def __init__(self, model_args):
super(GPNN_VCOCO_v2, self).__init__()
self.model_args = model_args.copy()
if model_args['resize_feature_to_message_size']:
# Resize large features
self.edge_feature_resize = torch.nn.Linear(
model_args['edge_feature_size'], model_args['message_size'])
self.node_feature_resize = torch.nn.Linear(
model_args['node_feature_size'], model_args['message_size'])
torch.nn.init.xavier_normal(self.edge_feature_resize.weight)
torch.nn.init.xavier_normal(self.node_feature_resize.weight)
model_args['edge_feature_size'] = model_args['message_size']
model_args['node_feature_size'] = model_args['message_size']
self.link_fun = units.LinkFunction('GraphConv', model_args)
self.sigmoid = torch.nn.Sigmoid()
self.message_fun = units.MessageFunction('linear_concat_relu',
model_args)
self.update_fun = units.UpdateFunction('gru', model_args)
self.readout_fun = units.ReadoutFunction(
'fc', {
'readout_input_size': model_args['node_feature_size'],
'output_classes': model_args['hoi_classes']
})
self.readout_fun2 = units.ReadoutFunction(
'fc', {
'readout_input_size': model_args['node_feature_size'],
'output_classes': model_args['roles_num']
})
self.propagate_layers = model_args['propagate_layers']
self._load_link_fun(model_args)
def __init__(self, model_args):
super(GPNN_CAD, self).__init__()
self.link_fun = units.LinkFunction('GraphConvLSTM', model_args)
self.message_fun = units.MessageFunction('linear_concat', model_args)
self.update_funs = torch.nn.ModuleList([])
self.update_funs.append(units.UpdateFunction('gru', model_args))
self.update_funs.append(units.UpdateFunction('gru', model_args))
self.subactivity_classes = model_args['subactivity_classes']
self.affordance_classes = model_args['affordance_classes']
self.readout_funs = torch.nn.ModuleList([])
self.readout_funs.append(
units.ReadoutFunction(
'fc_soft_max', {
'readout_input_size': model_args['node_feature_size'],
'output_classes': self.subactivity_classes
}))
self.readout_funs.append(
units.ReadoutFunction(
'fc_soft_max', {
'readout_input_size': model_args['node_feature_size'],
'output_classes': self.affordance_classes
}))
self.propagate_layers = model_args['propagate_layers']
self._load_link_fun(model_args)
def __init__(self, model_args):
super(HGNN_lstm, self).__init__()
self.model_args = model_args.copy()
self.resnet = units.Resnet.ResNet50(3)
self.link_fun = units.LinkFunction('GraphConv', model_args)
self.message_fun = units.MessageFunction('linear_concat_relu',
model_args)
self.update_fun = units.UpdateFunction('gru', model_args)
self.readout_fun = units.ReadoutFunction(
'fc', {
'readout_input_size': model_args['node_feature_size'],
'output_classes': model_args['big_attr_classes']
})
self.lstm = torch.nn.LSTM(model_args['roi_feature_size'],
self.model_args['lstm_hidden_size'])
self.lstm_readout = torch.nn.Linear(
self.model_args['lstm_hidden_size'],
self.model_args['big_attr_classes'])
self.propagate_layers = model_args['propagate_layers']
self.softmax = torch.nn.Softmax()
self.relu = torch.nn.ReLU()
self.sigmoid = torch.nn.Sigmoid()
self.maxpool = torch.nn.MaxPool2d((2, 2))
def __init__(self, model_args, args):
super(AttMat_msg_lstm, self).__init__()
self.model_args = model_args.copy()
# TODO: I don't think resnet34 is sufficient for this task,
# TODO: you'd better adopt resnet50 for feature extraction at least.
self.resnet = units.Resnet34(model_args['roi_feature_size'])
