")


batch_size: int = field(default=500)


learning_step: int = field(default=0)


learning_step_max: int = field(default=10000)



number_of_cpu_processes: int = field(default=-1)



number_of_spikes: int = field(default=0)


cooldown_after_number_of_spikes: int = field(default=0)



weight_path: str = field(default="./Weights/")


eps_xy_path: str = field(default="./EpsXY/")

    

reduction_cooldown: float = field(default=25.0)


epsilon_0: float = field(default=1.0)



update_after_x_batch: float = field(default=1.0)



## network_structure (required!)
Parameters of the network. The details about its layers and the number of output neurons.  


number_of_output_neurons: int = field(default=0)


forward_neuron_numbers: list[list[int]] = field(default_factory=list)


is_pooling_layer: list[bool] = field(default_factory=list)



forward_kernel_size: list[list[int]] = field(default_factory=list)


strides: list[list[int]] = field(default_factory=list)


dilation: list[list[int]] = field(default_factory=list)


padding: list[list[int]] = field(default_factory=list)



w_trainable: list[bool] = field(default_factory=list)


eps_xy_trainable: list[bool] = field(default_factory=list)


eps_xy_mean: list[bool] = field(default_factory=list)



## learning_parameters
Parameter required for training   


learning_active: bool = field(default=True)



loss_coeffs_mse: float = field(default=0.5)


loss_coeffs_kldiv: float = field(default=1.0)



optimizer_name: str = field(default="Adam")


learning_rate_gamma_w: float = field(default=-1.0)


learning_rate_gamma_eps_xy: float = field(default=-1.0)


learning_rate_threshold_w: float = field(default=0.00001)


learning_rate_threshold_eps_xy: float = field(default=0.00001)



lr_schedule_name: str = field(default="ReduceLROnPlateau")


lr_scheduler_factor_w: float = field(default=0.75)


lr_scheduler_patience_w: int = field(default=-1)



lr_scheduler_factor_eps_xy: float = field(default=0.75)


lr_scheduler_patience_eps_xy: int = field(default=-1)



number_of_batches_for_one_update: int = field(default=1)


overload_path: str = field(default="./Previous")



weight_noise_amplitude: float = field(default=0.01)


eps_xy_intitial: float = field(default=0.1)



test_every_x_learning_steps: int = field(default=50)


test_during_learning: bool = field(default=True)



alpha_number_of_iterations: int = field(default=0)


## augmentation
Parameters used for data augmentation.  


crop_width_in_pixel: int = field(default=2)



flip_p: float = field(default=0.5)



jitter_brightness: float = field(default=0.5)


jitter_contrast: float = field(default=0.1)


jitter_saturation: float = field(default=0.1)


jitter_hue: float = field(default=0.15)



use_on_off_filter: bool = field(default=True)


## ImageStatistics (please ignore)
(Statistical) information about the input. i.e. mean values and the x and y size of the input  


mean: list[float] = field(default_factory=list)


the_size: list[int] = field(default_factory=list)