Spaces:
Runtime error
Runtime error
File size: 9,590 Bytes
2fc2c1f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 |
#!/usr/local/bin/python3
# avenir-python: Machine Learning
# Author: Pranab Ghosh
#
# Licensed under the Apache License, Version 2.0 (the "License"); you
# may not use this file except in compliance with the License. You may
# obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
# implied. See the License for the specific language governing
# permissions and limitations under the License.
# Package imports
import os
import sys
import matplotlib.pyplot as plt
import numpy as np
import sklearn as sk
from sklearn.neighbors import KDTree
import matplotlib
import random
import jprops
from random import randint
import statistics
sys.path.append(os.path.abspath("../lib"))
from util import *
from mlutil import *
from tnn import *
from stats import *
"""
neural model calibration
"""
class ModelCalibration(object):
def __init__(self):
pass
@staticmethod
def findModelCalibration(model):
"""
pmodel calibration
"""
FeedForwardNetwork.prepValidate(model)
FeedForwardNetwork.validateModel(model)
yPred = model.yPred.flatten()
yActual = model.validOutData.flatten()
nsamp = len(yActual)
#print(yPred.shape)
#print(yActual.shape)
nBins = model.config.getIntConfig("calibrate.num.bins")[0]
prThreshhold = model.config.getFloatConfig("calibrate.pred.prob.thresh")[0]
minConf = yPred.min()
maxConf = yPred.max()
bsize = (maxConf - minConf) / nBins
#print("minConf {:.3f} maxConf {:.3f} bsize {:.3f}".format(minConf, maxConf, bsize))
blist = list(map(lambda i : None, range(nBins)))
#binning
for yp, ya in zip(yPred, yActual):
indx = int((yp - minConf) / bsize)
if indx == nBins:
indx = nBins - 1
#print("yp {:.3f} indx {}".format(yp, indx))
pair = (yp, ya)
plist = blist[indx]
if plist is None:
plist = list()
blist[indx] = plist
plist.append(pair)
x = list()
y = list()
yideal = list()
ece = 0
mce = 0
# per bin confidence and accuracy
b = 0
for plist in blist:
if plist is not None:
#confidence
ypl = list(map(lambda p : p[0], plist))
ypm = statistics.mean(ypl)
x.append(ypm)
#accuracy
ypcount = 0
for p in plist:
yp = 1 if p[0] > prThreshhold else 0
if (yp == 1 and p[1] == 1):
ypcount += 1
acc = ypcount / len(plist)
y.append(acc)
yideal.append(ypm)
ce = abs(ypm - acc)
ece += len(plist) * ce
if ce > mce:
mce = ce
else:
ypm = minConf + (b + 0.5) * bsize
x.append(ypm)
yideal.append(ypm)
y.append(0)
b += 1
#calibration plot
drawPairPlot(x, y, yideal, "confidence", "accuracy", "actual", "ideal")
print("confidence\taccuracy")
for z in zip(x,y):
print("{:.3f}\t{:.3f}".format(z[0], z[1]))
#expected calibration error
ece /= nsamp
print("expected calibration error\t{:.3f}".format(ece))
print("maximum calibration error\t{:.3f}".format(mce))
@staticmethod
def findModelCalibrationLocal(model):
"""
pmodel calibration based k nearest neghbors
"""
FeedForwardNetwork.prepValidate(model)
FeedForwardNetwork.validateModel(model)
yPred = model.yPred.flatten()
yActual = model.validOutData.flatten()
nsamp = len(yActual)
neighborCnt = model.config.getIntConfig("calibrate.num.nearest.neighbors")[0]
prThreshhold = model.config.getFloatConfig("calibrate.pred.prob.thresh")[0]
fData = model.validFeatData.numpy()
tree = KDTree(fData, leaf_size=4)
dist, ind = tree.query(fData, k=neighborCnt)
calibs = list()
#all data
for si, ni in enumerate(ind):
conf = 0
ypcount = 0
#all neighbors
for i in ni:
conf += yPred[i]
yp = 1 if yPred[i] > prThreshhold else 0
if (yp == 1 and yActual[i] == 1):
ypcount += 1
conf /= neighborCnt
acc = ypcount / neighborCnt
calib = (si, conf, acc)
calibs.append(calib)
#descending sort by difference between confidence and accuracy
calibs = sorted(calibs, key=lambda c : abs(c[1] - c[2]), reverse=True)
print("local calibration")
print("conf\taccu\trecord")
for i in range(19):
si, conf, acc = calibs[i]
rec = toStrFromList(fData[si], 3)
print("{:.3f}\t{:.3f}\t{}".format(conf, acc, rec))
@staticmethod
def findModelSharpness(model):
"""
pmodel calibration
"""
FeedForwardNetwork.prepValidate(model)
FeedForwardNetwork.validateModel(model)
yPred = model.yPred.flatten()
yActual = model.validOutData.flatten()
nsamp = len(yActual)
#print(yPred.shape)
#print(yActual.shape)
nBins = model.config.getIntConfig("calibrate.num.bins")[0]
