def __init__(self,
img_emb_size,
action_size,
args,
use_bb_in_input=True,
n_classes=0,
use_spatial_softmax=True):
super(BoundingBoxOnlyModel, self).__init__()
# 6 for other oject pose (6D) and 12 for both anchor and other bb.
bb_size = 6 + 12 if args.use_bb_in_input else 0
assert bb_size > 0
self.bb_action_model = build_mlp(
[bb_size + action_size, 512, 256, 64],
activation='relu',
)
use_regression = (args.loss_type == 'regr')
if use_regression:
self.delta_pose_model = build_mlp(
[64, 6],
activation='relu',
final_nonlinearity=False,
)
print("Using regression model.")
else:
assert n_classes > 0, "Invalid number of classes."
print("Using classif model with {} classes".format(n_classes))
self.delta_pose_model = build_mlp([64, n_classes],
activation='relu',
final_nonlinearity=False)
def __init__(self,
img_emb_size,
action_size,
args,
use_bb_in_input=True,
n_classes=0,
use_spatial_softmax=True):
super(UnscaledVoxelModel, self).__init__()
self.img_emb_size = img_emb_size
self.action_size = action_size
self.use_spatial_softmax = use_spatial_softmax
self.args = args
if args.add_xy_channels == 0:
input_nc = 3
elif args.add_xy_channels == 1:
input_nc = 6
else:
raise ValueError("Invalid add_xy_channels")
if img_emb_size <= 36:
self.voxels_to_rel_emb_model = CML_3(input_nc)
else:
self.voxels_to_rel_emb_model = CML_2(input_nc)
if use_spatial_softmax:
z_ss_size = 96
self.spatial_softmax_encoder = Spatial3DSoftmaxImageEncoder(
input_nc, z_ss_size)
else:
z_ss_size = 0
# 6 for other oject pose (6D) and 12 for both anchor and other bb.
bb_size = 6 + 12 if args.use_bb_in_input else 0
self.img_action_model = build_mlp(
[
img_emb_size + bb_size + action_size + z_ss_size,
img_emb_size // 2, img_emb_size // 2, img_emb_size // 4
],
activation='relu',
)
use_regression = (args.loss_type == 'regr')
if use_regression:
output_size = 6
self.delta_pose_model = build_mlp(
[img_emb_size // 4, 64, output_size],
activation='relu',
final_nonlinearity=False,
)
print("Using regression model.")
else:
assert n_classes > 0, "Invalid number of classes."
print("Using classif model with {} classes".format(n_classes))
self.delta_pose_model = build_mlp(
[img_emb_size // 4, 64, n_classes],
activation='relu',
final_nonlinearity=False)
def __init__(self,
resnet_klass,
img_emb_size,
action_size,
args,
use_bb_in_input=True,
n_classes=0,
use_spatial_softmax=True):
super(UnscaledResNetVoxelModel, self).__init__()
self.img_emb_size = img_emb_size
self.action_size = action_size
self.use_spatial_softmax = use_spatial_softmax
self.args = args
if args.add_xy_channels == 0:
input_nc = 3
elif args.add_xy_channels == 1:
input_nc = 6
else:
raise ValueError("Invalid add_xy_channels")
self.resnet = resnet_klass(emb_size=img_emb_size,
input_channels=input_nc)
z_ss_size = 0
bb_size = 6 + 12 if args.use_bb_in_input else 0
self.img_action_model = build_mlp(
[
img_emb_size + bb_size + action_size + z_ss_size,
img_emb_size // 2, img_emb_size // 4
],
activation='relu',
final_nonlinearity=True,
)
use_regression = (args.loss_type == 'regr')
if use_regression:
output_size = 6
self.delta_pose_model = build_mlp(
[img_emb_size // 4, output_size],
activation='relu',
final_nonlinearity=False,
)
print("Using regression model.")
else:
assert n_classes > 0, "Invalid number of classes."
