def __init__(self, nc, n, ps=0.5):
super().__init__()
layers = [AdaptiveConcatPool2d(), Mish(), Flatten()] + \
bn_drop_lin(nc*2, 512, True, ps, Mish()) + \
bn_drop_lin(512, n, True, ps)
self.fc = nn.Sequential(*layers)
self._init_weight()
def build_model():
model = Sequential()
model.add(Dense(96, input_dim=state_size, kernel_initializer='he_uniform'))
model.add(Mish())
model.add(Dense(48, kernel_initializer='he_uniform'))
model.add(Mish())
model.add(Dense(24, kernel_initializer='he_uniform'))
model.add(Mish())
model.add(Dense(action_size, kernel_initializer='he_uniform'))
model.compile(Adam(lr=0.001), loss='mse')
return model
def __init__(self,
features=[10, 256, 256],
bn=False,
activation_fn='mish',
p=0):
super().__init__()
self.net = nn.Sequential()
self.activation_fn = Mish()
self.features = features
self.bn = bn
self.activation_fn_name = activation_fn
self.p = p
for layer in range(1, len(features)):
self.net.add_module(
'fc%d' % layer, nn.Linear(features[layer - 1],
features[layer]))
self.net.add_module('sig%d' % layer, self.activation_fn)
if p > 0:
self.net.add_module('dp%d' % layer, nn.Dropout(p))
self.net.add_module('out', nn.Linear(features[-1], 1))
# self.net.add_module('out-nolinear', nn.Softplus(beta=5))
for m in self.modules():
if isinstance(m, nn.Linear):
m.weight.data = nn.init.xavier_normal_(m.weight.data)
m.bias.data = torch.nn.init.zeros_(m.bias.data)
def test_function(test_input):
x1, x2 = (test_input.clone().requires_grad_() for i in range(2))
m1 = Mish()
y1 = m1(x1)
l1 = y1.mean()
exp, = torch.autograd.grad(l1, x1)
y2 = MishImplementation.apply(x2)
l2 = y2.mean()
res, = torch.autograd.grad(l2, x2)
assert_allclose(res, exp)
def __init__(self, vgg_name):
super(VGG, self).__init__()
self.features = self._make_layers(cfg[vgg_name])
self.avgpool = nn.AdaptiveAvgPool2d((7, 7))
self.classifier = nn.Sequential(
nn.Linear(512 * 7 * 7, 4096),
act_fn,
nn.Dropout(),
nn.Linear(4096, 4096),
Mish(),
nn.Dropout(),
nn.Linear(4096, 2), )
def test_module(test_input):
x1, x2 = (test_input.clone().requires_grad_() for i in range(2))
m1 = Mish()
y1 = m1(x1)
l1 = y1.mean()
exp, = torch.autograd.grad(l1, x1)
m2 = MemoryEfficientMish()
y2 = m2(x2)
l2 = y2.mean()
res, = torch.autograd.grad(l2, x2)
assert_allclose(y1, y2)
assert_allclose(res, exp)
def __init__(self,
inplanes,
planes,
stride=1,
downsample=None,
groups=1,
norm_layer=None):
super(BasicBlock, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self.conv1 = conv3x3(inplanes, planes, stride)
self.bn1 = norm_layer(planes)
self.relu = Mish()
self.conv2 = conv3x3(planes, planes)
self.bn2 = norm_layer(planes)
self.downsample = downsample
self.stride = stride
def __init__(self,
inplanes,
planes,
stride=1,
downsample=None,
groups=1,
norm_layer=None,
First=False):
super(Bottleneck, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self.first = First
if self.first:
self.ocb1 = FirstOctaveCBR(inplanes,
planes,
kernel_size=(1, 1),
norm_layer=norm_layer,
padding=0)
else:
self.ocb1 = OctaveCBR(inplanes,
planes,
kernel_size=(1, 1),
norm_layer=norm_layer,
padding=0)
self.ocb2 = OctaveCBR(planes,
planes,
kernel_size=(3, 3),
stride=stride,
groups=groups,
norm_layer=norm_layer)
self.ocb3 = OctaveCB(planes,
planes,
kernel_size=(1, 1),
norm_layer=norm_layer,
padding=0)
self.relu = Mish()
self.downsample = downsample
self.stride = stride
def __init__(self):
super().__init__()
act = lambda: Mish()
self.model = Sequential(
Conv2d(in_channels=3, out_channels=96,
kernel_size=(11, 11), stride=(4, 4), padding=2),
LocalResponseNorm(size=5, k=2, alpha=1e-4, beta=0.75),
MaxPool2d(kernel_size=(3, 3), stride=2),
act(),
Conv2d(in_channels=96, out_channels=256,
kernel_size=(5, 5), padding=2),
LocalResponseNorm(size=5, k=2, alpha=1e-4, beta=0.75),
MaxPool2d(kernel_size=(3, 3), stride=2),
act(),