prThreshhold = model.config.getFloatConfig("calibrate.pred.prob.thresh")[0]
minConf = yPred.min()
maxConf = yPred.max()
bsize = (maxConf - minConf) / nBins
#print("minConf {:.3f} maxConf {:.3f} bsize {:.3f}".format(minConf, maxConf, bsize))
blist = list(map(lambda i : None, range(nBins)))
#binning
for yp, ya in zip(yPred, yActual):
indx = int((yp - minConf) / bsize)
if indx == nBins:
indx = nBins - 1
#print("yp {:.3f} indx {}".format(yp, indx))
pair = (yp, ya)
plist = blist[indx]
if plist is None:
plist = list()
blist[indx] = plist
plist.append(pair)
y = list()
ypgcount = 0
# per bin confidence and accuracy
for plist in blist:
#ypl = list(map(lambda p : p[0], plist))
#ypm = statistics.mean(ypl)
#x.append(ypm)
ypcount = 0
for p in plist:
yp = 1 if p[0] > prThreshhold else 0
if (yp == 1 and p[1] == 1):
ypcount += 1
ypgcount += 1
acc = ypcount / len(plist)
y.append(acc)
print("{} {}".format(ypgcount, nsamp))
accg = ypgcount / nsamp
accgl = [accg] * nBins
x = list(range(nBins))
drawPairPlot(x, y, accgl, "discretized confidence", "accuracy", "local", "global")
contrast = list(map(lambda acc : abs(acc - accg), y))
contrast = statistics.mean(contrast)
print("contrast {:.3f}".format(contrast))
"""
neural model robustness
"""
class ModelRobustness(object):
def __init__(self):
pass
def localPerformance(self, model, fpath, nsamp, neighborCnt):
"""
local performnance sampling
"""
#load data
fData, oData = FeedForwardNetwork.prepData(model, fpath)
#print(type(fData))
#print(type(oData))
#print(fData.shape)
dsize = fData.shape[0]
ncol = fData.shape[1]
#kdd
tree = KDTree(fData, leaf_size=4)
scores = list()
indices = list()
for _ in range(nsamp):
indx = randomInt(0, dsize - 1)
indices.append(indx)
frow = fData[indx]
frow = np.reshape(frow, (1, ncol))
dist, ind = tree.query(frow, k=neighborCnt)
ind = ind[0]
vfData = fData[ind]
voData = oData[ind]
#print(type(vfData))
#print(vfData.shape)
#print(type(voData))
#print(voData.shape)
model.setValidationData((vfData, voData), False)
score = FeedForwardNetwork.validateModel(model)
scores.append(score)
#performance distribution
m, s = basicStat(scores)
print("model performance: mean {:.3f}\tstd dev {:.3f}".format(m,s))
drawHist(scores, "model accuracy", "accuracy", "frequency")
#worst performance
lscores = sorted(zip(indices, scores), key=lambda s : s[1])
print(lscores[:5])
lines = getFileLines(fpath, None)
print("worst performing features regions")
for i,s in lscores[:5]:
print("score {:.3f}\t{}".format(s, lines[i]))
"""
conformal prediction for regression
"""
class ConformalRegressionPrediction(object):
def __init__(self):
self.calibration = dict()
def calibrate(self, ypair, confBound):
""" n
calibration for conformal prediction
"""
cscores = list()
ymax = None
ymin = None
for yp, ya in ypair:
cscore = abs(yp - ya)
cscores.append(cscore)
if ymax is None:
ymax = ya
ymin = ya
else:
ymax = ya if ya > ymax else ymax
ymin = ya if ya < ymin else ymin
cscores.sort()
drawHist(cscores, "conformal score distribution", "conformal score", "frequency", 20)
cbi = int(confBound * len(cscores))
scoreConfBound = cscores[cbi]
self.calibration["scoreConfBound"] = scoreConfBound
self.calibration["ymin"] = ymin
self.calibration["ymax"] = ymax
print(self.calibration)
def saveCalib(self, fPath):
"""
saves scoformal score calibration
"""
saveObject(self.calibration, fPath)
def restoreCalib(self, fPath):
"""
saves scoformal score calibration
"""
self.calibration = restoreObject(fPath)
print(self.calibration)
def getPredRange(self, yp, nstep=100):
"""
get prediction range and related data
"""
ymin = self.calibration["ymin"]
ymax = self.calibration["ymax"]
step = (ymax - ymin) / nstep
scoreConfBound = self.calibration["scoreConfBound"]
rmin = None
rmax = None
rcount = 0
#print(ymin, ymax, step)
for ya in np.arange(ymin, ymax, step):
cscore = abs(yp - ya)
if cscore < scoreConfBound:
if rmin is None:
#lower bound
rmin = ya
rmax = ya
else:
#keep updating upper bound
rmax = ya if ya > rmax else rmax
rcount += 1
else:
if rmax is not None and rcount > 0:
#past upper bound
break
res = dict()
res["predRangeMin"] = rmin
res["predRangeMax"] = rmax
accepted = yp >= rmin and yp <= rmax
res["status"] = "accepted" if accepted else "rejected"
conf = 1.0 - (rmax - rmin) / (ymax - ymin)
res["confidence"] = conf
return res
|