print("Using classif model with {} classes".format(n_classes))
self.delta_pose_model = build_mlp([img_emb_size // 4, n_classes],
activation='relu',
final_nonlinearity=False)
def __init__(self,
img_emb_size,
args,
use_bb_in_input=True,
n_classes=0,
use_spatial_softmax=True):
super(UnscaledPrecondVoxelModel, self).__init__()
self.img_emb_size = img_emb_size
self.use_spatial_softmax = use_spatial_softmax
self.args = args
if args.add_xy_channels == 0:
input_nc = 3
elif args.add_xy_channels == 1:
input_nc = 6
else:
raise ValueError("Invalid add_xy_channels")
self.voxels_to_rel_emb_model = CML_2(input_nc)
if use_spatial_softmax:
z_ss_size = 96
self.spatial_softmax_encoder = Spatial3DSoftmaxImageEncoder(
input_nc, z_ss_size)
else:
z_ss_size = 0
# 6 for other oject pose (6D) and 12 for both anchor and other bb.
bb_size = 12 if args.use_bb_in_input else 0
self.img_emb_pred_model = build_mlp(
[img_emb_size + bb_size + z_ss_size, 256, 64, 1],
activation='relu',
final_nonlinearity=False)
def __init__(self, img_emb_size, action_size, args, use_bb_in_input=True):
super(SimpleModel, self).__init__()
self.img_emb_size = img_emb_size
self.action_size = action_size
self.args = args
if args.add_xy_channels == 0:
input_nc = 3
elif args.add_xy_channels in (1, 2):
input_nc = 5
self.img_to_rel_emb_model = ConvImageEncoder(input_nc,
img_emb_size,
inp_size=256)
z_ss_size = 128
self.spatial_softmax_encoder = SpatialSoftmaxImageEncoder(
input_nc, z_ss_size)
# self.img_to_rel_emb_model = ResnetGenerator_Encoder(
# input_nc=input_nc,
# output_nc=3,
# n_blocks=1,
# ngf=2,
# n_downsampling=8,
# final_channels=img_emb_size//2)
bb_size = 8 if args.use_bb_in_input else 0
self.img_action_model = build_mlp(
[
img_emb_size + bb_size + action_size + z_ss_size, 256, 256,
img_emb_size
],
activation='relu',
final_nonlinearity=True,
)
self.delta_pose_model = build_mlp(
[img_emb_size, 128, 4],
activation='relu',
final_nonlinearity=False,
)
self.inv_model = build_mlp(
[img_emb_size + 4, 256, 256, img_emb_size + z_ss_size])
def __init__(self,
img_emb_size,
action_size,
args,
use_bb_in_input=True,
n_classes=0,
use_spatial_softmax=True):
super(SmallEmbUnscaledVoxelModel, self).__init__()
self.img_emb_size = img_emb_size
self.action_size = action_size
self.use_spatial_softmax = use_spatial_softmax
self.args = args
if args.add_xy_channels == 0:
input_nc = 3
elif args.add_xy_channels == 1:
input_nc = 6
else:
raise ValueError("Invalid add_xy_channels")
# Why do we need an extra +3 ?
# self.voxels_to_rel_emb_model = CML_4(input_nc, img_emb_size+3, out_emb=None)
self.voxels_to_rel_emb_model = CML_4(input_nc,
img_emb_size,
out_emb=None)
if use_spatial_softmax:
z_ss_size = 96
self.spatial_softmax_encoder = Spatial3DSoftmaxImageEncoder(
input_nc, z_ss_size)
else:
z_ss_size = 0
# 6 for other oject pose (6D) and 12 for both anchor and other bb.
bb_size = 6 + 12 if args.use_bb_in_input else 0
delta_pose_model_inp_size = img_emb_size \
+ bb_size \
+ action_size \
+ z_ss_size
use_regression = (args.loss_type == 'regr')
delta_pose_model_out_size = 6 if use_regression else n_classes
if (delta_pose_model_inp_size // 4) >= 32:
delta_pose_model_layers = [
delta_pose_model_inp_size, delta_pose_model_inp_size // 2,
delta_pose_model_inp_size // 4, delta_pose_model_out_size
]
else:
delta_pose_model_layers = [
delta_pose_model_inp_size, delta_pose_model_inp_size // 2,
delta_pose_model_out_size
]
if use_regression:
output_size = 6
self.delta_pose_model = build_mlp(
delta_pose_model_layers,
activation='relu',
final_nonlinearity=False,
)
print("Using regression model.")
else:
assert n_classes > 0, "Invalid number of classes."
