def SkipNet18_Combo(c, **kargs):
dl = n.DeepLoss
cmk = n.CorrFix
dec = lambda x: n.DecorrMin(x, num_to_keep = True)
return n.Seq(n.ResNet([2,2,2,2], extra = [ (cmk(20),2),(dec(10),2)
,(cmk(10),3),(dec(5),3),(dl(S.Until(90, S.Lin(0, 0.2, 50, 40), 0)), 3)
,(cmk(5),4),(dec(2),4)], bias = True, ibp_init=True, skip_net = True), n.FFNN([512, 512, c], bias=True, last_lin=True, last_zono = True, ibp_init=True, **kargs))
def ResNetTiny_FewCombo(
c, **kargs
): # resnetWide also used by mixtrain and scaling provable adversarial defenses
def wb(c, bias=True, **kargs):
return n.WideBlock(c,
False,
bias=bias,
ibp_init=True,
batch_norm=False,
**kargs)
dl = n.DeepLoss
cmk = n.CorrMaxK
cm2d = n.CorrMaxPool2D
cm3d = n.CorrMaxPool3D
dec = lambda x: n.DecorrMin(x, num_to_keep=True)
return n.Seq(
cmk(32), n.Conv(16, 3, padding=1, bias=True, ibp_init=True), dec(8),
wb(16), dec(4), wb(32), n.Concretize(), wb(32), wb(32), wb(32),
cmk(10),
n.FFNN([500, c],
bias=True,
last_lin=True,
ibp_init=True,
last_zono=True,
**kargs))
def ResNet34(c, **kargs):
return n.Seq(
n.ResNet([3, 4, 6, 3], bias=True, ibp_init=True),
n.FFNN([512, 512, c],
bias=True,
last_lin=True,
last_zono=True,
ibp_init=True,
**kargs))
def SkipNet18(c, **kargs):
return n.Seq(
n.ResNet([2, 2, 2, 2], bias=True, ibp_init=True, skip_net=True),
n.FFNN([512, 512, c],
bias=True,
last_lin=True,
last_zono=True,
ibp_init=True,
**kargs))
def ResNetTiny(c, **kargs): # resnetWide also used by mixtrain and scaling provable adversarial defenses
def wb(c, bias = True, **kargs):
return n.WideBlock(c, False, bias=bias, ibp_init=True, batch_norm = False, **kargs)
return n.Seq(n.Conv(16, 3, padding=1, bias=True, ibp_init = True),
wb(16),
wb(32),
wb(32),
wb(32),
wb(32),
n.FFNN([500, c], bias=True, last_lin=True, ibp_init = True, last_zono = True, **kargs))
def ResNetWong(c, **kargs):
return n.Seq(
n.Conv(16, 3, padding=1, bias=False), n.WideBlock(16), n.WideBlock(16),
n.WideBlock(32, True), n.WideBlock(64, True),
n.FFNN([1000, c],
ibp_init=True,
bias=True,
last_lin=True,
last_zono=True,
**kargs))
def ResNetLarge_LargeCombo(c, **kargs): # resnetWide also used by mixtrain and scaling provable adversarial defenses
def wb(c, bias = True, **kargs):
return n.WideBlock(c, False, bias=bias, ibp_init=True, batch_norm = False, **kargs)
dl = n.DeepLoss
cmk = n.CorrMaxK
cm2d = n.CorrMaxPool2D
cm3d = n.CorrMaxPool3D
dec = lambda x: n.DecorrMin(x, num_to_keep = True)
return n.Seq(n.Conv(16, 3, padding=1, bias=True, ibp_init = True), cmk(4),
wb(16), cmk(4), dec(4),
wb(32), cmk(4), dec(4),
wb(32), dl(S.Until(1, 0, S.Lin(0.5, 0, 50, 3))),
wb(32), cmk(4), dec(4),
wb(64), cmk(4), dec(2),
wb(64), dl(S.Until(24, S.Lin(0, 0.1, 20, 4), S.Lin(0.1, 0, 50))),
wb(64),
n.FFNN([1000, c], bias=True, last_lin=True, ibp_init = True, **kargs))
def SemanticGenerator(latent_dim, bias=False, normal=False):
conv_layers = [(latent_dim * 2, 5, 5, 2, 'ReLU'), (1, 4, 4, 2, 'tanh')]
def transfer(tp):
if isinstance(tp, InferModule):
return tp
if isinstance(tp[0], str):
return MaxPool2D(*tp[1:])
return Seq(
ConvTranspose2D(out_channels=tp[0],
kernel_size=tp[1],
stride=tp[-2] if len(tp) == 5 else 1,
bias=bias,
normal=normal,
padding=0), activation(tp[-1]))
conv = [transfer(s) for s in conv_layers]
return n.Seq(
#input_shape=(Z_DIM,)
n.FFNN([latent_dim * 7 * 7 * 4], batch_norm=True, bias=bias),
Unflatten2d((7, 7, LATENT_DIM * 4)),
*conv)
def resnet34(c, **kargs):
return n.Seq(n.ResNet([3, 4, 6, 3]),
n.FFNN([512, 512, c], bias=False, last_lin=True, **kargs))
def resnet18(c, **kargs):
return n.Seq(n.ResNet([2, 2, 2, 2]),
n.FFNN([512, 512, c], bias=False, last_lin=True, **kargs))
def resnet18small(c, **kargs):
return n.Seq(n.ResNet([2, 2, 2]),
n.FFNN([100, c], bias=True, last_lin=False, **kargs))
def CharLevelAGSub(c, **kargs):
return n.Seq(n.EmbeddingWithSub(64, 64, 3), n.Conv(64, 10, bias=True), n.AvgPool2D(10), n.ReduceToZono(), n.FFNN([64, 64, c], last_lin=True, last_zono=True, **kargs)) #
def WordLevelSST2(c, **kargs):
return n.Seq(n.Embedding(1, 300), n.Conv(100, 5, bias=True), n.AvgPool2D(5), n.ReduceToZono(), n.FFNN([c], last_lin=True, last_zono=True, **kargs))
def MyLenet():
return n.Seq(n.Flatten((28, 28)), n.PrintActivation(), n.Linear(784, 128),
n.activation('ReLU'), n.Dropout(128, p=0.2),
n.Linear(128, 10), n.activation('softmax'))
