class EmbedNet(nn.Sequential):
def __init__(self, configer):
super(EmbedNet, self).__init__()
self.configer = configer
self.num_classes = self.configer.get('data', 'num_classes')
self.backbone = BackboneSelector(configer).get_backbone()
num_features = self.backbone.get_num_features()
if 'caffe' in self.configer.get('network', 'backbone'):
self.low_features = nn.Sequential(
self.backbone.conv1,
self.backbone.bn1,
self.backbone.relu1,
self.backbone.conv2,
self.backbone.bn2,
self.backbone.relu2,
self.backbone.conv3,
self.backbone.bn3,
self.backbone.relu3,
self.backbone.maxpool,
self.backbone.layer1,
)
else:
self.low_features = nn.Sequential(
self.backbone.conv1,
self.backbone.bn1,
self.backbone.relu,
self.backbone.maxpool,
self.backbone.layer1,
)
self.high_features1 = nn.Sequential(self.backbone.layer2,
self.backbone.layer3)
self.high_features2 = nn.Sequential(self.backbone.layer4)
self.embed_conv = EmbedModule(1024,
bn_type=self.configer.get(
'network', 'bn_type'))
self.decoder = PPMBilinearDeepsup(num_class=self.num_classes,
fc_dim=num_features,
bn_type=self.configer.get(
'network', 'bn_type'))
def forward(self, x):
low = self.low_features(x)
aux = self.high_features1(low)
incr = self.embed_conv(aux)
x = self.high_features2(aux + incr)
x, aux = self.decoder([x, aux])
x = F.interpolate(x,
scale_factor=8,
mode="bilinear",
align_corners=False)
return x, aux, incr
class OpenPose(nn.Module):
def __init__(self, configer):
super(OpenPose, self).__init__()
self.configer = configer
self.backbone = BackboneSelector(configer).get_backbone()
self.pose_model = PoseModel(configer, self.backbone.get_num_features())
def forward(self, x):
x = self.backbone(x)
out = self.pose_model(x)
return out
class PSPNet(nn.Sequential):
def __init__(self, configer):
super(PSPNet, self).__init__()
self.configer = configer
self.num_classes = self.configer.get('data', 'num_classes')
self.backbone = BackboneSelector(configer).get_backbone()
num_features = self.backbone.get_num_features()
self.dsn = nn.Sequential(
_ConvBatchNormReluBlock(num_features // 2,
num_features // 4,
3,
1,
norm_type=self.configer.get(
'network', 'norm_type')),
nn.Dropout2d(0.1),
nn.Conv2d(num_features // 4, self.num_classes, 1, 1, 0))
self.ppm = PPMBilinearDeepsup(fc_dim=num_features,
norm_type=self.configer.get(
'network', 'norm_type'))
self.cls = nn.Sequential(
nn.Conv2d(num_features + 4 * 512,
512,
kernel_size=3,
padding=1,
bias=False),
ModuleHelper.BNReLU(512,
norm_type=self.configer.get(
'network', 'norm_type')),
nn.Dropout2d(0.1), nn.Conv2d(512, self.num_classes, kernel_size=1))
def forward(self, data_dict):
x = self.backbone(data_dict['img'])
aux_x = self.dsn(x[-2])
x = self.ppm(x[-1])
x = self.cls(x)
aux_x = F.interpolate(aux_x,
size=(data_dict['img'].size(2),
data_dict['img'].size(3)),
mode="bilinear",
align_corners=True)
x = F.interpolate(x,
size=(data_dict['img'].size(2),
data_dict['img'].size(3)),
mode="bilinear",
align_corners=True)
return dict(aux_out=aux_x, out=x)
class PSPNet(nn.Sequential):
def __init__(self, configer):
super(PSPNet, self).__init__()
self.configer = configer
self.num_classes = self.configer.get('data', 'num_classes')
self.backbone = BackboneSelector(configer).get_backbone()
num_features = self.backbone.get_num_features()
self.low_features = nn.Sequential(
self.backbone.conv1,
self.backbone.bn1,
self.backbone.relu,
self.backbone.maxpool,
self.backbone.layer1,
)
self.high_features1 = nn.Sequential(self.backbone.layer2,
self.backbone.layer3)
self.high_features2 = nn.Sequential(self.backbone.layer4)
self.decoder = PPMBilinearDeepsup(num_class=self.num_classes,
fc_dim=num_features,
bn_type=self.configer.get(
'network', 'bn_type'))
def forward(self, x_):
low = self.low_features(x_)
