def __init__(self, skeleton, vertex_num):
super(SkinRefineNet, self).__init__()
self.joint_num = len(skeleton)
self.vertex_num = vertex_num
self.fc_pose = make_linear_layers(
[self.joint_num * 3, 256, vertex_num * 3], relu_final=False)
self.fc_id = make_linear_layers([cfg.id_code_dim, 256, vertex_num * 3],
relu_final=False)
def __init__(self, joint_num):
super(ParamRegressor, self).__init__()
self.joint_num = joint_num
self.fc = make_linear_layers([self.joint_num * 3, 1024, 512],
use_bn=True)
self.fc_pose = make_linear_layers(
[512, 24 * 6], relu_final=False) # joint orientation
self.fc_shape = make_linear_layers([512, 10],
relu_final=False) # shape parameter
def __init__(self, joint_num):
super(PoseNet, self).__init__()
self.joint_num = joint_num # single hand
self.joint_deconv_1 = make_deconv_layers([2048,256,256,256])
self.joint_conv_1 = make_conv_layers([256,self.joint_num*cfg.output_hm_shape[0]],kernel=1,stride=1,padding=0,bnrelu_final=False)
self.joint_deconv_2 = make_deconv_layers([2048,256,256,256])
self.joint_conv_2 = make_conv_layers([256,self.joint_num*cfg.output_hm_shape[0]],kernel=1,stride=1,padding=0,bnrelu_final=False)
self.root_fc = make_linear_layers([2048,512,cfg.output_root_hm_shape],relu_final=False)
self.hand_fc = make_linear_layers([2048,512,2],relu_final=False)
def __init__(self, joint_num):
super(ParamRegressor, self).__init__()
self.joint_num = joint_num
self.fc = make_linear_layers([self.joint_num * 3, 1024, 512],
use_bn=True)
if 'FreiHAND' in cfg.trainset_3d + cfg.trainset_2d + [cfg.testset]:
self.fc_pose = make_linear_layers(
[512, 16 * 6], relu_final=False) # hand joint orientation
else:
self.fc_pose = make_linear_layers(
[512, 24 * 6], relu_final=False) # body joint orientation
self.fc_shape = make_linear_layers([512, 10],
relu_final=False) # shape parameter
def __init__(self, class_num):
super(Classifier, self).__init__()
self.class_num = class_num
if cfg.mode == 'rgb_only':
self.fc = make_linear_layers(
[cfg.resnet_feat_dim[cfg.img_resnet_type], self.class_num],
relu_final=False)
elif cfg.mode == 'pose_only':
self.fc = make_linear_layers(
[cfg.resnet_feat_dim[cfg.pose_resnet_type], self.class_num],
relu_final=False)
else:
self.fc = make_linear_layers([cfg.agg_feat_dim, self.class_num],
relu_final=False)
self.dropout = nn.Dropout(0.5)
def __init__(self):
super(Aggregator, self).__init__()
self.img_resnet_dim = cfg.resnet_feat_dim[cfg.img_resnet_type]
self.pose_resnet_dim = cfg.resnet_feat_dim[cfg.pose_resnet_type]
## temporal strided conv to fuse with RGB
self.pose_frame_num = (cfg.frame_per_seg -
1) * cfg.pose_frame_factor + 1
self.pose_temporal_conv = make_conv3d_layers(
[self.pose_resnet_dim, self.pose_resnet_dim],
kernel=(5, 1, 1),
stride=(cfg.pose_frame_factor, 1, 1),
padding=(2, 0, 0))
## pose gate layer
self.pose_gate_fc = make_linear_layers(
[self.pose_resnet_dim, cfg.agg_feat_dim], relu_final=False)
## aggregation layer
self.img_conv = make_conv_layers(
[self.img_resnet_dim, cfg.agg_feat_dim], kernel=1, padding=0)
self.img_norm = nn.LayerNorm([cfg.agg_feat_dim, 1, 1])
self.pose_conv = make_conv_layers(
[self.pose_resnet_dim, cfg.agg_feat_dim], kernel=1, padding=0)
self.pose_norm = nn.LayerNorm([cfg.agg_feat_dim, 1, 1])
def __init__(self, skeleton):
super(PoseNet, self).__init__()
self.joint_num = len(skeleton)
self.register_buffer(
'DoF',
torch.cat([
torch.from_numpy(skeleton[i]['DoF'])[None, :]
for i in range(self.joint_num)
]).cuda().float())
self.dof_num = int(torch.sum(self.DoF))
self.fc = make_linear_layers(
[cfg.backbone_img_feat_dim, 512, self.dof_num], relu_final=False)
def __init__(self, skeleton):
super(SkeletonRefineNet, self).__init__()
self.joint_num = len(skeleton)
self.fc_skeleton = make_linear_layers(
[cfg.id_code_dim, 64, self.joint_num * 3], relu_final=False)
def __init__(self):
super(BackboneNet, self).__init__()
self.resnet = ResNetBackbone(cfg.resnet_type)
self.fc = make_linear_layers([2048, cfg.backbone_img_feat_dim])
