-
Notifications
You must be signed in to change notification settings - Fork 50
/
convolve.py
227 lines (195 loc) · 8.8 KB
/
convolve.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
import numpy as np
from scipy import fftpack
# numbers of the form 2^n3^m5^r, even only and r<=1
fastFFT = np.array(
[2, 4, 6, 8, 10, 12, 16, 20, 24, 32, 40, 48, 64, 80, 96, 128, 144, 160, 192, 256, 288, 320, 384, 432, 480,
512, 576, 640, 720, 768, 864, 960, 1024, 1152, 1280, 1440, 1536, 1728, 1920, 2048, 2304, 2560, 2880, 3072,
3456, 3840, 4096, 4608, 5120, 5760, 6144, 6912, 7680, 8192, 9216, 10240, 11520, 12288, 13824, 15360, 16384,
18432, 20480, 23040, 24576, 27648, 30720, 32768, 36864, 40960, 46080, 49152, 55296, 61440, 65536, 73728, 81920,
92160, 98304, 110592, 122880, 131072, 147456, 163840, 184320, 196608, 221184, 245760, 262144, 294912, 327680,
368640, 393216, 442368, 491520, 524288, 589824, 655360, 737280, 786432, 884736, 983040, 1048576, 1179648,
1310720, 1474560, 1572864, 1769472, 1966080, 2097152, 2359296, 2621440, 2949120, 3145728, 3538944, 3932160,
4194304, 4718592, 5242880, 5898240, 6291456, 7077888, 7864320, 8388608, 9437184, 10485760, 11796480, 12582912,
14155776, 15728640, 16777216, 18874368, 20971520, 23592960, 25165824, 28311552, 31457280, 33554432, 37748736,
41943040, 47185920, 50331648, 56623104, 62914560, 67108864, 75497472, 83886080, 94371840, 100663296, 113246208,
125829120, 134217728, 150994944, 167772160, 188743680, 201326592, 226492416, 234881024, 251658240, 268435456,
301989888, 335544320, 377487360, 402653184, 452984832, 503316480, 536870912, 603979776, 671088640, 754974720,
805306368, 905969664, 1006632960, 1207959552, 1342177280, 1358954496, 1509949440, 1610612736, 1811939328,
2013265920], dtype=np.int)
def nearestFFTnumber(x):
return np.maximum(x, fastFFT[np.searchsorted(fastFFT, x)])
def convolve1D(x, y, mode, largest_size=0, cache=None, cache_args=[1, 2]):
if min(x.shape[0], y.shape[0]) > 1000:
return convolveFFT(x, y, mode, largest_size=largest_size, cache=cache, cache_args=[1, 2])
else:
return np.convolve(x, y, mode)
def convolve2D(x, y, mode, largest_size=0, cache=None, cache_args=[1, 2]):
return convolveFFTn(x, y, mode, largest_size, cache, cache_args=cache_args)
def convolveFFT(x, y, mode='same', yfft=None, xfft=None, largest_size=0, cache=None, cache_args=[1, 2]):
"""
convolution of x with y; fft cans be cached.
Be careful with caches, key uses id which can be reused for different object if object is freed.
"""
size = x.size + y.size - 1
fsize = nearestFFTnumber(np.maximum(largest_size, size))
if yfft is None:
if cache is not None and 2 in cache_args:
key = (fsize, y.size, id(y))
yfft = cache.get(key)
if yfft is None:
yfft = np.fft.rfft(y, fsize)
if cache is not None and 2 in cache_args: cache[key] = yfft
if xfft is None:
if cache is not None and 1 in cache_args:
key = (fsize, x.size, id(x))
xfft = cache.get(key)
if xfft is None:
xfft = np.fft.rfft(x, fsize)
if cache is not None and 1 in cache_args: cache[key] = xfft
res = np.fft.irfft(xfft * yfft)[0:size]
if mode == 'same':
return res[(y.size - 1) // 2:(y.size - 1) // 2 + x.size]
elif mode == 'full':
return res
elif mode == 'valid':
return res[y.size - 1:x.size]
def convolveFFTn(in1, in2, mode="same", largest_size=0, cache=None, yfft=None, xfft=None, cache_args=[1, 2]):
s1 = np.array(in1.shape)
s2 = np.array(in2.shape)
size = s1 + s2 - 1
fsize = nearestFFTnumber(np.maximum(largest_size, size))
if cache is not None:
if xfft is None and 1 in cache_args:
key = (tuple(fsize), tuple(in1.shape), id(in1))
xfft = cache.get(key)
if yfft is None and 2 in cache_args:
key2 = (tuple(fsize), tuple(in2.shape), id(in2))
yfft = cache.get(key2)
if xfft is None:
xfft = np.fft.rfftn(in1, fsize)
if cache is not None and 1 in cache_args: cache[key] = xfft
if yfft is None:
yfft = np.fft.rfftn(in2, fsize)
if cache is not None and 2 in cache_args: cache[key2] = yfft
fslice = tuple([slice(0, int(sz)) for sz in size])
ret = np.fft.irfftn(xfft * yfft, fsize)[fslice]
if mode == "full":
return ret
elif mode == "same":
return _centered(ret, s1)
elif mode == "valid":
return _centered(ret, s1 - s2 + 1)
def _centered(arr, newsize):
