def seg_pth_to_my_pth():
# from collections import OrderedDict
# new_state_dict = OrderedDict()
# state_dict = torch.load('./drn_d_38_cityscapes.pth')
# for k, v in state_dict.items():
# name = k.replace("base.", "layer")
# name = name.replace("seg", "fc")
# if 'up' in name:
# print("pass a up")
# continue
# print(k, name)
# new_state_dict[name] = v
# torch.save(new_state_dict, './new_drn_d_38_seg_pretrain.pth')
import drn, models
model = drn.drn_d_38()
model.fc = None
for k, v in model.state_dict().items():
print("org model key:", k)
model.load_state_dict(torch.load('./new_drn_d_38_seg_pretrain.pth'))
img = torch.randn((1, 3, 512, 512))
model.fc = models.conv1x1(512, 6)
out = model(img)
print(model)
print(out.shape)
def _build_network_head(self, outs):
self.subnets = nn.ModuleList()
for _ in range(self.num_subnets):
# create the subnets
self.subnets.append(SubUNet(self.subnet_config))
if self.num_subnets == 2:
# In case that we only have two subnets we produce only one output
# with sigmoid, the second probability is implicit in this case
count = self.num_subnets - 1
else:
count = self.num_subnets
# create the probability weights, this can be seen as p(z|x), i.e. the probability
# of a decision given the input image. To obtain a weighted "average" of the
# predictions of the subnets, we multiply this weight to their output.
self.weight_probabilities = conv1x1(outs, count)
def _build_network_head(self, outs):
self.subnets = nn.ModuleList()
for _ in range(self.num_subnets):
# We create each requested subnet
self.subnets.append(SubUNet(self.subnet_config))
self.final_conv = conv1x1(outs, self.num_subnets)
def _build_network_head(self, outs):
self.network_head = conv1x1(outs, self.in_channels)
def _build_network_head(self, outs):
# Mean and std as for each input channel
self.conv_final = conv1x1(outs, self.in_channels * 2)