我正在尝试使用FFT加速神经模拟器的计算。
等式是:
(1)\ sum(j = 1到N)(w(i - j)* s_NMDA [j])
其中s_NMDA是长度为N的向量,w由以下定义:
(2)w(j)= tanh [1 /(2 * sigma * p)] * exp(-abs(j)/(sigma * p)]
其中sigma和p是常量。
(有没有更好的方法在stackoverflow上渲染方程?)
必须对N个神经元进行计算。由于(1)仅取决于绝对距离abs(i-j),因此应该可以使用FFT(卷积定理)计算它。
我尝试使用FFTW实现此功能,但结果与预期结果不符。我之前从未使用过FFTW,现在我不确定如果我对卷积定理的假设是假的,我的实现是不正确的。
void f_I_NMDA_FFT(
const double **states, // states[i][6] == s_NMDA[i]
const unsigned int numNeurons)
{
fftw_complex *distances, *sNMDAs, *convolution;
fftw_complex *distances_f, *sNMDAs_f, *convolution_f;
fftw_plan p, pinv;
const double scale = 1./numNeurons;
distances = (fftw_complex *)fftw_malloc(sizeof(fftw_complex) * numNeurons);
sNMDAs = (fftw_complex *)fftw_malloc(sizeof(fftw_complex) * numNeurons);
convolution = (fftw_complex *)fftw_malloc(sizeof(fftw_complex) * numNeurons);
distances_f = (fftw_complex *)fftw_malloc(sizeof(fftw_complex) * numNeurons);
sNMDAs_f = (fftw_complex *)fftw_malloc(sizeof(fftw_complex) * numNeurons);
convolution_f = (fftw_complex *)fftw_malloc(sizeof(fftw_complex) * numNeurons);
// fill input array for distances
for (unsigned int i = 0; i < numNeurons; ++i)
{
distances[i][0] = w(i);
distances[i][1] = 0;
}
// fill input array for sNMDAs
for (unsigned int i = 0; i < numNeurons; ++i)
{
sNMDAs[i][0] = states[i][6];
sNMDAs[i][1] = 0;
}
p = fftw_plan_dft_1d(numNeurons,
distances,
distances_f,
FFTW_FORWARD,
FFTW_ESTIMATE);
fftw_execute(p);
p = fftw_plan_dft_1d(numNeurons,
sNMDAs,
sNMDAs_f,
FFTW_FORWARD,
FFTW_ESTIMATE);
fftw_execute(p);
// convolution in frequency domain
for(unsigned int i = 0; i < numNeurons; ++i)
{
convolution_f[i][0] = (distances_f[i][0] * sNMDAs_f[i][0]
- distances_f[i][1] * sNMDAs_f[i][1]) * scale;
convolution_f[i][1] = (distances_f[i][0] * sNMDAs_f[i][1]
- distances_f[i][1] * sNMDAs_f[i][0]) * scale;
}
pinv = fftw_plan_dft_1d(numNeurons,
convolution_f,
convolution,
FFTW_FORWARD,
FFTW_ESTIMATE);
fftw_execute(pinv);
// compute and compare with expected result
for (unsigned int i = 0; i < numNeurons; ++i)
{
double expected = 0;
for (int j = 0; j < numNeurons; ++j)
{
expected += w(i - j) * states[j][6];
}
printf("i=%d, FFT: r%f, i%f : Expected: %f\n", i, convolution[i][0], convolution[i][1], expected);
}
fftw_destroy_plan(p);
fftw_destroy_plan(pinv);
fftw_free(distances), fftw_free(sNMDAs), fftw_free(convolution);
fftw_free(distances_f), fftw_free(sNMDAs_f), fftw_free(convolution_f);
这是20个神经元的示例输出:
i=0, FFT: r0.042309, i0.000000 : Expected: 0.041504
i=1, FFT: r0.042389, i0.000000 : Expected: 0.042639
i=2, FFT: r0.042466, i0.000000 : Expected: 0.043633
i=3, FFT: r0.042543, i0.000000 : Expected: 0.044487
i=4, FFT: r0.041940, i0.000000 : Expected: 0.045203
i=5, FFT: r0.041334, i0.000000 : Expected: 0.045963
i=6, FFT: r0.041405, i0.000000 : Expected: 0.046585
i=7, FFT: r0.041472, i0.000000 : Expected: 0.047070
i=8, FFT: r0.041537, i0.000000 : Expected: 0.047419
i=9, FFT: r0.041600, i0.000000 : Expected: 0.047631
i=10, FFT: r0.041660, i0.000000 : Expected: 0.047708
i=11, FFT: r0.041717, i0.000000 : Expected: 0.047649
i=12, FFT: r0.041773, i0.000000 : Expected: 0.047454
i=13, FFT: r0.041826, i0.000000 : Expected: 0.047123
i=14, FFT: r0.041877, i0.000000 : Expected: 0.046656
i=15, FFT: r0.041926, i0.000000 : Expected: 0.046052
i=16, FFT: r0.041294, i0.000000 : Expected: 0.045310
i=17, FFT: r0.042059, i0.000000 : Expected: 0.044430
i=18, FFT: r0.042144, i0.000000 : Expected: 0.043412
i=19, FFT: r0.042228, i0.000000 : Expected: 0.042253
结果似乎几乎是正确的,但误差随着神经元的数量而增加。而且,对于非常低或非常高的位置(i),结果似乎更准确。这里发生了什么?
