def __init__(self, config):
super().__init__()
self.config = config
self.name = self.__class__.__name__
self.backbone = backbone(config.model)
if hasattr(self.config.model, 'backbone_freeze'):
if self.config.model.backbone_freeze:
print('freeze backbone weights' + '*' * 30)
freeze_layer(self.backbone)
self.upsample_layer = self.config.model.upsample_layer
self.class_number = self.config.model.class_number
self.input_shape = self.config.model.input_shape
self.dataset_name = self.config.dataset.name
self.ignore_index = self.config.dataset.ignore_index
self.edge_class_num = self.config.dataset.edge_class_num
self.midnet_input_shape = self.backbone.get_output_shape(
self.upsample_layer, self.input_shape)
self.midnet_out_channels = 2 * self.midnet_input_shape[1]
self.midnet = get_midnet(self.config, self.midnet_input_shape,
self.midnet_out_channels)
# out feature channels 512
self.branch_edge = get_suffix_net(self.config,
self.midnet_out_channels,
self.edge_class_num)
# out feature channels 512
self.branch_seg = get_suffix_net(self.config, self.midnet_out_channels,
self.class_number)
# input=concat(512,512)
self.feature_conv = conv_bn_relu(in_channels=512 + 512,
out_channels=512,
kernel_size=1,
stride=1,
padding=0)
self.seg = get_suffix_net(self.config, 512, self.class_number)
def __init__(self,config):
super().__init__()
decoder_config=dict2edict(config)
self.input_shape=config.input_shape
self.upsample_layer=config['upsample_layer']
self.backbone=backbone(config,use_none_layer=config.use_none_layer)
self.midnet_input_shape=self.backbone.get_output_shape(self.upsample_layer,self.input_shape)
self.midnet_out_channels=2+self.midnet_input_shape[1]
self.concat_size=self.midnet_input_shape[2:]
self.class_number=config.class_number
self.decoder=get_suffix_net(decoder_config,
self.midnet_out_channels,
self.class_number)
def __init__(self, config):
super().__init__()
self.config = config
self.name = self.__class__.__name__
self.upsample_layer = config.model.upsample_layer
self.class_number = config.model.class_number
self.backbone = backbone(config.model,
use_none_layer=config.model.use_none_layer)
self.panet = panet(config.model)
self.midnet = transform_panet(self.backbone, self.panet, self.config)
self.midnet_out_channels = self.backbone.get_feature_map_channel(
self.upsample_layer)
self.decoder = get_suffix_net(config, self.midnet_out_channels,
self.class_number)
def __init__(self, config):
super().__init__()
self.config = config
self.name = self.__class__.__name__
self.backbone = backbone(config.model,
use_none_layer=config.model.use_none_layer)
self.upsample_layer = self.config.model.upsample_layer
self.class_number = self.config.model.class_number
self.input_shape = self.config.model.input_shape
self.dataset_name = self.config.dataset.name
self.ignore_index = self.config.dataset.ignore_index
self.midnet_input_shape = self.backbone.get_output_shape(
self.upsample_layer, self.input_shape)
self.midnet_out_channels = self.midnet_input_shape[1]
self.decoder = get_suffix_net(config, self.midnet_out_channels,
self.class_number)
def __init__(self, config):
super().__init__()
self.config = config
self.name = self.__class__.__name__
print('warnniing: use_none_layer is false for motionnet')
self.backbone = backbone(config.model,
use_none_layer=config.model.use_none_layer)
self.upsample_type = self.config.model.upsample_type
self.upsample_layer = self.config.model.upsample_layer
self.class_number = self.config.model.class_number
self.input_shape = self.config.model.input_shape
self.dataset_name = self.config.dataset.name
self.ignore_index = self.config.dataset.ignore_index
