def __init__(self, args):
super(ResNet18Image2IMU, self).__init__()
assert args.sequence_length == 2, "ResNet18Image2IMU supports seq-len=2"
assert args.input_length == 2, "input length not currect"
assert args.output_length == 1, "output length not currect"
self.class_weights = args.dataset.CLASS_WEIGHTS[torch.LongTensor(
args.imus)]
self.output_length = args.output_length
self.lr = args.lrm
self.step_size = args.step_size
self.decay = args.weight_decay
resnet_model = torchvision_resnet18(pretrained=args.pretrain)
# Remove the last fully connected layer.
del resnet_model.fc
self.resnet = resnet_model
num_features = 512
num_frames = 2
num_classes = args.num_classes
# Make num_imu fc layers
self.imus = args.imus
for i in self.imus:
setattr(self, 'imu{}'.format(i),
nn.Linear(num_frames * num_features, num_classes))
self.dropout = nn.Dropout(p=args.dropout_ratio)
def __init__(self, args):
super(ResNet18Image2IMUOneTowerPrediction, self).__init__()
self.class_weights = args.dataset.CLASS_WEIGHTS[torch.LongTensor(
args.imus)]
resnet_model = torchvision_resnet18(pretrained=args.pretrain)
# Remove the last fully connected layer.
del resnet_model.fc
self.resnet = resnet_model
num_features = 512
num_frames = args.input_length
self.args = args
self.resnet.conv1 = nn.Conv2d(num_frames * 3,
64,
kernel_size=(7, 7),
stride=(2, 2),
padding=(3, 3),
bias=False)
num_classes = args.num_classes
# Make num_imu fc layers
self.imus = args.imus
for i in self.imus:
setattr(self, 'imu{}'.format(i),
nn.Linear(num_features, num_classes))
self.dropout = nn.Dropout()
def __init__(self, args):
super(FullyConvolutional, self).__init__()
self.class_weights = args.dataset.CLASS_WEIGHTS
self.args = args
self.lr = args.lrm
self.decay = args.weight_decay
self.step_size = args.step_size
resnet_model = torchvision_resnet18(pretrained=args.pretrain)
# Remove the last fully connected layer.
del resnet_model.fc
self.resnet = resnet_model
num_features = 512
num_frames = 2
self.imus = args.imus
self.relu = nn.ReLU()
self.dropout = nn.Dropout(p=args.dropout_ratio)
self.conv4 = conv(512, 256, stride=2, transposed=True, kernel_size=3)
self.bn4 = nn.BatchNorm2d(256)
self.conv3 = conv(256, 128, stride=2, transposed=True, kernel_size=3)
self.bn3 = nn.BatchNorm2d(128)
self.conv2 = conv(128, 64, stride=2, transposed=True, kernel_size=3)
self.bn2 = nn.BatchNorm2d(64)
self.conv1 = conv(64, 1, kernel_size=3)
def __init__(self, args):
super(ResNetSceneCategorization, self).__init__()
self.class_weights = args.dataset.CLASS_WEIGHTS
self.args = args
self.dropout = nn.Dropout()
resnet_model = torchvision_resnet18(pretrained=args.pretrain)
# Remove the last fully connected layer.
del resnet_model.fc
self.resnet = resnet_model
self.fc = nn.Linear(512, 397)
def __init__(self, args):
super(FeatureLearnerModule, self).__init__()
self.base_lr = args.base_lr
self.lrm = args.lrm
self.read_features = args.read_features
self.number_of_trained_resnet_blocks = args.number_of_trained_resnet_blocks
resnet_model = torchvision_resnet18(pretrained=args.pretrain)
del resnet_model.fc
self.resnet = resnet_model
self.detach_level = args.detach_level
self.fixed_feature_weights = args.fixed_feature_weights
self.intermediate_features = None
imagenet_feature_testing = args.pretrain and self.fixed_feature_weights and 'imagenet' in args.title and self.detach_level == 0
imagenet_feature_training = args.pretrain and 'imagenet_train_all_the_way' in args.title
assert imagenet_feature_testing or (
not args.pretrain) or imagenet_feature_training
def __init__(self, args):
super(ResNet18Image2IMURegress, self).__init__()
assert args.sequence_length == 1, "ResNet18Image2IMURegress supports seq-len=1"
assert args.regression, "--regression must be true"
self.class_weights = args.dataset.CLASS_WEIGHTS[torch.LongTensor(
args.imus)]
resnet_model = torchvision_resnet18(pretrained=args.pretrain)
# Remove the last fully connected layer.
del resnet_model.fc
self.resnet = resnet_model
num_features = 512
num_frames = 2
num_output = 4
# Make num_imu fc layers
self.imus = args.imus
for i in self.imus:
setattr(self, 'imu{}'.format(i),
nn.Linear(num_frames * num_features, num_output))