def __init__(self, image_dim, num_joints):
super(JORNet_light, self).__init__()
in_features = ceil(image_dim * 0.125)**2 * 256
self.num_joints = num_joints # hand and object
self.out_poses1 = cudafy(
nn.Linear(in_features=in_features, out_features=1000),
self.use_cuda)
self.out_poses2 = cudafy(
nn.Linear(in_features=1000, out_features=self.num_joints),
self.use_cuda)
def __init__(self, params_dict):
super(JORNet_light, self).__init__(params_dict)
if params_dict['hand_only']:
self.num_joints = 60 # hand
else:
self.num_joints = 66 # hand and object
self.out_poses1 = cudafy(
nn.Linear(in_features=160000, out_features=1000), self.use_cuda)
self.out_poses2 = cudafy(
nn.Linear(in_features=1000, out_features=self.num_joints),
self.use_cuda)
def map_out_conv(self, in_channels):
return cudafy(
hand_detection_net.HALNetConvBlock(kernel_size=3,
stride=1,
filters=21,
in_channels=in_channels,
padding=1), self.use_cuda)
def __init__(self, dims, num_joints=20):
super(JORNet, self).__init__()
res3out, res4out, conv4e, conv4f, h, w = dims_fun(dims[0], dims[1])
# self.innerprod1_size = 256 * h * w
self.crop_res = dims
self.num_joints = num_joints
# self.main_loss_conv = cudafy(HALNet.HALNetConvBlock(
# kernel_size=3, stride=1, filters=21, in_channels=256, padding=1),
# self.use_cuda)
self.main_loss_deconv1 = cudafy(
nn.Upsample(size=self.crop_res, mode='bilinear'), self.use_cuda)
#self.main_loss_deconv2 = cudafy(nn.Upsample(scale_factor=1, mode='bilinear'),
# self.use_cuda)
if self.cross_entropy:
self.softmax_final = cudafy(HALNet.SoftmaxLogProbability2D(),
self.use_cuda)
self.innerproduct1_joint1 = cudafy(
nn.Linear(in_features=res3out, out_features=200), self.use_cuda)
self.innerproduct2_joint1 = cudafy(
nn.Linear(in_features=200, out_features=self.num_joints * 3),
self.use_cuda)
self.innerproduct1_joint2 = cudafy(
nn.Linear(in_features=res4out, out_features=200), self.use_cuda)
self.innerproduct2_joint2 = cudafy(
nn.Linear(in_features=200, out_features=self.num_joints * 3),
self.use_cuda)
self.innerproduct1_joint3 = cudafy(
nn.Linear(in_features=conv4e, out_features=200), self.use_cuda)
self.innerproduct2_joint3 = cudafy(
nn.Linear(in_features=200, out_features=self.num_joints * 3),
self.use_cuda)
self.innerproduct1_joint_main = cudafy(
nn.Linear(in_features=conv4f, out_features=200), self.use_cuda)
self.innerproduct2_join_main = cudafy(
nn.Linear(in_features=200, out_features=self.num_joints * 3),
self.use_cuda)
def map_out_to_loss(self, innerprod1_size):
return cudafy(nn.Linear(in_features=innerprod1_size, out_features=200),
self.use_cuda)
def __init__(self, params_dict):
super(HALNet, self).__init__()
# initialize variables
self.joint_ixs = parse_model_param(params_dict,
'joint_ixs',
default_value="Mandatory")
self.use_cuda = parse_model_param(params_dict,
'use_cuda',
default_value=False)
self.num_joints = len(self.joint_ixs)
self.num_channels = params_dict['num_channels']
self.cross_entropy = parse_model_param(params_dict,
'cross_entropy',
default_value=True)
# build network
self.conv1 = cudafy(
HALNetConvBlock(kernel_size=7,
stride=1,
filters=64,
in_channels=self.num_channels,
padding=3), self.use_cuda)
self.mp1 = cudafy(nn.MaxPool2d(kernel_size=3, stride=2, padding=1),
self.use_cuda)
self.res2a = cudafy(
HALNetResBlockConv(stride=1,
filters1=64,
filters2=256,
padding_right1=1), self.use_cuda)
self.res2b = cudafy(HALNetResBlockIDSkip(filters1=64, filters2=256),
self.use_cuda)
self.res2c = cudafy(HALNetResBlockIDSkip(filters1=64, filters2=256),
self.use_cuda)
self.res3a = cudafy(
HALNetResBlockConv(stride=2,
filters1=128,
filters2=512,