self.midnet = transform_segnet(self.backbone, self.config)
self.midnet_out_channels = self.backbone.get_feature_map_channel(
self.upsample_layer)
self.decoder = get_suffix_net(config, self.midnet_out_channels,
self.class_number)
def __init__(self, config):
super().__init__()
self.config = config
self.name = self.__class__.__name__
self.backbone = backbone(config.model,
use_none_layer=config.model.use_none_layer)
self.upsample_layer = self.config.model.upsample_layer
self.class_number = self.config.model.class_number
self.input_shape = self.config.model.input_shape
self.dataset_name = self.config.dataset.name
self.ignore_index = self.config.dataset.ignore_index
self.auxnet_layer = self.config.model.auxnet_layer
assert self.auxnet_layer >= 3
assert self.upsample_layer >= 3
assert self.upsample_layer >= self.auxnet_layer
assert config.model.use_none_layer == True
# use modified backbone, the output shape is the same for upsample_layer=[3,4,5]
self.auxnet_input_shape = self.backbone.get_output_shape(
self.auxnet_layer, self.input_shape)
self.midnet_input_shape = self.backbone.get_output_shape(
self.upsample_layer, self.input_shape)
self.midnet_out_channels = 2 * self.midnet_input_shape[1]
self.auxnet_out_channels = self.auxnet_input_shape[1]
self.midnet = get_midnet(self.config, self.midnet_input_shape,
self.midnet_out_channels)
self.auxnet = get_suffix_net(self.config,
self.auxnet_out_channels,
self.class_number,
aux=True)
self.decoder = get_suffix_net(self.config, self.midnet_out_channels,
self.class_number)
if config.model.use_lr_mult:
if config.model.use_none_layer and config.model.backbone_pretrained and self.upsample_layer >= 4:
backbone_optmizer_params = get_backbone_optimizer_params(
config.model.backbone_name,
self.backbone.model,
unchanged_lr_mult=1,
changed_lr_mult=config.model.changed_lr_mult,
new_lr_mult=config.model.new_lr_mult)
else:
backbone_optmizer_params = [{
'params':
[p for p in self.backbone.parameters() if p.requires_grad],
'lr_mult':
1
}]
self.optimizer_params = backbone_optmizer_params + [
{
'params': self.midnet.parameters(),
'lr_mult': config.model.new_lr_mult
}, {
'params': self.auxnet.parameters(),
'lr_mult': config.model.new_lr_mult
}, {
'params': self.decoder.parameters(),
'lr_mult': config.model.new_lr_mult
}
]
else:
self.optimizer_params = [{
'params':
[p for p in self.backbone.parameters() if p.requires_grad],
'lr_mult':
1
}, {
'params': self.midnet.parameters(),
'lr_mult': 1
}, {
'params': self.auxnet.parameters(),
'lr_mult': 1
}, {
'params': self.decoder.parameters(),
'lr_mult': 1
}]
if not hasattr(config.model, 'aux_base_weight'):
config.model.aux_base_weight = 0.4
def __init__(self, config):
super().__init__()
self.config = config
self.name = self.__class__.__name__
use_none_layer = config.model.use_none_layer
self.backbone = backbone(config.model, use_none_layer=use_none_layer)
self.upsample_layer = self.config.model.upsample_layer
self.class_number = self.config.model.class_number
self.input_shape = self.config.model.input_shape
self.dataset_name = self.config.dataset.name
self.ignore_index = self.config.dataset.ignore_index
self.midnet_input_shape = self.backbone.get_output_shape(
self.upsample_layer, self.input_shape)
# self.midnet_out_channels=self.config.model.midnet_out_channels
self.midnet_out_channels = 2 * self.midnet_input_shape[1]
self.midnet = get_midnet(self.config, self.midnet_input_shape,
self.midnet_out_channels)
self.decoder = get_suffix_net(config, self.midnet_out_channels,
self.class_number)
self.center_channels = self.decoder.center_channels