Conv2d(in_channels=256, out_channels=384,
kernel_size=(3, 3), padding=1),
act(),
Conv2d(in_channels=384, out_channels=384,
kernel_size=(3, 3), padding=1),
act(),
Conv2d(in_channels=384, out_channels=256,
kernel_size=(3, 3), padding=1),
LocalResponseNorm(size=5, k=2, alpha=1e-4, beta=0.75),
MaxPool2d(kernel_size=(3, 3), stride=2),
act(),
Flatten(),
Linear(in_features=9216, out_features=4096),
act(),
Linear(in_features=4096, out_features=4096),
act(),
Linear(in_features=4096, out_features=10))
def __init__(self, arch, num_classes=1, pretrained=True):
super().__init__()
# load EfficientNet
if arch == 'se_resnext50_32x4d':
if pretrained:
self.base = se_resnext50_32x4d()
else:
self.base = se_resnext50_32x4d(pretrained=None)
self.nc = self.base.last_linear.in_features
elif arch == 'se_resnext101_32x4d':
if pretrained:
self.base = se_resnext101_32x4d()
else:
self.base = se_resnext101_32x4d(pretrained=None)
self.nc = self.base.last_linear.in_features
elif arch == 'inceptionv4':
if pretrained:
self.base = inceptionv4()
else:
self.base = inceptionv4(pretrained=None)
self.nc = self.base.last_linear.in_features
elif arch == 'inceptionresnetv2':
if pretrained:
self.base = inceptionresnetv2()
else:
self.base = inceptionresnetv2(pretrained=None)
self.nc = self.base.last_linear.in_features
self.logit = nn.Sequential(AdaptiveConcatPool2d(1), Flatten(),
nn.BatchNorm1d(2 * self.nc),
nn.Dropout(0.5),
nn.Linear(2 * self.nc, 512), Mish(),
nn.BatchNorm1d(512), nn.Dropout(0.5),
nn.Linear(512, 1))
def mish_test_cpu():
x = torch.randn(*shape, requires_grad=True)
m = Mish()
y = m(x)
def __init__(self,
block,
layers,
zero_init_residual=False,
groups=1,
norm_layer=None):
super(OctaveIID, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self.inplanes = 32
self.groups = groups
self.conv1 = nn.Conv2d(3,
self.inplanes,
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.bn1 = norm_layer(self.inplanes)
self.relu = Mish()
self.layer1 = self._make_layer(block,
64,
layers[0],
norm_layer=norm_layer,
First=True)
self.octaveCBR1 = OctaveCBR(64, 128)
self.layer2 = self._make_layer(block,
128,
layers[1],
norm_layer=norm_layer)
self.layer3 = self._make_layer(block,
128,
layers[2],
norm_layer=norm_layer)
self.octaveCBR2 = OctaveCBR(128, 64)
self.layer4 = self._make_layer(block,
64,
layers[3],
norm_layer=norm_layer)
self.basic_block_h = BasicBlock(32, 32)
self.basic_block_l = BasicBlock(32, 32)
self.Upsample = nn.Upsample(scale_factor=2)
self.conv2_h = nn.Conv2d(32,
3,
kernel_size=3,
stride=1,
padding=1,
bias=False)
self.conv2_l = nn.Conv2d(32,
3,
kernel_size=3,
stride=1,
padding=1,
bias=False)
self.sigmoid = nn.Sigmoid()
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight,
mode='fan_out',
nonlinearity='leaky_relu')
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
if zero_init_residual:
for m in self.modules():
if isinstance(m, Bottleneck):
nn.init.constant_(m.bn3.weight, 0)
elif isinstance(m, BasicBlock):
nn.init.constant_(m.bn2.weight, 0)
i = torch.Tensor(1, 3, 512, 512).cuda()
FOCconv = FirstOctaveConv(kernel_size=(3, 3),
in_channels=3,
out_channels=128,
dilation=3).cuda()
x_out, y_out = FOCconv(i)
print("First: ", x_out.size(), y_out.size())
# test last Octave Cov
LOCconv = LastOctaveConv(kernel_size=(3, 3),
in_channels=256,
out_channels=128,
alpha=0.75).cuda()
i = high, low
out = LOCconv(i)
print("Last: ", out.size())
# test OCB
ocb = OctaveCB(in_channels=256,
out_channels=128,
alpha=0.75,
norm_layer=nn.GroupNorm).cuda()
i = high, low
x_out_h, y_out_l = ocb(i)
print("OCB:", x_out_h.size(), y_out_l.size())
# test last OCB
ocb_last = LastOCtaveCBR(256, 128, alpha=0.75, activation=Mish()).cuda()
i = high, low
x_out_h = ocb_last(i)
print("Last OCB", x_out_h.size())