padding_right3=1,
first_in_channels=256), self.use_cuda)
self.interm_loss1 = cudafy(
HALNetConvBlock(kernel_size=3,
stride=1,
filters=self.num_joints,
in_channels=512,
padding=1), self.use_cuda)
self.interm_loss1_deconv = cudafy(
nn.Upsample(scale_factor=4, mode='bilinear', align_corners=True),
self.use_cuda)
self.interm_loss1_softmax = cudafy(SoftmaxLogProbability2D(),
self.use_cuda)
self.res3b = cudafy(HALNetResBlockIDSkip(filters1=128, filters2=512),
self.use_cuda)
self.res3c = cudafy(HALNetResBlockIDSkip(filters1=128, filters2=512),
self.use_cuda)
self.res4a = cudafy(
HALNetResBlockConv(stride=2,
filters1=256,
filters2=1024,
padding_right3=1,
first_in_channels=512), self.use_cuda)
self.interm_loss2 = cudafy(
HALNetConvBlock(kernel_size=3,
stride=1,
filters=self.num_joints,
in_channels=1024,
padding=1), self.use_cuda)
self.interm_loss2_deconv = cudafy(
nn.Upsample(scale_factor=8, mode='bilinear', align_corners=True),
self.use_cuda)
self.interm_loss2_softmax = cudafy(SoftmaxLogProbability2D(),
self.use_cuda)
self.res4b = cudafy(HALNetResBlockIDSkip(filters1=256, filters2=1024),
self.use_cuda)
self.res4c = cudafy(HALNetResBlockIDSkip(filters1=256, filters2=1024),
self.use_cuda)
self.res4d = cudafy(HALNetResBlockIDSkip(filters1=256, filters2=1024),
self.use_cuda)
self.conv4e = cudafy(
HALNetConvBlock(kernel_size=3,
stride=1,
filters=512,
in_channels=1024,
padding=1), self.use_cuda)
self.interm_loss3 = cudafy(
HALNetConvBlock(kernel_size=3,
stride=1,
filters=self.num_joints,
in_channels=512,
padding=1), self.use_cuda)
self.interm_loss3_deconv = cudafy(
nn.Upsample(scale_factor=8, mode='bilinear', align_corners=True),
self.use_cuda)
self.interm_loss3_softmax = cudafy(SoftmaxLogProbability2D(),
self.use_cuda)
self.conv4f = cudafy(
HALNetConvBlock(kernel_size=3,
stride=1,
filters=256,
in_channels=512,
padding=1), self.use_cuda)
self.main_loss_conv = cudafy(
HALNetConvBlock(kernel_size=3,
stride=1,
filters=self.num_joints,
in_channels=256,
padding=1), self.use_cuda)
self.main_loss_deconv = cudafy(
nn.Upsample(scale_factor=8, mode='bilinear', align_corners=True),
self.use_cuda)
if self.cross_entropy:
self.softmax_final = cudafy(SoftmaxLogProbability2D(),
self.use_cuda)
def __init__(self, params_dict):
super(JORNet, self).__init__(params_dict)
self.num_joints = 20
self.main_loss_conv = cudafy(HALNet.HALNetConvBlock(
kernel_size=3, stride=1, filters=21, in_channels=256, padding=1),
self.use_cuda)
self.main_loss_deconv1 = cudafy(nn.Upsample(size=self.crop_res, mode='bilinear'), self.use_cuda)
#self.main_loss_deconv2 = cudafy(nn.Upsample(scale_factor=1, mode='bilinear'),
# self.use_cuda)
if self.cross_entropy:
self.softmax_final = cudafy(HALNet.
SoftmaxLogProbability2D(), self.use_cuda)
self.innerproduct1_joint1 = cudafy(
nn.Linear(in_features=524288, out_features=200), self.use_cuda)
self.innerproduct2_joint1 = cudafy(
nn.Linear(in_features=200, out_features=self.num_joints * 3), self.use_cuda)
self.innerproduct1_joint2 = cudafy(
nn.Linear(in_features=262144, out_features=200), self.use_cuda)
self.innerproduct2_joint2 = cudafy(
nn.Linear(in_features=200, out_features=self.num_joints * 3), self.use_cuda)
self.innerproduct1_joint3 = cudafy(
nn.Linear(in_features=131072, out_features=200), self.use_cuda)
self.innerproduct2_joint3 = cudafy(
nn.Linear(in_features=200, out_features=self.num_joints * 3), self.use_cuda)
self.innerproduct1_joint_main = cudafy(
nn.Linear(in_features=65536, out_features=200), self.use_cuda)
self.innerproduct2_join_main = cudafy(
nn.Linear(in_features=200, out_features=self.num_joints * 3), self.use_cuda)