def convert_norm_to_float(net):
'''
BatchNorm layers to have parameters in single precision.
Find all layers and convert them back to float. This can't
be done with built in .apply as that function will apply
fn to all modules, parameters, and buffers. Thus we wouldn't
be able to guard the float conversion based on the module type.
'''
if isinstance(net, torch.nn.modules.batchnorm._BatchNorm):
net.float()
for child in net.children():
VoxelNet.convert_norm_to_float(child)
return net
def build(model_cfg: second_pb2.VoxelNet, voxel_generator,
target_assigner) -> VoxelNet:
"""build second pytorch instance.
"""
if not isinstance(model_cfg, second_pb2.VoxelNet):
raise ValueError('model_cfg not of type ' 'second_pb2.VoxelNet.')
vfe_num_filters = list(model_cfg.voxel_feature_extractor.num_filters)
vfe_with_distance = model_cfg.voxel_feature_extractor.with_distance
grid_size = voxel_generator.grid_size
dense_shape = [1] + grid_size[::-1].tolist() + [vfe_num_filters[-1]]
num_class = model_cfg.num_class
num_input_features = model_cfg.num_point_features
if model_cfg.without_reflectivity:
num_input_features = 3
loss_norm_type_dict = {
0: LossNormType.NormByNumExamples,
1: LossNormType.NormByNumPositives,
2: LossNormType.NormByNumPosNeg,
}
loss_norm_type = loss_norm_type_dict[model_cfg.loss_norm_type]
losses = losses_builder.build(model_cfg.loss)
encode_rad_error_by_sin = model_cfg.encode_rad_error_by_sin
cls_loss_ftor, loc_loss_ftor, cls_weight, loc_weight, _ = losses
pos_cls_weight = model_cfg.pos_class_weight
neg_cls_weight = model_cfg.neg_class_weight
direction_loss_weight = model_cfg.direction_loss_weight
net = VoxelNet(
dense_shape,
num_class=num_class,
vfe_class_name=model_cfg.voxel_feature_extractor.module_class_name,
vfe_num_filters=vfe_num_filters,
middle_class_name=model_cfg.middle_feature_extractor.module_class_name,
middle_num_filters_d1=list(
model_cfg.middle_feature_extractor.num_filters_down1),
middle_num_filters_d2=list(
model_cfg.middle_feature_extractor.num_filters_down2),
rpn_class_name=model_cfg.rpn.module_class_name,
rpn_layer_nums=list(model_cfg.rpn.layer_nums),
rpn_layer_strides=list(model_cfg.rpn.layer_strides),
rpn_num_filters=list(model_cfg.rpn.num_filters),
rpn_upsample_strides=list(model_cfg.rpn.upsample_strides),
rpn_num_upsample_filters=list(model_cfg.rpn.num_upsample_filters),
use_norm=True,
use_rotate_nms=model_cfg.use_rotate_nms,
multiclass_nms=model_cfg.use_multi_class_nms,
nms_score_threshold=model_cfg.nms_score_threshold,
nms_pre_max_size=model_cfg.nms_pre_max_size,
nms_post_max_size=model_cfg.nms_post_max_size,
nms_iou_threshold=model_cfg.nms_iou_threshold,
use_sigmoid_score=model_cfg.use_sigmoid_score,
encode_background_as_zeros=model_cfg.encode_background_as_zeros,
use_direction_classifier=model_cfg.use_direction_classifier,
use_bev=model_cfg.use_bev,
num_input_features=num_input_features,
num_groups=model_cfg.rpn.num_groups,
use_groupnorm=model_cfg.rpn.use_groupnorm,
with_distance=vfe_with_distance,
cls_loss_weight=cls_weight,
loc_loss_weight=loc_weight,
pos_cls_weight=pos_cls_weight,
neg_cls_weight=neg_cls_weight,
direction_loss_weight=direction_loss_weight,
loss_norm_type=loss_norm_type,
encode_rad_error_by_sin=encode_rad_error_by_sin,
loc_loss_ftor=loc_loss_ftor,
cls_loss_ftor=cls_loss_ftor,
target_assigner=target_assigner,
voxel_size=voxel_generator.voxel_size,
pc_range=voxel_generator.point_cloud_range)
return net
def build(model_cfg: second_pb2.VoxelNet, fv_generator, #voxel_generator,
save_path=None, RGB_embedding=False) -> VoxelNet:
"""build second pytorch instance.
