def __init__(self, **kwargs):
self._numhid = kwargs.pop('numhid', 64)
self._outnum = kwargs.pop('outnum', 10)
self._outact = kwargs.pop('outact', None)
self._patest = kwargs.pop('patest', 'linear')
self._outest = kwargs.pop('outest', self._patest)
super().__init__(**kwargs)
with self.name_scope():
# _numhid x 28 x 28
self.add(
Conv2D(self._numhid,
4,
strides=2,
padding=1,
use_bias=False,
regimes=estimators[self._patest](),
isinput=True))
self.add(ReLU(regimes=estimators[self._patest]()))
# _numhid x 14 x 14
self.add(
Conv2D(self._numhid * 2,
4,
strides=2,
padding=1,
use_bias=False,
regimes=estimators[self._patest]()))
self.add(ReLU(regimes=estimators[self._patest]()))
# _numhid x 7 x 7
self.add(
Conv2D(self._numhid * 4,
4,
strides=1,
padding=0,
use_bias=False,
regimes=estimators[self._patest]()))
self.add(ReLU(regimes=estimators[self._patest]()))
# _numhid x 4 x 4
self.add(
Conv2D(self._numhid * 8,
4,
strides=1,
padding=0,
use_bias=False,
regimes=estimators[self._patest]()))
self.add(ReLU(regimes=estimators[self._patest]()))
# filters x 1 x 1
self.add(Dense(self._outnum, regimes=estimators[self._outest]()))
if self._outact == 'relu':
self.add(ReLU(regimes=estimators[self._outest]()))
else:
self.add(Identity(regimes=estimators[self._outest]()))
def __init__(self, **kwargs):
self._numhid = kwargs.pop('numhid', 64)
self._outnum = kwargs.pop('outnum', 10)
self._patest = kwargs.pop('patest', 'linear')
self._outest = kwargs.pop('outest', self._patest)
super().__init__(**kwargs)
with self.name_scope():
# 3 x 28 x 28
self.add(
Conv2D(self._numhid,
5,
strides=1,
padding=0,
use_bias=True,
regimes=estimators[self._patest]()))
self.add(ReLU(regimes=estimators[self._patest]()))
# filt x 22 x 22
self.add(MaxPool2D(pool_size=2, strides=2))
# filt x 11 x 11
self.add(
Conv2D(self._numhid * 2,
4,
strides=1,
padding=0,
use_bias=True,
regimes=estimators[self._patest]()))
# filt x 8 x 8
self.add(ReLU(regimes=estimators[self._patest]()))
self.add(MaxPool2D(pool_size=2, strides=2))
# filt x 4 x 4
self.add(
Dense(self._numhid * 2, regimes=estimators[self._patest]()))
self.add(Dense(self._outnum, regimes=estimators[self._outest]()))
self.add(Identity(regimes=estimators[self._outest]()))
def __init__(self, **kwargs):
outnum = kwargs.pop('outnum', 1)
outact = kwargs.pop('outact', None)
numhid = kwargs.pop('numhid', 512)
droprate = kwargs.pop('droprate', 0.25)
use_bias = kwargs.pop('use_bias', False)
# patest = dict(relu='relu', out='clip', pixel='relu', gauss='relu')
patest = dict(relu='linear',
out='linear',
pixel='linear',
gauss='linear')
patest.update(kwargs.pop('patest', {}))
explain = dict(relu='zplus', out='zplus', pixel='zb', gauss='wsquare')
explain.update(kwargs.pop('explain', {}))
super().__init__(**kwargs)
with self.name_scope():
# 28 x 28
self.add(
Conv2D(128,
3,
strides=1,
padding=1,
use_bias=use_bias,
explain=explain['pixel'],
regimes=estimators[patest['pixel']]()))
self.add(ReLU())
self.add(MaxPool2D(pool_size=2, strides=2))
# 14 x 14
self.add(
Conv2D(128,
3,
strides=1,
padding=1,
use_bias=use_bias,
explain=explain['pixel'],
regimes=estimators[patest['pixel']]()))
self.add(ReLU())
self.add(MaxPool2D(pool_size=2, strides=2))