aux = self.high_features1(low)
x = self.high_features2(aux)
x, aux = self.decoder([x, aux])
x = F.interpolate(x,
size=(x_.size(2), x_.size(3)),
mode="bilinear",
align_corners=False)
return x, aux
class DenseASPP(nn.Module):
"""
* output_scale can only set as 8 or 16
"""
def __init__(self, configer):
super(DenseASPP, self).__init__()
self.configer = configer
dropout0 = 0.1
dropout1 = 0.1
self.backbone = BackboneSelector(configer).get_backbone()
num_features = self.backbone.get_num_features()
self.trans = _Transition(num_input_features=self.num_features,
num_output_features=self.num_features // 2,
bn_type=self.configer.get('network', 'bn_type'))
self.num_features = self.num_features // 2
self.ASPP_3 = _DenseAsppBlock(input_num=num_features, num1=256, num2=64,
dilation_rate=3, drop_out=dropout0,
bn_type=self.configer.get('network', 'bn_type'))
self.ASPP_6 = _DenseAsppBlock(input_num=num_features + 64 * 1, num1=256, num2=64,
dilation_rate=6, drop_out=dropout0,
bn_type=self.configer.get('network', 'bn_type'))
self.ASPP_12 = _DenseAsppBlock(input_num=num_features + 64 * 2, num1=256, num2=64,
dilation_rate=12, drop_out=dropout0,
bn_type=self.configer.get('network', 'bn_type'))
self.ASPP_18 = _DenseAsppBlock(input_num=num_features + 64 * 3, num1=256, num2=64,
dilation_rate=18, drop_out=dropout0,
bn_type=self.configer.get('network', 'bn_type'))
self.ASPP_24 = _DenseAsppBlock(input_num=num_features + 64 * 4, num1=256, num2=64,
dilation_rate=24, drop_out=dropout0,
bn_type=self.configer.get('network', 'bn_type'))
num_features = num_features + 5 * 64
self.classification = nn.Sequential(
nn.Dropout2d(p=dropout1),
nn.Conv2d(in_channels=num_features,
out_channels=self.configer.get('network', 'out_channels'), kernel_size=1, padding=0)
)
def forward(self, x):
feature = self.backbone(x)
aspp3 = self.ASPP_3(feature)
feature = torch.cat((aspp3, feature), dim=1)
aspp6 = self.ASPP_6(feature)
feature = torch.cat((aspp6, feature), dim=1)
aspp12 = self.ASPP_12(feature)
feature = torch.cat((aspp12, feature), dim=1)
aspp18 = self.ASPP_18(feature)
feature = torch.cat((aspp18, feature), dim=1)
aspp24 = self.ASPP_24(feature)
feature = torch.cat((aspp24, feature), dim=1)
cls = self.classification(feature)
out = F.interpolate(cls, scale_factor=8, mode='bilinear', align_corners=False)
return out
class DenseASPP(nn.Module):
"""
* output_scale can only set as 8 or 16
"""
def __init__(self, configer):
super(DenseASPP, self).__init__()
self.configer = configer
dropout0 = 0.1
dropout1 = 0.1
self.backbone = BackboneSelector(configer).get_backbone()
num_features = self.backbone.get_num_features()
self.trans = _Transition(num_input_features=self.num_features,
num_output_features=self.num_features // 2,
bn_type=self.configer.get('network', 'bn_type'))
self.num_features = self.num_features // 2
self.ASPP_3 = _DenseAsppBlock(input_num=num_features, num1=256, num2=64,
dilation_rate=3, drop_out=dropout0,
bn_type=self.configer.get('network', 'bn_type'))
self.ASPP_6 = _DenseAsppBlock(input_num=num_features + 64 * 1, num1=256, num2=64,
dilation_rate=6, drop_out=dropout0,
bn_type=self.configer.get('network', 'bn_type'))
self.ASPP_12 = _DenseAsppBlock(input_num=num_features + 64 * 2, num1=256, num2=64,
dilation_rate=12, drop_out=dropout0,
bn_type=self.configer.get('network', 'bn_type'))
self.ASPP_18 = _DenseAsppBlock(input_num=num_features + 64 * 3, num1=256, num2=64,
dilation_rate=18, drop_out=dropout0,
bn_type=self.configer.get('network', 'bn_type'))
self.ASPP_24 = _DenseAsppBlock(input_num=num_features + 64 * 4, num1=256, num2=64,
dilation_rate=24, drop_out=dropout0,
bn_type=self.configer.get('network', 'bn_type'))
num_features = num_features + 5 * 64
self.classification = nn.Sequential(