# Return the center newsize portion of the array.
startind = (np.array(arr.shape) - newsize) // 2
endind = startind + newsize
myslice = [slice(startind[k], endind[k]) for k in range(len(endind))]
return arr[tuple(myslice)]
def autoCorrelation(x, n=None, normalized=True, start_index=0):
"""
Calculates auto-correlation of x, result[k] = sum_{i=0}^{n-k} x_i x_{i+k}/(n-k)
if normalized=True, divides by the variance (for if start_index=0, first number is one)
"""
result = autoConvolve(x - x.mean(), n, normalize=True)
if normalized:
result /= result[0]
return result[start_index:]
def autoConvolve(x, n=None, normalize=True):
"""
Calculates auto-covariance of x, result[k] = sum_i x_i x_{i+k}
n is maximum size to return (k = 0..n-1)
if normalize=True then normalize convolution by the number of terms for each k
(can input x-mean(x) and divide result by variance to get auto correlation)
"""
s = nearestFFTnumber(2 * x.size)
# yt = np.fft.rfft(x, s)
# yt *= yt.conj()
# return np.fft.irfft(yt)[0:x.size]
xt = fftpack.rfft(x, s)
auto = np.empty((xt.size // 2) + 1)
auto[0] = xt[0] ** 2
auto[-1] = xt[-1] ** 2
auto[1:-1] = (xt[1:-2:2] ** 2 + xt[2:-1:2] ** 2)
n = n or x.size
res = fftpack.idct(auto, type=1)[0:n] / s
if normalize:
res /= np.arange(x.size, x.size - n, -1)
return res
def convolveGaussianDCT(x, sigma, pad_sigma=4, mode='same', cache={}):
"""
1D convolution of x with Gaussian of width sigma pixels
If pad_sigma>0, pads ends with zeros by int(pad_sigma*sigma) pixels
Otherwise does unpadded fast cosine transform, hence reflection from the ends
"""
fill = int(pad_sigma * sigma)
actual_size = x.size + fill * 2
if fill > 0:
s = nearestFFTnumber(actual_size)
fill2 = s - x.size - fill
padded_x = np.pad(x, (fill, fill2), mode='constant')
else:
padded_x = x
s = padded_x.size
hnorm = sigma / float(s)
gauss = cache.get((s, hnorm))
if gauss is None:
gauss = np.exp(-(np.arange(0, s) * (np.pi * hnorm)) ** 2 / 2.)
cache[(s, hnorm)] = gauss
res = fftpack.idct(fftpack.dct(padded_x, overwrite_x=fill > 0) * gauss, overwrite_x=fill > 0) / (2 * s)
if fill == 0: return res
if mode == 'same':
return res[fill:-fill2]
elif mode == 'valid':
return res[fill * 2:-fill2 - fill]
else:
raise ValueError('mode not supported for convolveGaussianDCT')
def convolveGaussian(x, sigma, sigma_range=4, cache=None):
"""
1D convolution of x with Gaussian of width sigma pixels
x_max = int(sigma_range*sigma) the zero padding range at ends
This uses periodic boundary conditions, and mode = 'same'
This is the fastest fft version
"""
fill = int(sigma_range * sigma)
actual_size = x.size + 2 * fill
if fill > 0:
s = nearestFFTnumber(actual_size)
else:
s = actual_size
gauss = None if cache is None else cache.get((fill, actual_size, sigma))
if gauss is None:
hnorm = sigma / float(s)
ps = np.arange(1, s + 1) // 2
gauss = np.exp(-(ps * (np.pi * hnorm)) ** 2 * 2)
if cache is not None:
cache[(fill, actual_size, sigma)] = gauss
res = fftpack.irfft(fftpack.rfft(x, s) * gauss, s)
return res[:x.size]
def convolveGaussianTrunc(x, sigma, sigma_range=4, mode='same', cache=None):
"""
1D convolution of x with Gaussian of width sigma pixels
x_max = int(sigma_range*sigma) determines the finite support (in pixels) of the truncated gaussian
This uses normalized finite range approximation to Gaussian
"""
fill = int(sigma_range * sigma)
actual_size = x.size + 2 * fill
s = nearestFFTnumber(actual_size)
gauss = None if cache is None else cache.get((fill, actual_size, sigma))
if gauss is None:
points = np.arange(-fill, fill + 1)
Win = np.exp(-(points / sigma) ** 2 / 2.)
Win /= np.sum(Win)
gauss = np.fft.rfft(Win, s)
if cache is not None:
cache[(fill, actual_size, sigma)] = gauss
res = np.fft.irfft(np.fft.rfft(x, s) * gauss, s)[:actual_size]
if mode == 'same':
return res[fill:-fill]
elif mode == 'full':
return res
elif mode == 'valid':
return res[2 * fill:-2 * fill]
def dct2d(a):
return fftpack.dct(fftpack.dct(a, axis=0), axis=1)
def idct2d(a):
return fftpack.idct(fftpack.idct(a, axis=1), axis=0)