# TODO: The init of the ConvLSTM is important, double check the init procedure.
self.link_fun = units.LinkFunction('GraphConv', model_args)
self.base_model = torch.nn.ModuleList([])
self.load_base_model = args.load_base_model
self.base_model_weight = args.base_model_weight
self.freeze_base_model = args.freeze_base_model
self.message_fun = units.MessageFunction('linear_concat_relu',
model_args)
self.update_fun = units.UpdateFunction('gru', model_args)
self.readout_fun = units.ReadoutFunction(
'fc', {
'readout_input_size': model_args['node_feature_size'],
'output_classes': model_args['big_attr_classes']
})
self.base_model.append(self.resnet)
self.base_model.append(self.link_fun)
self.base_model.append(self.message_fun)
self.base_model.append(self.update_fun)
self.lstm = torch.nn.LSTM(model_args['roi_feature_size'],
self.model_args['lstm_hidden_size'])
self.lstm_readout = units.ReadoutFunction(
'fc', {
'readout_input_size': model_args['lstm_hidden_size'],
'output_classes': model_args['big_attr_classes']
})
self.propagate_layers = model_args['propagate_layers']
self.softmax = torch.nn.Softmax()
self.relu = torch.nn.ReLU()
self.sigmoid = torch.nn.Sigmoid()
self.maxpool = torch.nn.MaxPool2d((2, 2))
if self.load_base_model:
self._load_base_model_weight(self.base_model_weight)
if self.freeze_base_model:
for i, param in self.base_model[0].named_parameters():
param.requires_grad = False
def __init__(self, model_args, args):
super(Atomic_HGNN, self).__init__()
self.model_args = model_args.copy()
self.resnet = units.Resnet.ResNet50(6)
self._load_pretrained_weight(
os.path.join(args.tmp_root, 'checkpoints', 'attmat',
'model_best.pth'))
self.freeze_res_layer(layer_num=9)
num_ftrs = self.resnet.resnet.fc.in_features
self.resnet.resnet.fc = torch.nn.Linear(num_ftrs, 64 * 3)
self.conv3d1 = torch.nn.Conv3d(
in_channels=64 * 3,
out_channels=6,
kernel_size=(5, 6, 1),
stride=1,
padding=0
) #Conv3d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True)
self.conv3d2 = torch.nn.Conv3d(in_channels=6,
out_channels=6,
kernel_size=(1, 1, 1),
stride=1,
padding=0)
self.link_fun = units.LinkFunction('GraphConv', model_args)
self.message_fun = units.MessageFunction('linear_concat_relu',
model_args)
self.update_fun = units.UpdateFunction('gru', model_args)
self.readout_fun = units.ReadoutFunction(
'fc', {
'readout_input_size': model_args['node_feature_size'],
'output_classes': model_args['big_attr_classes']
})
self.lstm = torch.nn.LSTM(model_args['roi_feature_size'],
self.model_args['lstm_hidden_size'])
self.lstm_readout = torch.nn.Linear(
self.model_args['lstm_hidden_size'],
self.model_args['big_attr_classes'])
self.propagate_layers = model_args['propagate_layers']
self.softmax = torch.nn.Softmax()
self.relu = torch.nn.ReLU()
self.sigmoid = torch.nn.Sigmoid()
self.maxpool = torch.nn.MaxPool2d((2, 2))
def __init__(self, model_args):
super(GPNN_HICO, self).__init__()
self.model_args = model_args.copy()
if model_args['resize_feature_to_message_size']:
# Resize large features into message size with Linear functions and xavier normalization
self.edge_feature_resize = torch.nn.Linear(model_args['edge_feature_size'], model_args['message_size'])
self.node_feature_resize = torch.nn.Linear(model_args['node_feature_size'], model_args['message_size'])
torch.nn.init.xavier_normal(self.edge_feature_resize.weight)
torch.nn.init.xavier_normal(self.node_feature_resize.weight)
model_args['edge_feature_size'] = model_args['message_size']
model_args['node_feature_size'] = model_args['message_size']
self.link_fun = units.LinkFunction('GraphConv', model_args) # if needs temporal operation, then 'GraphConvLSTM'
self.sigmoid = torch.nn.Sigmoid()
self.message_fun = units.MessageFunction('linear_concat_relu', model_args) # concatenate messages from all neghbor nodes
self.update_fun = units.UpdateFunction('gru', model_args)
self.readout_fun = units.ReadoutFunction('fc', {'readout_input_size': model_args['node_feature_size'], 'output_classes': model_args['hoi_classes']})
self.propagate_layers = model_args['propagate_layers']
self._load_link_fun(model_args)