# for pretrained module, use small lr_mult=1
# for modified module, use middle lr_mult=10
# for new module, use largest lr_mult=20
# for resnet, the begin layers is newed and the end layers is changed
if self.config.model.backbone_freeze:
self.optimizer_params = [{
'params': self.midnet.parameters(),
'lr_mult': 1
}, {
'params': self.decoder.parameters(),
'lr_mult': 1
}]
elif config.model.use_lr_mult:
if use_none_layer and config.model.backbone_pretrained and self.upsample_layer >= 4:
backbone_optmizer_params = get_backbone_optimizer_params(
config.model.backbone_name,
self.backbone,
unchanged_lr_mult=config.model.pre_lr_mult,
changed_lr_mult=config.model.changed_lr_mult,
new_lr_mult=config.model.new_lr_mult)
else:
warnings.warn(
'config.model.use_lr_mult is True but not fully wored')
backbone_optmizer_params = [{
'params':
[p for p in self.backbone.parameters() if p.requires_grad],
'lr_mult':
config.model.pre_lr_mult
}]
self.optimizer_params = backbone_optmizer_params + [
{
'params': self.midnet.parameters(),
'lr_mult': config.model.new_lr_mult
}, {
'params': self.decoder.parameters(),
'lr_mult': config.model.new_lr_mult
}
]
else:
self.optimizer_params = [{
'params':
[p for p in self.backbone.parameters() if p.requires_grad],
'lr_mult':
1
}, {
'params': self.midnet.parameters(),
'lr_mult': 1
}, {
'params': self.decoder.parameters(),
'lr_mult': 1
}]
def __init__(self, config):
super().__init__()
self.config = config
self.name = self.__class__.__name__
self.backbone = backbone(config.model)
if hasattr(self.config.model, 'backbone_freeze'):
if self.config.model.backbone_freeze:
print('freeze backbone weights' + '*' * 30)
freeze_layer(self.backbone)
self.upsample_layer = self.config.model.upsample_layer
self.class_number = self.config.model.class_number
self.input_shape = self.config.model.input_shape
self.dataset_name = self.config.dataset.name
self.ignore_index = self.config.dataset.ignore_index
self.edge_class_num = self.config.dataset.edge_class_num
self.midnet_input_shape = self.backbone.get_output_shape(
self.upsample_layer, self.input_shape)
self.midnet_out_channels = 2 * self.midnet_input_shape[1]
self.midnet = get_midnet(self.config, self.midnet_input_shape,
self.midnet_out_channels)
if hasattr(self.config.model, 'edge_seg_order'):
self.edge_seg_order = self.config.model.edge_seg_order
print('the edge and seg order is %s' % self.edge_seg_order,
'*' * 30)
assert self.edge_seg_order in [
'same', 'first', 'later'
], 'unexcepted edge seg order %s' % self.edge_seg_order
else:
self.edge_seg_order = 'same'
if self.edge_seg_order == 'same':
self.seg_decoder = get_suffix_net(self.config,
self.midnet_out_channels,
self.class_number)
self.edge_decoder = get_suffix_net(self.config,
self.midnet_out_channels,
self.edge_class_num)
elif self.edge_seg_order == 'later':
self.seg_decoder = get_suffix_net(self.config,
self.midnet_out_channels,
self.class_number)
self.edge_decoder = get_suffix_net(self.config, 512,
self.edge_class_num)
else:
self.edge_decoder = get_suffix_net(self.config,
self.midnet_out_channels,
self.edge_class_num)
self.feature_conv = conv_bn_relu(
in_channels=self.midnet_out_channels,
out_channels=512,
kernel_size=1,
stride=1,
padding=0)
# the input is torch.cat[self.edge_class_num,self.class_number]
self.seg_conv = conv_bn_relu(in_channels=512 + 512,
out_channels=512,
kernel_size=1,
stride=1,
padding=0)
self.seg_decoder = get_suffix_net(self.config, 512,
self.class_number)
def __init__(self, config):
super().__init__()
self.config = config
self.name = self.__class__.__name__
self.backbone = backbone(config.model)
if hasattr(self.config.model, 'backbone_freeze'):