"""
if not isinstance(model_cfg, second_pb2.VoxelNet):
raise ValueError('model_cfg not of type ' 'second_pb2.VoxelNet.')
vfe_num_filters = list(model_cfg.voxel_feature_extractor.num_filters)
vfe_with_distance = model_cfg.voxel_feature_extractor.with_distance
# grid_size = voxel_generator.grid_size
fv_dim = fv_generator.fv_dim
dense_shape = [1] + fv_dim[::-1].tolist() + [vfe_num_filters[-1]]
num_class = model_cfg.num_class
num_input_features = model_cfg.num_point_features
if model_cfg.without_reflectivity:
num_input_features = 3
loss_norm_type_dict = {
0: LossNormType.NormByNumExamples,
1: LossNormType.NormByNumPositives,
2: LossNormType.NormByNumPosNeg,
}
loss_norm_type = loss_norm_type_dict[model_cfg.loss_norm_type]
losses = losses_builder.build(model_cfg.loss)
encode_rad_error_by_sin = model_cfg.encode_rad_error_by_sin
cls_loss_ftor, loc_loss_ftor, cls_weight, loc_weight, _ = losses
pos_cls_weight = model_cfg.pos_class_weight
neg_cls_weight = model_cfg.neg_class_weight
direction_loss_weight = model_cfg.direction_loss_weight
net = VoxelNet(
dense_shape,
num_class=num_class,
# vfe_class_name=model_cfg.voxel_feature_extractor.module_class_name,
# vfe_num_filters=vfe_num_filters,
# middle_class_name=model_cfg.middle_feature_extractor.module_class_name,
# middle_num_filters_d1=list(
# model_cfg.middle_feature_extractor.num_filters_down1),
# middle_num_filters_d2=list(
# model_cfg.middle_feature_extractor.num_filters_down2),
# rpn_class_name=model_cfg.rpn.module_class_name,
# rpn_layer_nums=list(model_cfg.rpn.layer_nums),
# rpn_layer_strides=list(model_cfg.rpn.layer_strides),
# rpn_num_filters=list(model_cfg.rpn.num_filters),
# rpn_upsample_strides=list(model_cfg.rpn.upsample_strides),
# rpn_num_upsample_filters=list(model_cfg.rpn.num_upsample_filters),
# use_norm=True,
# use_rotate_nms=model_cfg.use_rotate_nms,
# multiclass_nms=model_cfg.use_multi_class_nms,
# nms_score_threshold=model_cfg.nms_score_threshold,
# nms_pre_max_size=model_cfg.nms_pre_max_size,
# nms_post_max_size=model_cfg.nms_post_max_size,
# nms_iou_threshold=model_cfg.nms_iou_threshold,
use_sigmoid_score=model_cfg.use_sigmoid_score,
encode_background_as_zeros=model_cfg.encode_background_as_zeros,
# use_direction_classifier=model_cfg.use_direction_classifier,
# use_bev=model_cfg.use_bev,
# num_input_features=num_input_features,
# num_groups=model_cfg.rpn.num_groups,
# use_groupnorm=model_cfg.rpn.use_groupnorm,
# with_distance=vfe_with_distance,
# cls_loss_weight=cls_weight,
# loc_loss_weight=loc_weight,
# pos_cls_weight=pos_cls_weight,
# neg_cls_weight=neg_cls_weight,
# direction_loss_weight=direction_loss_weight,
# loss_norm_type=loss_norm_type,
# encode_rad_error_by_sin=encode_rad_error_by_sin,
# loc_loss_ftor=loc_loss_ftor,
# cls_loss_ftor=cls_loss_ftor,
# voxel_size=voxel_generator.voxel_size,
# pc_range=voxel_generator.point_cloud_range,
# grid_size=fv_generator.grid_size,
# cartesian_coord_range=fv_generator.cartesian_coord_range,
save_path=save_path,
RGB_embedding=RGB_embedding
)
return net
def build(model_cfg: second_pb2.VoxelNet,
voxel_generator,
target_assigner,
measure_time=False) -> VoxelNet:
"""build second pytorch instance.
"""
if not isinstance(model_cfg, second_pb2.VoxelNet):
raise ValueError('model_cfg not of type ' 'second_pb2.VoxelNet.')
vfe_num_filters = list(model_cfg.voxel_feature_extractor.num_filters)
vfe_with_distance = model_cfg.voxel_feature_extractor.with_distance
grid_size = voxel_generator.grid_size
dense_shape = [1] + grid_size[::-1].tolist() + [vfe_num_filters[-1]]
num_class = len(target_assigner.classes)
class_cfg = model_cfg.target_assigner
use_iou_param_partaa = class_cfg.use_iou_param_partaa
num_input_features = model_cfg.num_point_features
if model_cfg.without_reflectivity:
num_input_features = 3
loss_norm_type_dict = {
0: LossNormType.NormByNumExamples,
1: LossNormType.NormByNumPositives,
2: LossNormType.NormByNumPosNeg,
}
loss_norm_type = loss_norm_type_dict[model_cfg.loss_norm_type]
losses = losses_builder.build(model_cfg.loss)
encode_rad_error_by_sin = model_cfg.encode_rad_error_by_sin
cls_loss_ftor, loc_loss_ftor, iou_loss_ftor, \
cls_weight, loc_weight, iou_weight, _ = losses
pos_cls_weight = model_cfg.pos_class_weight
neg_cls_weight = model_cfg.neg_class_weight
direction_loss_weight = model_cfg.direction_loss_weight
# if model_cfg.use_rpn:
net = VoxelNet(
dense_shape,
num_class=num_class,
vfe_class_name=model_cfg.voxel_feature_extractor.module_class_name,
vfe_num_filters=vfe_num_filters,
middle_class_name=model_cfg.middle_feature_extractor.module_class_name,
middle_num_input_features=model_cfg.middle_feature_extractor.