# 7 x 7
self.add(
Conv2D(128,
3,
strides=1,
padding=1,
use_bias=use_bias,
explain=explain['pixel'],
regimes=estimators[patest['pixel']]()))
self.add(ReLU())
self.add(MaxPool2D(pool_size=2, strides=2))
# 3 x 3
self.add(
Conv2D(128,
3,
strides=1,
padding=1,
use_bias=use_bias,
explain=explain['pixel'],
regimes=estimators[patest['pixel']]()))
self.add(ReLU())
self.add(MaxPool2D(pool_size=2, strides=2))
# 2 x 2
self.add(Flatten())
self.add(
Dense(numhid,
explain=explain['relu'],
regimes=estimators[patest['relu']]()))
self.add(ReLU())
self.add(Dropout(droprate))
self.add(
Dense(outnum,
explain=explain['relu'],
regimes=estimators[patest['relu']]()))
if outact == 'relu':
self.add(ReLU())
else:
self.add(Identity())
def __init__(self, **kwargs):
outnum = kwargs.pop('outnum', 1)
outact = kwargs.pop('outact', None)
numhid = kwargs.pop('numhid', 64)
leakage = kwargs.pop('leakage', 0.1)
use_bias = kwargs.pop('use_bias', False)
# patest = dict(relu='relu', out='clip', pixel='relu', gauss='relu')
patest = dict(relu='linear',
out='linear',
pixel='linear',
gauss='linear')
patest.update(kwargs.pop('patest', {}))
explain = dict(relu='zplus', out='zplus', pixel='zb', gauss='wsquare')
explain.update(kwargs.pop('explain', {}))
super().__init__(**kwargs)
with self.name_scope():
# _numhid x 32 x 32
self.add(
Conv2D(numhid,
4,
strides=2,
padding=1,
use_bias=use_bias,
explain=explain['pixel'],
regimes=estimators[patest['pixel']]()))
self.add(LeakyReLU(leakage))
# _numhid x 16 x 14
self.add(
Conv2D(numhid * 2,
4,
strides=2,
padding=1,
use_bias=use_bias,
explain=explain['relu'],
regimes=estimators[patest['relu']]()))
self.add(BatchNorm())
self.add(LeakyReLU(leakage))
# _numhid x 8 x 8
self.add(
Conv2D(numhid * 4,
4,
strides=2,
padding=1,
use_bias=use_bias,
explain=explain['relu'],
regimes=estimators[patest['relu']]()))
self.add(BatchNorm())
self.add(LeakyReLU(leakage))
# _numhid x 4 x 4
self.add(
Conv2D(outnum,
4,
strides=1,
padding=0,
use_bias=use_bias,
explain=explain['out'],
regimes=estimators[patest['out']]()))
self.add(Flatten())
# self.add(BatchNorm())
# self.add(LeakyReLU(leakage))
# # filters x 1 x 1
# self.add(Dense(outnum,
# explain=explain['relu'], regimes=estimators[patest['relu']]()))
if outact == 'relu':
self.add(ReLU())
else:
self.add(Identity())
#patterntest
import numpy as np
from mxnet import nd
from ecGAN.layer import Conv2D
from ecGAN.explain.pattern.estimator import estimators
lay = Conv2D(20, 2, strides=2, padding=0, regimes=estimators['linear']())
lay.initialize()
data = nd.random.normal(5, shape=[1000, 3, 8, 8])
out = lay(data)
lay.init_pattern()
lay.collect_pparams().initialize()
for mdat in [data[i::100] for i in range(100)]:
lay.forward_logged(mdat)
lay.learn_pattern()
lay.compute_pattern()
resdat = data.reshape([1000, 3, 4, 2, 4,
2]).transpose([0, 2, 4, 1, 3,
5]).reshape([1000 * 4 * 4, 3 * 4])
resout = out.transpose([0, 2, 3, 1]).reshape([1000 * 4 * 4, 20])
rescov = nd.dot((resout - resout.mean(0)).T,
(resdat - resdat.mean(0))) / resout.shape[0]
#TODO check whether correlation is correct!
var_y = (lay.weight.data().flatten() * rescov).mean(1, keepdims=True)
std_y = (resout - resout.mean(0)).mean(0)