print("Using classif model with {} classes".format(n_classes))
self.delta_pose_model = build_mlp(delta_pose_model_layers,
activation='relu',
final_nonlinearity=False)
def __init__(self,
img_emb_size,
action_size,
args,
use_bb_in_input=True,
use_spatial_softmax=True):
super(UnscaledClassifVoxelModel, self).__init__()
self.img_emb_size = img_emb_size
self.action_size = action_size
self.use_spatial_softmax = use_spatial_softmax
self.args = args
if args.add_xy_channels == 0:
input_nc = 3
elif args.add_xy_channels == 1:
input_nc = 6
else:
raise ValueError("Invalid add_xy_channels")
self.voxels_to_rel_emb_model = CML_2(input_nc)
if use_spatial_softmax:
z_ss_size = 96
self.spatial_softmax_encoder = Spatial3DSoftmaxImageEncoder(
input_nc, z_ss_size)
else:
z_ss_size = 0
# 6 for other oject pose (6D) and 12 for both anchor and other bb.
bb_size = 6 + 12 if args.use_bb_in_input else 0
self.img_action_model = build_mlp(
[
img_emb_size + bb_size + action_size + z_ss_size,
img_emb_size // 2, img_emb_size // 2, img_emb_size // 4
],
activation='relu',
)
if args.pos_loss_type == 'regr':
output_size = 3
self.delta_pose_model = build_mlp(
[img_emb_size // 4, 64, output_size],
activation='relu',
final_nonlinearity=False,
)
print("Using regression model.")
else:
n_classes = args.pos_classif_num_classes
assert n_classes > 0
self.delta_pos_x = build_mlp([img_emb_size // 4, 64, n_classes],
activation='relu',
final_nonlinearity=False)
self.delta_pos_y = build_mlp([img_emb_size // 4, 64, n_classes],
activation='relu',
final_nonlinearity=False)
self.delta_pos_z = build_mlp([img_emb_size // 4, 64, n_classes],
activation='relu',
final_nonlinearity=False)
if args.orient_loss_type == 'regr':
output_size = 3
self.delta_orient_model = build_mlp(
[img_emb_size // 4, 64, output_size],
activation='relu',
final_nonlinearity=False,
)
print("Using regression model.")
else:
n_classes = args.orient_classif_num_classes
assert n_classes > 0
self.delta_orient_x = build_mlp([img_emb_size // 4, 64, n_classes],
activation='relu',
final_nonlinearity=False)
self.delta_orient_y = build_mlp([img_emb_size // 4, 64, n_classes],
activation='relu',
final_nonlinearity=False)
self.delta_orient_z = build_mlp([img_emb_size // 4, 64, n_classes],
activation='relu',
final_nonlinearity=False)
def __init__(self,
resnet_klass,
img_emb_size,
action_size,
args,
use_bb_in_input=True,
use_spatial_softmax=True):
super(UnscaledClassifResNetVoxelModel, self).__init__()
self.img_emb_size = img_emb_size
self.action_size = action_size
self.use_spatial_softmax = use_spatial_softmax
self.args = args
if args.add_xy_channels == 0:
input_nc = 3
elif args.add_xy_channels == 1:
input_nc = 6
else:
raise ValueError("Invalid add_xy_channels")
self.resnet = resnet_klass(emb_size=img_emb_size,
input_channels=input_nc)
z_ss_size = 0
bb_size = 6 + 12 if args.use_bb_in_input else 0
self.img_action_model = build_mlp(
[
img_emb_size + bb_size + action_size + z_ss_size,
img_emb_size // 2, img_emb_size // 4
],
activation='relu',
final_nonlinearity=True,
)
if args.pos_loss_type == 'regr':
output_size = 3
self.delta_pose_model = build_mlp(
[img_emb_size // 4, 64, output_size],
activation='relu',
final_nonlinearity=False,
)
print("Using regression model.")
else:
n_classes = args.pos_classif_num_classes
assert n_classes > 0
self.delta_pos_x = build_mlp([img_emb_size // 4, 64, n_classes],
activation='relu',
final_nonlinearity=False)
self.delta_pos_y = build_mlp([img_emb_size // 4, 64, n_classes],
activation='relu',
final_nonlinearity=False)
self.delta_pos_z = build_mlp([img_emb_size // 4, 64, n_classes],
activation='relu',
final_nonlinearity=False)
if args.orient_loss_type == 'regr':
output_size = 3
self.delta_orient_model = build_mlp(
[img_emb_size // 4, 64, output_size],
activation='relu',
final_nonlinearity=False,
)
print("Using regression model.")
else:
n_classes = args.orient_classif_num_classes
assert n_classes > 0
self.delta_orient_x = build_mlp([img_emb_size // 4, 64, n_classes],
activation='relu',
final_nonlinearity=False)
self.delta_orient_y = build_mlp([img_emb_size // 4, 64, n_classes],
activation='relu',
final_nonlinearity=False)
self.delta_orient_z = build_mlp([img_emb_size // 4, 64, n_classes],
activation='relu',
final_nonlinearity=False)