更新:正如Oli Charlesworth所建议的那样,我在八度音程中实现了算法,看它是否是一个实现或数学问题:
input = [0.186775; 0.186775; 0.186775; 0.186775; 0.186775; 0; 0.186775; 0.186775; 0.186775; 0.186775];
function ret = _w(i)
ret = tanh(1 / (2* 1 * 32)) * exp(-abs(i) / (1 * 32));
end
for i = linspace(1, 10, 10)
expected = 0;
for j = linspace(1, 10, 10)
expected += _w(i-j) * input(j);
end
expected
end
distances = _w(transpose(linspace(0, 9, 10)));
input_f = fft(input);
distances_f = fft(distances);
convolution_f = input_f .* distances_f;
convolution = ifft(convolution_f)
结果:
expected = 0.022959
expected = 0.023506
expected = 0.023893
expected = 0.024121
expected = 0.024190
expected = 0.024100
expected = 0.024034
expected = 0.023808
expected = 0.023424
expected = 0.022880
convolution =
0.022959
0.023036
0.023111
0.023183
0.023253
0.022537
0.022627
0.022714
0.022798
0.022880
结果非常相似。因此,我对卷积定理/ FFT的理解肯定有问题。
答案 0 :(得分:8)
要通过FFT卷积2个信号,通常需要这样做:
在您的代码中,我在所有3个FFT中都看到FFTW_FORWARD
。我猜这不是问题,它是它的一部分。最后一个FFT应该是“向后”,而不是“向前”。
另外,我认为你在第二个表达式中需要“+”,而不是“ - ”:
convolution_f[i][0] = (distances_f[i][0] * sNMDAs_f[i][0]
- distances_f[i][1] * sNMDAs_f[i][1]) * scale;
convolution_f[i][1] = (distances_f[i][0] * sNMDAs_f[i][1]
- distances_f[i][1] * sNMDAs_f[i][0]) * scale;
答案 1 :(得分:1)
我终于解决了这个问题,非常感谢Alex和Oli Charlesworth的建议!
function ret = _w(i)
ret = tanh(1 / (2* 1 * 32)) * exp(-abs(i) / (1 * 32));
end
_input = [0.186775; 0.186775; 0.186775; 0.186775; 0.186775; 0; 0.186775; 0.186775; 0.186775; 0.186775];
n = size(_input)(1);
input = _input;
for i = linspace(1, n, n)
expected = 0;
for j = linspace(1, n, n)
expected += _w(i-j) * input(j);
end
expected
end
input = vertcat(_input, zeros((2*n)-n-1,1));
distances = _w(transpose(linspace(0, (2*n)-n-1, n)));
distances = vertcat(flipud(distances), distances(2:end));
input_f = fft(input);
distances_f = fft(distances);
convolution_f = input_f .* distances_f;
convolution = ifft(convolution_f)(n:end)
结果:
expected = 0.022959
expected = 0.023506
expected = 0.023893
expected = 0.024121
expected = 0.024190
expected = 0.024100
expected = 0.024034
expected = 0.023808
expected = 0.023424
expected = 0.022880
convolution =
0.022959
0.023506
0.023893
0.024121
0.024190
0.024100
0.024034
0.023808
0.023424
0.022880
我基本上忘了以正确的方式命令距离数组。如果有人有兴趣,我可以在以后提供更详细的解释。
更新:(说明)
这是我的距离矢量(对于5个神经元)最初看起来很喜欢的东西:
i = 1 2 3 4 5
| _w(0) | _w(1) | _w(2) | _w(3) | _w(4) |
在这个载体上,我应用了一个内核,例如:
| 0.1 | 0.1 | 0.0 | 0.2 | 0.3 |
由于我使用循环卷积,第一个神经元_w(0)的结果是:
0.0 * _w(2)+ 0.1 * _w(1)+ 0.1 * _w(0)+ 0.1 * _w(1)+ 0.0 * _w(2)
但这是不正确的,结果应该是:
0.1 * _w(0)+ 0.1 * _w(1)+ 0.0 * _w(2)+ 0.2 * _w(3)+ 0.3 * _w(4)
为了实现这一点,我必须“镜像”我的距离向量并在内核中添加一些填充:
input = vertcat(_input, zeros((2*n)-n-1,1));
distances = _w(transpose(linspace(0, (2*n)-n-1, n)));
distances = _w(transpose(linspace(0, (2*n)-n-1, n)));
i = 1 2 3 4 5 6 7 8 9
| _w(4) | _w(3) | _w(2) | _w(1) | _w(0) | _w(1) | _w(2) | _w(3) | _w(4) |
| 0.1 | 0.1 | 0.0 | 0.2 | 0.3 | 0 | 0 | 0 | 0 |
现在,如果我应用卷积,i = [5:9]的结果正是我想要的结果,所以我只需要丢弃[1:4]并且我已经完成了:)
convolution = ifft(convolution_f)(n:end)