def __init__(self, num_class=2, deploy=False, width_multiplier=[0.75, 0.75, 0.75, 2.5], num_blocks=[2, 4, 14, 1], override_groups_map=None):
super(FusionNet, self).__init__()
self.deploy = deploy
self.cur_layer_idx = 1
self.in_planes = 384
self.override_groups_map = override_groups_map or dict()
assert 0 not in self.override_groups_map
self.color_moudle = Net(num_class=num_class, is_first_bn=True)
self.depth_moudle = Net(num_class=num_class, is_first_bn=True)
self.ir_moudle = Net(num_class=num_class, is_first_bn=True)
self.color_SE = SEModule(128,reduction=16)
self.depth_SE = SEModule(128,reduction=16)
self.ir_SE = SEModule(128,reduction=16)
self.res_0 = self._make_layer(BasicBlock, 384, 256, 2, stride=2)
self.res_1 = self._make_layer(BasicBlock, 256, 512, 2, stride=2)
# self.res_0 = self._make_RepVGG_layer(384, num_blocks[2], stride=2)
# self.res_1 = self._make_RepVGG_layer(int(512 * width_multiplier[3]), num_blocks[3], stride=2)
self.fc = nn.Sequential(nn.Dropout(0.5),
# nn.Linear(int(512 * width_multiplier[3]), 256),
nn.Linear(int(512), 256),
nn.ReLU(inplace=True),
nn.Linear(256, num_class))
def __init__(self, num_class=2, modality='fusion'):
super(FusionNet, self).__init__()
# Net是model_baseline中的net，返回
# logit.shape: torch.Size([batch_size, 2])
# logit.shape: torch.Size([batch_size, 300])
# fea.shape: torch.Size([batch_size, 512])
self.modality = modality
if self.modality == 'fusion':
self.color_moudle = Net(num_class=num_class, is_first_bn=True)
self.depth_moudle = Net(num_class=num_class, is_first_bn=True)
self.ir_moudle = Net(num_class=num_class, is_first_bn=True)
# SEModule，输入channels和reduction,这个channel要和前一个网络的输出维度一致
self.color_SE = SEModule(128, reduction=16)
self.depth_SE = SEModule(128, reduction=16)
self.ir_SE = SEModule(128, reduction=16)
# 采用resnet的方式创建两个层
self.res_0 = self._make_layer(BasicBlock, 384, 256, 2, stride=2)
else:
self.color_moudle = Net(num_class=num_class, is_first_bn=True)
self.color_SE = SEModule(128, reduction=16)
self.res_0 = self._make_layer(BasicBlock, 128, 256, 2, stride=2)
self.res_1 = self._make_layer(BasicBlock, 256, 512, 2, stride=2)
self.fc = nn.Sequential(nn.Dropout(0.5), nn.Linear(512, 256),
nn.ReLU(inplace=True),
nn.Linear(256, num_class))
def __init__(self, num_class=2):
super(FusionNet, self).__init__()
self.color_moudle = Net(num_class=num_class, is_first_bn=True)
self.depth_moudle = Net(num_class=num_class, is_first_bn=True)
self.ir_moudle = Net(num_class=num_class, is_first_bn=True)
self.res_0 = self._make_layer(BasicBlock, 384, 256, 2, stride=2)
self.res_1 = self._make_layer(BasicBlock, 256, 512, 2, stride=2)
self.fc = nn.Sequential(nn.Dropout(0.5), nn.Linear(512, 256),
nn.ReLU(inplace=True),
nn.Linear(256, num_class))
def get_model(model_name, num_class,is_first_bn):
if model_name == 'baseline':
from model.model_baseline import Net
elif model_name == 'model_A':
from model.FaceBagNet_model_A import Net
elif model_name == 'model_B':
from model.FaceBagNet_model_B import Net
elif model_name == 'model_C':
from model.FaceBagNet_model_C import Net
net = Net(num_class=num_class,is_first_bn=is_first_bn)
return net
def run_check_net():
batch_size = 32
C, H, W = 3, 128, 128
num_class = 2
input = np.random.uniform(0, 1, (batch_size, C, H, W)).astype(np.float32)
truth = np.random.choice(num_class, batch_size).astype(np.float32)
#------------
input = torch.from_numpy(input).float().cuda()
truth = torch.from_numpy(truth).long().cuda()
input = to_var(input)
truth = to_var(truth)
#---
criterion = softmax_cross_entropy_criterion
net = Net(num_class).cuda()
net.set_mode('backup')
print(net)
logit = net.forward(input)
loss = criterion(logit, truth)
def run_check_net():
num_class = 2
net = Net(num_class)
print(net)
def run_check_net():
num_class = 2
x = torch.rand(36, 9, 48, 48)
net = Net(num_class)
output = net.forward(x)