if self.config.model.backbone_freeze:
print('freeze backbone weights' + '*' * 30)
freeze_layer(self.backbone)
self.upsample_layer = self.config.model.upsample_layer
self.class_number = self.config.model.class_number
self.input_shape = self.config.model.input_shape
self.dataset_name = self.config.dataset.name
self.ignore_index = self.config.dataset.ignore_index
self.edge_class_num = self.config.dataset.edge_class_num
self.cross_merge_times = self.config.model.cross_merge_times
self.midnet_input_shape = self.backbone.get_output_shape(
self.upsample_layer, self.input_shape)
self.midnet_out_channels = 2 * self.midnet_input_shape[1]
self.midnet = get_midnet(self.config, self.midnet_input_shape,
self.midnet_out_channels)
self.seg0 = get_suffix_net(self.config, self.midnet_out_channels,
self.class_number)
self.edge0 = get_suffix_net(self.config, self.midnet_out_channels,
self.edge_class_num)
seg_list = []
edge_list = []
# before concat
seg_conv_list = []
edge_conv_list = []
feature_channel = 512
concat_channel = feature_channel // 2
#TODO use psp other than conv
for i in range(self.cross_merge_times):
seg_conv_list.append(
conv_bn_relu(in_channels=feature_channel,
out_channels=concat_channel,
kernel_size=1,
stride=1,
padding=0))
edge_conv_list.append(
conv_bn_relu(in_channels=feature_channel,
out_channels=concat_channel,
kernel_size=1,
stride=1,
padding=0))
seg_list.append(
get_suffix_net(self.config, 2 * concat_channel,
self.class_number))
edge_list.append(
get_suffix_net(self.config, 2 * concat_channel,
self.edge_class_num))
self.seg_list = TN.ModuleList(seg_list)
self.edge_list = TN.ModuleList(edge_list)
self.seg_conv_list = TN.ModuleList(seg_conv_list)
self.edge_conv_list = TN.ModuleList(edge_conv_list)
def __init__(self, config):
super().__init__()
self.config = config
self.name = self.__class__.__name__
use_none_layer = config.model.use_none_layer
self.backbone = backbone(config.model, use_none_layer=use_none_layer)
if hasattr(self.config.model, 'backbone_freeze'):
if self.config.model.backbone_freeze:
# print('freeze backbone weights'+'*'*30)
freeze_layer(self.backbone)
self.upsample_layer = self.config.model.upsample_layer
self.class_number = self.config.model.class_number
self.input_shape = self.config.model.input_shape
self.dataset_name = self.config.dataset.name
self.ignore_index = self.config.dataset.ignore_index
self.edge_class_num = self.config.dataset.edge_class_num
self.midnet_input_shape = self.backbone.get_output_shape(
self.upsample_layer, self.input_shape)
# self.midnet_out_channels=self.config.model.midnet_out_channels
self.midnet_out_channels = 2 * self.midnet_input_shape[1]
self.midnet = get_midnet(self.config, self.midnet_input_shape,
self.midnet_out_channels)
self.decoder = get_suffix_net(config, self.midnet_out_channels,
self.class_number)
layer_shapes = self.backbone.get_layer_shapes(self.input_shape)
print('layer shapes', layer_shapes)
edge_aux_list = []
for i in range(self.upsample_layer - 1):
edge_aux_list.append(
TN.Conv2d(in_channels=layer_shapes[i + 1][1],
out_channels=self.edge_class_num,
kernel_size=1))
self.edge_aux_list = TN.ModuleList(edge_aux_list)
self.edge_fusion_conv = TN.Conv2d(in_channels=self.edge_class_num *
(self.upsample_layer - 1),
out_channels=self.edge_class_num,
kernel_size=1)
self.optimizer_params = [{
'params':
[p for p in self.backbone.parameters() if p.requires_grad],
'lr_mult':
1
}, {
'params': self.edge_aux_list.parameters(),
'lr_mult': 1
}, {
'params': self.edge_fusion_conv.parameters(),
'lr_mult': 1
}, {
'params': self.midnet.parameters(),
'lr_mult': 10
}, {
'params': self.decoder.parameters(),
'lr_mult': 10
}]