nn.Dropout2d(p=dropout1),
nn.Conv2d(in_channels=num_features,
out_channels=self.configer.get('network', 'out_channels'), kernel_size=1, padding=0)
)
self.upsample = nn.Sequential(nn.Upsample(scale_factor=8, mode='bilinear', align_corners=True))
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.xavier_uniform_(m.weight.data)
elif isinstance(m, ModuleHelper.BatchNorm2d(bn_type=self.configer.get('network', 'bn_type'))):
m.weight.data.fill_(1)
m.bias.data.zero_()
def forward(self, x):
feature = self.backbone(x)
aspp3 = self.ASPP_3(feature)
feature = torch.cat((aspp3, feature), dim=1)
aspp6 = self.ASPP_6(feature)
feature = torch.cat((aspp6, feature), dim=1)
aspp12 = self.ASPP_12(feature)
feature = torch.cat((aspp12, feature), dim=1)
aspp18 = self.ASPP_18(feature)
feature = torch.cat((aspp18, feature), dim=1)
aspp24 = self.ASPP_24(feature)
feature = torch.cat((aspp24, feature), dim=1)
cls = self.classification(feature)
out = self.upsample(cls)
return out
class DenseASPP(nn.Module):
def __init__(self, configer):
super(DenseASPP, self).__init__()
self.configer = configer
det_features = self.configer.get('details', 'num_feature_list')
self.num_classes = self.configer.get('data', 'num_classes')
self.features = BackboneSelector(configer).get_backbone()
num_features = self.features.get_num_features()
self.ASPP_3 = nn.Sequential(nn.ReLU(inplace=True),
nn.Conv2d(in_channels=num_features, out_channels=256, kernel_size=1),
nn.BatchNorm2d(num_features=256),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=256, out_channels=64, kernel_size=3, dilation=3, padding=3),
nn.Dropout2d(p=0.1))
self.ASPP_6 = nn.Sequential(nn.BatchNorm2d(num_features=num_features + 64 * 1),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=num_features + 64 * 1, out_channels=256, kernel_size=1),
nn.BatchNorm2d(num_features=256),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=256, out_channels=64, kernel_size=3, dilation=6, padding=6),
nn.Dropout2d(p=0.1))
self.ASPP_12 = nn.Sequential(nn.BatchNorm2d(num_features=num_features + 64 * 2),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=num_features + 64 * 2, out_channels=256, kernel_size=1),
nn.BatchNorm2d(num_features=256),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=256, out_channels=64, kernel_size=3, dilation=12,
padding=12),
nn.Dropout2d(p=0.1))
self.ASPP_18 = nn.Sequential(nn.BatchNorm2d(num_features=num_features + 64 * 3),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=num_features + 64 * 3, out_channels=256, kernel_size=1),
nn.BatchNorm2d(num_features=256),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=256, out_channels=64, kernel_size=3, dilation=18,
padding=18),
nn.Dropout2d(p=0.1))
self.ASPP_24 = nn.Sequential(nn.BatchNorm2d(num_features=num_features + 64 * 4),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=num_features + 64 * 4, out_channels=256, kernel_size=1),
nn.BatchNorm2d(num_features=256),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=256, out_channels=64, kernel_size=3, dilation=24, padding=24))
num_features += 5 * 64
self.det_feature1 = _RevertedResDetBlock(num_features, det_features[0], stride=1, expand_ratio=2)
self.det_feature2 = _RevertedResDetBlock(det_features[0], det_features[1], stride=2, expand_ratio=3)
self.det_feature3 = _RevertedResDetBlock(det_features[1], det_features[2], stride=2, expand_ratio=3)
self.det_feature4 = _RevertedResDetBlock(det_features[2], det_features[3], stride=2, expand_ratio=3)
self.det_feature5 = _RevertedResDetBlock(det_features[3], det_features[4], stride=2, expand_ratio=3)
self.multibox_layer = ShareMultiBoxLayer(configer)
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_uniform_(m.weight.data)
