spikingjelly.configure module#

SpikingJelly package-level configuration.

API Language - 中文 | English

  • 中文

SpikingJelly 的库级选项通过 SJ_* 环境变量配置。环境变量在本模块首次导入时读取， 因此必须在启动 Python 前设置。未设置的选项使用下列文档中给出的默认值。

SJ_CUDA_THREADS=256 SJ_SAVE_DATASETS_COMPRESSED=0 python train.py

布尔环境变量遵循 PyTorch 的 0 / 1 风格，其中 0 表示 False， 1 表示 True。

  • English

SpikingJelly package-level options are configured through SJ_* environment variables. They are read when this module is first imported, so set them before starting Python. Unset options use the defaults documented below.

SJ_CUDA_THREADS=256 SJ_SAVE_DATASETS_COMPRESSED=0 python train.py

Boolean environment variables follow PyTorch's 0 / 1 convention: 0 means False and 1 means True.

spikingjelly.configure.max_threads_number_for_datasets_preprocess = 16#

数据集预处理使用的最大线程数，由 SJ_MAX_THREADS_NUMBER_FOR_DATASETS_PREPROCESS 配置，默认值为 16。

Maximum number of threads used for dataset preprocessing, configured by SJ_MAX_THREADS_NUMBER_FOR_DATASETS_PREPROCESS. The default is 16.

spikingjelly.configure.cuda_threads = 512#

CUDA kernel 默认使用的线程数，由 SJ_CUDA_THREADS 配置，默认值为 512。 建议使用 2 的幂。

Default number of threads used by CUDA kernels, configured by SJ_CUDA_THREADS. The default is 512; a power of two is recommended.

spikingjelly.configure.cuda_compiler_options = ('-use_fast_math',)#

传递给 NVRTC 或 NVCC 的编译选项，由 SJ_CUDA_COMPILER_OPTIONS 以逗号分隔的 形式配置，默认值为 ("-use_fast_math",)。空字符串表示不传递选项。

Compiler options passed to NVRTC or NVCC, configured as a comma-separated value in SJ_CUDA_COMPILER_OPTIONS. The default is ("-use_fast_math",); an empty value produces an empty tuple.

spikingjelly.configure.cuda_compiler_backend = 'nvrtc'#

CuPy CUDA 编译后端，由 SJ_CUDA_COMPILER_BACKEND 配置，允许 nvrtc 或 nvcc，默认值为 nvrtc。

CuPy CUDA compiler backend, configured by SJ_CUDA_COMPILER_BACKEND. Valid values are nvrtc and nvcc; the default is nvrtc.

spikingjelly.configure.save_datasets_compressed = True#

是否以压缩 NPZ 格式保存事件和帧，由 SJ_SAVE_DATASETS_COMPRESSED 配置， 默认值为 1。压缩格式占用更少磁盘空间，但读取更慢。

Whether to save events and frames as compressed NPZ files, configured by SJ_SAVE_DATASETS_COMPRESSED. The default is 1. Compression uses less disk space but takes longer to read.

spikingjelly.configure.save_spike_as_bool_in_neuron_kernel = False#

是否在神经元的 CuPy kernel 中以布尔形式保存反向传播所需的脉冲，由 SJ_SAVE_SPIKE_AS_BOOL_IN_NEURON_KERNEL 配置，默认值为 0。

Whether CuPy neuron kernels save spikes for backward as booleans, configured by SJ_SAVE_SPIKE_AS_BOOL_IN_NEURON_KERNEL. The default is 0.

spikingjelly.configure.save_bool_spike_level = 0#

布尔脉冲保存级别，由 SJ_SAVE_BOOL_SPIKE_LEVEL 配置，默认值为 0。 0 使用逐元素布尔值，1 将每 8 个脉冲打包到一个 uint8 中。

Boolean spike storage level, configured by SJ_SAVE_BOOL_SPIKE_LEVEL. The default is 0. Level 0 stores element-wise booleans; level 1 packs eight spikes into each uint8.

spikingjelly.configure.triton_neuron_kernel_static_range_max_T = 64#

Triton 多步神经元 kernel 使用 tl.static_range 的最大序列长度，由 SJ_TRITON_NEURON_KERNEL_STATIC_RANGE_MAX_T 配置，默认值为 64。超过该值时 使用 tl.range，以降低长序列的编译开销。

Maximum sequence length for which Triton multi-step neuron kernels use tl.static_range, configured by SJ_TRITON_NEURON_KERNEL_STATIC_RANGE_MAX_T. The default is 64. Larger sequences use tl.range to reduce compilation overhead.