#! --------------------- #! PATHS #! --------------------- DATA_PATH = '../data/' ADJ_PATH = '../data/MAPK-adjacency_matrix.pt' GENEORDER_PATH = '../data/MAPK-gene_order.pkl' EXPR_PATH = '../data/expr/' OUTPUT_PATH = '../output/' GO_MATRIX_PATH = '../data/MAPK&overlap_pathway_matrix.pt' #! --------------------- #! MODEL ARCHITECTURE #! --------------------- ACTIVATION = Mish() # F.leaky_relu #! --------------------- #! PRETRAINING - WEIGHT TRANSFER #! --------------------- USE_PRETRAINED_WEIGHTS = True STATE_DICT_PATH = '../pretrained_weights/model_state_dict-EPOCH_18.pkl' #GCN_WEIGHTS_PATH = '../pretrained_weights/cancer127_state_dict.pkl' #! --------------------- #! TRAINING PARAMS #! --------------------- USE_SEED = False SEED = 0
def __init__(self, ngpu, use_decoder, use_mixconv, act_func):
super(InvGenerator, self).__init__()
self.ngpu = ngpu
self.use_decoder = use_decoder
self.use_mixconv = use_mixconv
self.enc_act_func, self.dec_act_func = act_func
if self.enc_act_func == 'mish':
EncActFuncs = [Mish() for i in range(4)]
elif self.enc_act_func == 'relu':
EncActFuncs = [nn.LeakyReLU(0.3, inplace=True) for i in range(4)]
if self.dec_act_func == 'mish':
DecActFuncs = [Mish() for i in range(4)]
elif self.dec_act_func == 'relu':
DecActFuncs = [nn.ReLU(inplace=True) for i in range(4)]
C1 = C2 = C3 = C4 = 5
encoder = [
# input is (nc=1) x 40 x 40
nn.Conv2d(nc, ndf * C1, 6, 1, 0, bias=False),
nn.BatchNorm2d(ndf * C1),
EncActFuncs[0],
# state size. (ndf * 2) x 35 x 35
nn.Conv2d(ndf * C1, ndf * C2, 5, 2, 1, bias=False),
nn.BatchNorm2d(ndf * C2),
EncActFuncs[1],
# state size. (ndf*2) x 16 x 16
nn.Conv2d(ndf * C2, ndf * C3, 4, 2, 1, bias=False),
nn.BatchNorm2d(ndf * C3),
EncActFuncs[2],
# state size. (ndf*2) x 8 x 8
nn.Conv2d(ndf * C3, ndf * C4, 4, 2, 1, bias=False),
nn.BatchNorm2d(ndf * C4),
EncActFuncs[3],
# state size. (ndf*4) x 4 x 4
# nn.Conv2d(ndf * C4, nz, 4, 1, 0, bias=False),
# nn.BatchNorm2d(ndf * 4),
# nn.ReLU(True),
# state size. (ndf*8) x 3 x 3
# nn.Conv2d(ndf * 4, nz, 3, 1, 0, bias=False),
# 480 * 1 * 1
]
if self.use_decoder:
encoder.append(nn.Conv2d(ndf * C4, nz, 4, 1, 0, bias=False))
else:
encoder.append(nn.Conv2d(ndf * C4, 40 * 12, 4, 1, 0, bias=False))
encoder.append(Flatten(40, 12))
self.encoder = nn.Sequential(*encoder)
decoder = [
nn.ConvTranspose2d(nz, ngf * 8, 3, 1, 0, bias=False), # 13
nn.BatchNorm2d(ngf * 8),
DecActFuncs[0],
# image size: (ngf*8) x 3 x 3
nn.ConvTranspose2d(ngf * 8, ngf * 4, 3, (2, 1), 1,
bias=False), # 16
nn.BatchNorm2d(ngf * 4),
DecActFuncs[1],
# image size: (ngf*4) x 5 x 3
nn.ConvTranspose2d(ngf * 4, ngf * 2, 4, 2, 1, bias=False), # 19
nn.BatchNorm2d(ngf * 2),
DecActFuncs[2],
# image size: (ngf*2) x 10 x 6
nn.ConvTranspose2d(ngf * 2, ngf, (4, 3), (2, 1), 1,
bias=False), # 22
Power(2),
nn.BatchNorm2d(ngf),
DecActFuncs[3],
]
if self.use_mixconv:
# image size: (ngf*2) x 20 x 5
decoder.append(
nn.ConvTranspose2d(ngf,
args.num_mix_comps,
4,
2,
1,
bias=False)) # 25
else:
decoder.append(
nn.ConvTranspose2d(ngf, num_img_channels, 4, 2, 1,
bias=False)) # 25
decoder.append(nn.Tanh())
# 1 x 40 x 12
# Flatten()
self.decoder = nn.Sequential(*decoder)
# @Software: PyCharm
import torch
import torch.nn as nn
# __all__=['']
# vgg相关参数的配置，这里BN层几乎是所有CNN的标配了，故这里不再实现无BN层的网络
from mish import Mish
cfg = {
'VGG11': [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'],
'VGG13': [64, 64, 'M', 128, 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'],
'VGG16': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M'],
'VGG19': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512, 'M'],
}
act_fn = Mish()
class VGG(nn.Module):
def __init__(self, vgg_name):
super(VGG, self).__init__()
self.features = self._make_layers(cfg[vgg_name])
self.avgpool = nn.AdaptiveAvgPool2d((7, 7))
self.classifier = nn.Sequential(
nn.Linear(512 * 7 * 7, 4096),
act_fn,
nn.Dropout(),
nn.Linear(4096, 4096),
Mish(),
nn.Dropout(),
nn.Linear(4096, 2), )