-
Notifications
You must be signed in to change notification settings - Fork 589
/
kld.py
210 lines (182 loc) 路 8.56 KB
/
kld.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
# -*- coding: utf-8 -*-
import logging as logging
import numpy as np
import math
import functools
import vaex.vaexfast
import vaex.utils
from vaex.utils import Timer
logger = logging.getLogger("vaex.kld")
def mutual_information(data):
Q = vaex.utils.disjoined(data)
P = data
P = P / P.sum()
Q = Q / Q.sum()
mask = (P > 0) & (Q > 0)
information = np.sum(P[mask] * np.log(P[mask] / Q[mask])) # * np.sum(dx)
return information
def kl_divergence(P, Q, axis=None):
P = P / P.sum(axis=axis)
Q = Q / Q.sum(axis=axis)
# mask = (P > 0) & (Q>0)
# information = np.sum(P[mask] * np.log(P[mask]/Q[mask]), axis=axis)# * np.sum(dx)
# mask = (P > 0) & (Q>0)
information = np.sum(P * np.log(P / Q), axis=axis) # * np.sum(dx)
return information
class KlDivergenceShuffle(object):
def __init__(self, dataset, pairs, gridsize=128):
self.dataset = dataset
self.pairs = pairs
self.dimension = len(self.pairs[0])
self.logger = logger.getLogger('kld')
self.gridsize = gridsize
logger.debug("dimension: %d, pairs: %s" % (self.dimension, self.pairs))
def get_jobs(self):
def job(pair):
pass
def to_disjoined(counts):
shape = counts.shape
assert len(counts.shape) == 2
counts_0 = counts.sum(axis=1).reshape((shape[0], 1))
counts_1 = counts.sum(axis=0).reshape((1, shape[1]))
counts_disjoined = counts_0 * counts_1
return counts_disjoined
def kld_shuffled(columns, Ngrid=128, datamins=None, datamaxes=None, offset=1):
if datamins is None:
datamins = np.array([np.nanmin(column) for column in columns])
if datamaxes is None:
datamaxes = np.array([np.nanmax(column) for column in columns])
dim = len(columns)
counts = np.zeros((Ngrid, ) * dim, dtype=np.float64)
counts_shuffled = np.zeros((Ngrid, ) * dim, dtype=np.float64)
D_kl = -1
if len(columns) == 2:
x, y = columns
# print x
# print y
print((x, y, counts, counts_shuffled, datamins[0], datamaxes[0], datamins[1], datamaxes[1], offset))
try:
vaex.histogram.hist2d_and_shuffled(x, y, counts, counts_shuffled, datamins[0], datamaxes[0], datamins[1], datamaxes[1], offset)
except:
args = [x, y, counts, counts_shuffled, datamins[0], datamaxes[0], datamins[1], datamaxes[1], offset]
sig = [numba.dispatcher.typeof_pyval(a) for a in args]
print(sig)
raise
print(("counts", sum(counts)))
deltax = [float(datamaxes[i] - datamins[i]) for i in range(dim)]
dx = np.array([deltax[d] / counts.shape[d] for d in range(dim)])
density = counts / np.sum(counts) # * np.sum(dx)
density_shuffled = counts_shuffled / np.sum(counts_shuffled) # * np.sum(dx)
mask = (density_shuffled > 0) & (density > 0)
# print density
D_kl = np.sum(density[mask] * np.log(density[mask] / density_shuffled[mask])) # * np.sum(dx)
# if D_kl < 0:
# import pdb
# pdb.set_trace()
return D_kl
def kld_shuffled_grouped(dataset, range_map, pairs, feedback=None, size_grid=32, use_mask=True, bytes_max=int(1024**3 / 2)):
dimension = len(pairs[0])
bytes_per_grid = size_grid ** dimension * 8 # 8==sizeof (double)
grids_per_iteration = min(128, int(bytes_max / bytes_per_grid))
iterations = int(math.ceil(len(pairs) * 1. / grids_per_iteration))
jobsManager = vaex.dataset.JobsManager()
ranges = [None] * len(pairs)
D_kls = []
class Wrapper(object):
pass
wrapper = Wrapper()
wrapper.N_done = 0
counts = np.zeros((grids_per_iteration,) + (size_grid,) * dimension, dtype=np.float64)
# counts_shuffled = np.zeros((grids_per_iteration,) + (size_grid,) * dimension, dtype=np.float64)
N_total = len(pairs) * len(dataset) * dimension + len(pairs) * counts.size