if m.bias is not None:
m.bias.data.zero_()
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
def forward(self, _input):
feature = self.features(_input)
aspp3 = self.ASPP_3(feature)
feature = torch.cat((aspp3, feature), dim=1)
aspp6 = self.ASPP_6(feature)
feature = torch.cat((aspp6, feature), dim=1)
aspp12 = self.ASPP_12(feature)
feature = torch.cat((aspp12, feature), dim=1)
aspp18 = self.ASPP_18(feature)
feature = torch.cat((aspp18, feature), dim=1)
aspp24 = self.ASPP_24(feature)
feature = torch.cat((aspp24, feature), dim=1)
det_feature = []
feature = self.det_feature1(feature)
det_feature.append(feature)
feature = self.det_feature2(feature)
det_feature.append(feature)
feature = self.det_feature3(feature)
det_feature.append(feature)
feature = self.det_feature4(feature)
det_feature.append(feature)
feature = self.det_feature5(feature)
det_feature.append(feature)
loc_preds, conf_preds = self.multibox_layer(det_feature)
return loc_preds, conf_preds
@staticmethod
def __extra_layer(num_in, num_out, num_c, stride, pad):
layer = nn.Sequential(
nn.Conv2d(num_in, num_c, kernel_size=1),
nn.ReLU(),
nn.Conv2d(num_c, num_out, kernel_size=3, stride=stride, padding=pad),
nn.ReLU(),
)
return layer
def load_pretrained_weight(self, net):
self.features.load_pretrained_weight(net)
class DenseASPP(nn.Module):
"""
* output_scale can only set as 8 or 16
"""
def __init__(self, configer):
super(DenseASPP, self).__init__()
self.configer = configer
dropout0 = MODEL_CONFIG['dropout0']
dropout1 = MODEL_CONFIG['dropout1']
d_feature0 = MODEL_CONFIG['d_feature0']
d_feature1 = MODEL_CONFIG['d_feature1']
self.backbone = BackboneSelector(configer).get_backbone()
num_features = self.backbone.get_num_features()
self.trans = _Transition(num_input_features=self.num_features,
num_output_features=self.num_features // 2,
norm_type=self.configer.get(
'network', 'norm_type'))
self.num_features = self.num_features // 2
self.ASPP_3 = _DenseAsppBlock(input_num=num_features,
num1=d_feature0,
num2=d_feature1,
dilation_rate=3,
drop_out=dropout0,
norm_type=self.configer.get(
'network', 'norm_type'))
self.ASPP_6 = _DenseAsppBlock(input_num=num_features + d_feature1 * 1,
num1=d_feature0,
num2=d_feature1,
dilation_rate=6,
drop_out=dropout0,
norm_type=self.configer.get(
'network', 'norm_type'))
self.ASPP_12 = _DenseAsppBlock(input_num=num_features + d_feature1 * 2,
num1=d_feature0,
num2=d_feature1,
dilation_rate=12,
drop_out=dropout0,
norm_type=self.configer.get(
'network', 'norm_type'))
self.ASPP_18 = _DenseAsppBlock(input_num=num_features + d_feature1 * 3,
num1=d_feature0,
num2=d_feature1,
dilation_rate=18,
drop_out=dropout0,
norm_type=self.configer.get(
'network', 'norm_type'))
self.ASPP_24 = _DenseAsppBlock(input_num=num_features + d_feature1 * 4,
num1=d_feature0,
num2=d_feature1,
dilation_rate=24,
drop_out=dropout0,
norm_type=self.configer.get(
'network', 'norm_type'))
num_features = num_features + 5 * d_feature1
self.classification = nn.Sequential(
nn.Dropout2d(p=dropout1),
nn.Conv2d(num_features,
self.configer.get('data', 'num_classes'),
kernel_size=1,
padding=0))
def forward(self, data_dict):
x = self.backbone(data_dict['img'])
feature = self.trans(x)
aspp3 = self.ASPP_3(feature)
feature = torch.cat((aspp3, feature), dim=1)
aspp6 = self.ASPP_6(feature)
feature = torch.cat((aspp6, feature), dim=1)
aspp12 = self.ASPP_12(feature)
feature = torch.cat((aspp12, feature), dim=1)
aspp18 = self.ASPP_18(feature)
feature = torch.cat((aspp18, feature), dim=1)
aspp24 = self.ASPP_24(feature)
feature = torch.cat((aspp24, feature), dim=1)
x = self.classification(feature)
x = F.interpolate(x,
size=(data_dict['img'].size(2),
data_dict['img'].size(3)),
mode="bilinear",
align_corners=True)
return dict(out=out)
