es-imagenet-master

Latest News: Now you can use SpikingJelly to process ES-ImageNet!

API reference: ES-imageNet API

code for generating data set ES-ImageNet with corresponding training code

dataset generator

• some codes of ODG algorithm
• The variables to be modified include datapath (data storage path after transformation, which needs to be created before transformation) and root_Path (root directory of training set before transformation)

file name	function
traconvert.py	converting training set of ISLVRC 2012 into event stream using ODG
trainlabel_dir.txt	It stores the corresponding relationship between the class name and label of the original Imagenet file
trainlabel.txt	It is generated during transformation and stores the label of training set
valconvert.py	Transformation code for test set.
valorigin.txt	Original test label of ImageNet-1K. Put it in the same folder with valconvert.py if you need.
vallabel.txt	It is generated during transformation and stores the label of training set.

dataset usage

• codes are in ./datasets
• some traing examples are provided for ES-imagenet in ./example An example code for easily using this dataset based on Pytorch

from __future__ import print_function
import sys
sys.path.append("..")
from datasets.es_imagenet_new import ESImagenet_Dataset
import torch.nn as nn
import torch

data_path = None #TODO:modify 
train_dataset = ESImagenet_Dataset(mode='train',data_set_path=data_path)
test_dataset = ESImagenet_Dataset(mode='test',data_set_path=data_path)

train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
test_sampler  = torch.utils.data.distributed.DistributedSampler(test_dataset)
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=False, num_workers=1,pin_memory=True,drop_last=True,sampler=train_sampler)
test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=batch_size, shuffle=False, num_workers=1,pin_memory=True)

for batch_idx, (inputs, targets) in enumerate(train_loader)
  pass
  # input = [batchsize,time,channel,width,height]
  
for batch_idx, (inputs, targets) in enumerate(test_loader):
  pass
  # input = [batchsize,time,channel,width,height]

training example and benchmarks

Requirements

• Python >= 3.5
• Pytorch >= 1.7
• CUDA >=10.0
• TenosrBoradX(optional)

Train the baseline models

$ cd example

$ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 example_ES_res18.py #LIAF/LIF-ResNet-18
$ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 example_ES_res34.py #LIAF/LIF-ResNet-34
$ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 compare_ES_3DCNN34.py #3DCNN-ResNet-34
$ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 compare_ES_3DCNN18.py #3DCNN-ResNet-18
$ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 compare_ES_2DCNN34.py #2DCNN-ResNet-34 
$ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 compare_ES_2DCNN18.py #2DCNN-ResNet-18
$ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 compare_CONVLSTM.py #ConvLSTM (no used in paper)
$ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 example_ES_res50.py #LIAF/LIF-ResNet-50 (no used in paper)

** note:** To select LIF mode, change the config files under /LIAFnet : self.actFun= torch.nn.LeakyReLU(0.2, inplace=False) #nexttest:selu to self.actFun= LIAF.LIFactFun.apply

baseline / Benchmark

Network	layer	Type Test Acc/%	# of Para	FP32+/GFLOPs	FP32x/GFLOPs
ResNet18	2D-CNN	41.030	11.68M	1.575	1.770
ResNet18	3D-CNN	38.050	28.56M	12.082	12.493
ResNet18	LIF	39.894	11.69M	12.668	0.269
ResNet18	LIAF	42.544	11.69M	12.668	14.159
ResNet34	2D-CNN	42.736	21.79M	3.211	3.611
ResNet34	3D-CNN	39.410	48.22M	20.671	21.411
ResNet34	LIF	43.424	21.80M	25.783	0.288
ResNet18+imagenet-pretrain (a)	LIF	43.74	11.69M	12.668	0.269
ConvECLIF2D-A	ECLIF	44.25	17.99M	-	-
Surrogate Module	LIF	44.76	-	-	-
ResNet34	LIAF	47.466	21.80M	25.783	28.901
ResNet18+self-pretrain	LIAF	50.54	11.69M	12.668	14.159
ResNet18+imagenet-pretrain (b)	LIAF	52.25	11.69M	12.668	14.159
ResNet34+imagenet-pretrain (c)	LIAF	51.83	21.80M	25.783	28.901

Note: model (a), (b) and (c) are stored in ./pretrained_model Some problems related to model loading can be referred to the issue. If you want to test 2D-CNN with reconstructed gray set, you can use this notebook

Download

• The datasets ES-ImageNet (100GB) for this study can be download in the Tsinghua Cloud1 or Openl

• The converted event-frame version (40GB, events are organized as 8 event-frames instead of lists) can be found in Tsinghua Cloud2, IT MAY BE BUGGY, STILL NEED TO BE CHECKED

• If you only need the validation set, you can download it in Tsinghua Cloud3 separately

Citation

If you find this code useful in your research, please consider citing and here is an example BibTeX entry:

@ARTICLE{ES_ImageNet2021,
AUTHOR={Lin, Yihan and Ding, Wei and Qiang, Shaohua and Deng, Lei and Li, Guoqi},   
TITLE={ES-ImageNet: A Million Event-Stream Classification Dataset for Spiking Neural Networks},      
JOURNAL={Frontiers in Neuroscience},      
VOLUME={15},      
PAGES={1546},     
YEAR={2021},      	  
URL={https://www.frontiersin.org/article/10.3389/fnins.2021.726582},       	
DOI={10.3389/fnins.2021.726582},      
ISSN={1662-453X},   
}