logger.debug("{iterations} iterations with {grids_per_iteration} grids per iteration".format(**locals()))
for part in range(iterations):
if part > 0: # next iterations reset the counts
counts.reshape(-1)[:] = 0
# counts_shuffled.reshape(-1)[:] = 0
i1, i2 = part * grids_per_iteration, (part + 1) * grids_per_iteration
if i2 > len(pairs):
i2 = len(pairs)
n_grids = i2 - i1
logger.debug("part {part} of {iterations}, from {i1} to {i2}".format(**locals()))
def grid(info, *blocks, **kwargs):
index = kwargs["index"]
if use_mask and dataset.mask is not None:
mask = dataset.mask[info.i1:info.i2]
blocks = [block[mask] for block in blocks]
else:
mask = None
ranges = []
minima = []
maxima = []
for dim in range(dimension):
ranges += list(range_map[pairs[index][dim]])
mi, ma = range_map[pairs[index][dim]]
minima.append(mi)
maxima.append(ma)
if len(blocks) == 2:
print(("mask", mask))
vaex.vaexfast.histogram2d(blocks[0], blocks[1], None, counts[index], *(ranges + [0, 0]))
# vaex.vaexfast.histogram2d(blocks[0], blocks[1], None, counts_shuffled[index], *(ranges + [1, 0]))
if len(blocks) == 3:
vaex.vaexfast.histogram3d(blocks[0], blocks[1], blocks[2], None, counts[index], *(ranges + [0, 0, 0]))
# vaex.vaexfast.histogramNd(list(blocks), None, counts[index], minima, maxima)
# for i in range(5):
# vaex.vaexfast.histogram3d(blocks[0], blocks[1], blocks[2], None, counts_shuffled[index], *(ranges + [2+i,1+i,0]))
if len(blocks) > 3:
vaex.vaexfast.histogramNd(list(blocks), None, counts[index], minima, maxima)
# for i in range(5):
# vaex.vaexfast.histogram3d(blocks[0], blocks[1], blocks[2], None, counts_shuffled[index], *(ranges + [2+i,1+i,0]))
if feedback:
wrapper.N_done += len(blocks[0]) * dimension
if feedback:
cancel = feedback(wrapper.N_done * 100. / N_total)
if cancel:
raise Exception("cancelled")
for index, pair in zip(list(range(i1, i2)), pairs[i1:i2]):
logger.debug("add job %r %r" % (index, pair))
jobsManager.addJob(0, functools.partial(grid, index=index - i1), dataset, *pair)
jobsManager.execute()
with Timer("D_kl"):
for i in range(n_grids):
if 0:
deltax = [float(range_map[pairs[i][d]][1] - range_map[pairs[i][d]][0]) for d in range(dimension)]
dx = np.array([deltax[d] / counts[i].shape[d] for d in range(dimension)])
density = counts[i] / np.sum(counts[i]) # * np.sum(dx)
counts_shuffled = to_disjoined(counts[i])
density_shuffled = counts_shuffled / np.sum(counts_shuffled) # * np.sum(dx)
mask = (density_shuffled > 0) & (density > 0)
print(("mask sum", np.sum(mask)))
print(("mask sum", np.sum((counts_shuffled > 0) & (counts[i] > 0))))
# print density
D_kl = np.sum(density[mask] * np.log(density[mask] / density_shuffled[mask])) # * np.sum(dx)
else:
D_kl = mutual_information(counts[i])
D_kls.append(D_kl)
if feedback:
wrapper.N_done += counts.size
cancel = feedback(wrapper.N_done * 100. / N_total)
if cancel:
return None
# raise Exception, "cancelled"
return D_kls
if __name__ == "__main__":
import vaex.dataset
import vaex.files
from optparse import OptionParser
parser = OptionParser() # usage="")
# parser.add_option("-n", "--name",
# help="dataset name [default=%default]", default="data", type=str)
# parser.add_option("-o", "--output",
# help="dataset output filename [by default the suffix of input filename will be replaced by hdf5]", default=None, type=str)
(options, args) = parser.parse_args()
if len(args) > 0:
filename = args[0]
else:
filename = "gaussian4d-1e7.hdf5"
path = vaex.files.get_datafile(filename)
dataset = vaex.dataset.Hdf5MemoryMapped(path)
# for column_name in dataset.column_names:
# print column_name
subspace = dataset.subspace(1)[:3] * dataset.subspace(1)[:3]
# subspace = Subspace(dataset, [("x",), ("y",), ("z",)]