num_input_features,
middle_num_filters_d1=list(
model_cfg.middle_feature_extractor.num_filters_down1),
middle_num_filters_d2=list(
model_cfg.middle_feature_extractor.num_filters_down2),
rpn_class_name=model_cfg.rpn.module_class_name,
rpn_num_input_features=model_cfg.rpn.num_input_features,
rpn_layer_nums=list(model_cfg.rpn.layer_nums),
rpn_layer_strides=list(model_cfg.rpn.layer_strides),
rpn_num_filters=list(model_cfg.rpn.num_filters),
rpn_upsample_strides=list(model_cfg.rpn.upsample_strides),
rpn_num_upsample_filters=list(model_cfg.rpn.num_upsample_filters),
use_norm=True,
use_voxel_classifier=model_cfg.use_aux_classifier,
use_rotate_nms=model_cfg.use_rotate_nms,
multiclass_nms=model_cfg.use_multi_class_nms,
nms_score_threshold=model_cfg.nms_score_threshold,
nms_pre_max_size=model_cfg.nms_pre_max_size,
nms_post_max_size=model_cfg.nms_post_max_size,
nms_iou_threshold=model_cfg.nms_iou_threshold,
use_sigmoid_score=model_cfg.use_sigmoid_score,
use_sparse_rpn=False,
encode_background_as_zeros=model_cfg.encode_background_as_zeros,
use_direction_classifier=model_cfg.use_direction_classifier,
use_bev=model_cfg.use_bev,
num_input_features=num_input_features,
num_groups=model_cfg.rpn.num_groups,
use_groupnorm=model_cfg.rpn.use_groupnorm,
with_distance=vfe_with_distance,
cls_loss_weight=cls_weight,
loc_loss_weight=loc_weight,
pos_cls_weight=pos_cls_weight,
neg_cls_weight=neg_cls_weight,
direction_loss_weight=direction_loss_weight,
loss_norm_type=loss_norm_type,
encode_rad_error_by_sin=encode_rad_error_by_sin,
loc_loss_ftor=loc_loss_ftor,
cls_loss_ftor=cls_loss_ftor,
target_assigner=target_assigner,
measure_time=measure_time,
use_iou_branch=model_cfg.use_iou_branch,
iou_dict=model_cfg.iou,
use_iou_param_partaa=use_iou_param_partaa,
iou_loss_weight=iou_weight,
iou_loss_ftor=iou_loss_ftor)
# elif model_cfg.use_rcnn:
# net = VoxelNet_Second(
# dense_shape,
# num_class=num_class,
# vfe_class_name=model_cfg.voxel_feature_extractor.module_class_name,
# vfe_num_filters=vfe_num_filters,
# middle_class_name=model_cfg.middle_feature_extractor.module_class_name,
# middle_num_input_features=model_cfg.middle_feature_extractor.num_input_features,
# middle_num_filters_d1=list(
# model_cfg.middle_feature_extractor.num_filters_down1),
# middle_num_filters_d2=list(
# model_cfg.middle_feature_extractor.num_filters_down2),
# rpn_class_name=model_cfg.rpn.module_class_name,
# rpn_num_input_features=model_cfg.rpn.num_input_features,
# rpn_layer_nums=list(model_cfg.rpn.layer_nums),
# rpn_layer_strides=list(model_cfg.rpn.layer_strides),
# rpn_num_filters=list(model_cfg.rpn.num_filters),
# rpn_upsample_strides=list(model_cfg.rpn.upsample_strides),
# rpn_num_upsample_filters=list(model_cfg.rpn.num_upsample_filters),
# use_norm=True,
# use_voxel_classifier=model_cfg.use_aux_classifier,
# use_rotate_nms=model_cfg.use_rotate_nms,
# multiclass_nms=model_cfg.use_multi_class_nms,
# nms_score_threshold=model_cfg.nms_score_threshold,
# nms_pre_max_size=model_cfg.nms_pre_max_size,
# nms_post_max_size=model_cfg.nms_post_max_size,
# nms_iou_threshold=model_cfg.nms_iou_threshold,
# use_sigmoid_score=model_cfg.use_sigmoid_score,
# use_sparse_rpn=False,
# encode_background_as_zeros=model_cfg.encode_background_as_zeros,
# use_direction_classifier=model_cfg.use_direction_classifier,
# use_bev=model_cfg.use_bev,
# num_input_features=num_input_features,
# num_groups=model_cfg.rpn.num_groups,
# use_groupnorm=model_cfg.rpn.use_groupnorm,
# with_distance=vfe_with_distance,
# cls_loss_weight=cls_weight,
# loc_loss_weight=loc_weight,
# pos_cls_weight=pos_cls_weight,
# neg_cls_weight=neg_cls_weight,
# direction_loss_weight=direction_loss_weight,
# loss_norm_type=loss_norm_type,
# encode_rad_error_by_sin=encode_rad_error_by_sin,
# loc_loss_ftor=loc_loss_ftor,
# cls_loss_ftor=cls_loss_ftor,
# target_assigner=target_assigner,
# measure_time=measure_time,
# )
return net