Spike Compressors#

将浮点数表示的脉冲张量转换为更紧凑的表示形式的压缩器。

Compressors that convert spike tensors represented in floating-point numbers into more compact representations.

class spikingjelly.activation_based.memopt.compress.BaseSpikeCompressor[源代码]#

基类：ABC

API Language: 中文 | English

中文
中文

脉冲压缩器的抽象基类。欲实现脉冲压缩器，需继承该抽象基类并实现 _compress 和 _decompress 方法。

English
English

Abstract base class for spike compressors. To implement a spike compressor, you need to inherit this abstract base class and implement the _compress and _decompress methods. :return: None :rtype: None

requires_strictly_binary = False#

compress(s_seq: Tensor) → Tensor[源代码]#

API Language: 中文 | English

中文

压缩缩脉冲序列。

参数:: s_seq (torch.Tensor) -- 输入脉冲序列
返回:: 压缩后的脉冲序列
返回类型:: torch.Tensor

English

Compress spike sequence.

参数:: s_seq (torch.Tensor) -- input spike sequence
返回:: compressed spike sequence
返回类型:: torch.Tensor

decompress(s_seq: Tensor, shape) → Tensor[源代码]#

API Language: 中文 | English

中文

解压缩脉冲序列。

参数:

s_seq (torch.Tensor) -- 压缩的脉冲序列
shape (tuple or torch.Size) -- 原始形状

返回:

解压缩后的脉冲序列

返回类型:

torch.Tensor

English

Decompress spike sequence.

参数:

s_seq (torch.Tensor) -- compressed spike sequence
shape (tuple or torch.Size) -- original shape

返回:

decompressed spike sequence

返回类型:

torch.Tensor

class spikingjelly.activation_based.memopt.compress.NullSpikeCompressor[源代码]#

基类：BaseSpikeCompressor

API Language: 中文 | English

中文
中文

空脉冲压缩器。压缩和解压缩过程都是恒等映射。

NullSpikeCompressor 是唯一能够无损处理非二进制张量的"脉冲压缩器"模块。例如，SNN的输入层应该始终使用 NullSpikeCompressor ，因为其输入是浮点张量而不是二值张量。

English
English

Null spike compressor. The compression and decompression process are identity mapping.

NullSpikeCompressor is the only compressor module that can deal with non-binary tensors losslessly. For instance, the input layer should always use NullSpikeCompressor, as its input is a float tensor rather than a binary tensor.

代码示例 | Example

import torch
from spikingjelly.activation_based.memopt.compress import (
    NullSpikeCompressor,
)

compressor = NullSpikeCompressor()
x = torch.randn(32, 10)
compressed = compressor.compress(x)
decompressed = compressor.decompress(compressed, x.shape)

返回:: None
返回类型:: None

requires_strictly_binary = False#

class spikingjelly.activation_based.memopt.compress.BooleanSpikeCompressor[源代码]#

基类：BaseSpikeCompressor

API Language: 中文 | English

中文
中文

布尔脉冲压缩器。

将脉冲序列转换为布尔类型以节省内存。要求输入必须是严格的二进制脉冲。

English
English

Boolean spike compressor.

Convert spike sequences to boolean type to save memory. Requires input to be strictly binary spikes.

代码示例 | Example

import torch
from spikingjelly.activation_based.memopt.compress import (
    BooleanSpikeCompressor,
)

compressor = BooleanSpikeCompressor()
spikes = torch.randint(0, 2, (32, 100)).float()
compressed = compressor.compress(spikes)
decompressed = compressor.decompress(compressed, spikes.shape)

返回:: None
返回类型:: None

requires_strictly_binary = True#

class spikingjelly.activation_based.memopt.compress.Uint8SpikeCompressor[源代码]#

基类：BaseSpikeCompressor

API Language: 中文 | English

中文
中文

Uint8脉冲压缩器。

将脉冲序列转换为uint8类型以节省内存。可以处理非二进制整数数值。

English
English

Uint8 spike compressor.

Convert spike sequences to uint8 type to save memory. Can handle non-binary integer values.

代码示例 | Example

import torch
from spikingjelly.activation_based.memopt.compress import (
    Uint8SpikeCompressor,
)

compressor = Uint8SpikeCompressor()
x = torch.randn(32, 10)
compressed = compressor.compress(x)
decompressed = compressor.decompress(compressed, x.shape)

返回:: None
返回类型:: None

requires_strictly_binary = False#

class spikingjelly.activation_based.memopt.compress.BitSpikeCompressor[源代码]#

基类：BaseSpikeCompressor

API Language: 中文 | English

中文
中文

比特脉冲压缩器。

使用位压缩技术将8个二进制脉冲压缩到一个字节中，实现极高的内存压缩比。要求输入必须是严格的二进制脉冲（0或1）。

English
English

Bit-level spike compressor.

Use bit compression technique to compress 8 binary spikes into one byte, achieving high memory compression ratio. Requires input to be strictly binary spikes (0 or 1).

代码示例 | Example

import torch
from spikingjelly.activation_based.memopt.compress import BitSpikeCompressor

compressor = BitSpikeCompressor()
spikes = torch.randint(0, 2, (32, 1000)).float()
compressed = compressor.compress(spikes)
decompressed = compressor.decompress(compressed, spikes.shape)

返回:: None
返回类型:: None

requires_strictly_binary = True#

class spikingjelly.activation_based.memopt.compress.SparseSpikeCompressor(dtype=torch.int64)[源代码]#

基类：BaseSpikeCompressor

API Language: 中文 | English

中文
中文

稀疏脉冲压缩器。

只存储非零脉冲的位置索引，适用于稀疏脉冲序列。要求输入必须是严格的二进制脉冲（0或1）。

参数:: dtype (torch.dtype) -- 索引数据类型，默认为 torch.int64

English
English

Sparse spike compressor.

Only store the position indices of non-zero spikes, suitable for sparse spike sequences. Requires input to be strictly binary spikes (0 or 1).

参数:: dtype (torch.dtype) -- index data type. Default to torch.int64

代码示例 | Example

import torch
from spikingjelly.activation_based.memopt.compress import (
    SparseSpikeCompressor,
)

compressor = SparseSpikeCompressor()
spikes = (torch.rand(32, 1000) < 0.04).float()
compressed = compressor.compress(spikes)
decompressed = compressor.decompress(compressed, spikes.shape)

返回:: None
返回类型:: None

requires_strictly_binary = True#