def __init__(self, pred_cls_channel, pred_reg_base_num,
pred_reg_channel_num, is_training, pred_attr_velo, bn,
pre_channel):
super().__init__()
self.pred_cls_channel = pred_cls_channel
self.pred_reg_base_num = pred_reg_base_num
self.pred_reg_channel_num = pred_reg_channel_num
self.is_training = is_training
self.pred_attr_velo = pred_attr_velo
self.pre_channel = pre_channel
self.bn = bn
cls_layers = []
cls_layers.append(pt_utils.Conv1d(self.pre_channel, 128, bn=self.bn))
cls_layers.append(
pt_utils.Conv1d(128,
self.pred_cls_channel,
bn=False,
activation=None))
self.cls_layers = nn.Sequential(*cls_layers)
self.angle_cls_num = cfg.MODEL.ANGLE_CLS_NUM
reg_layers = []
reg_layers.append(pt_utils.Conv1d(pre_channel, 128, bn=self.bn))
reg_channel = pred_reg_base_num * (pred_reg_channel_num +
self.angle_cls_num * 2)
reg_layers.append(
pt_utils.Conv1d(128, reg_channel, bn=False, activation=None))
self.reg_layers = nn.Sequential(*reg_layers)
def __init__(self, batch_size, anchor_num, head_idx, head_cfg, is_training):
super().__init__()
self.is_training = is_training
self.head_idx = head_idx
self.anchor_num = anchor_num
self.batch_size = batch_size
cur_head = head_cfg
self.xyz_index = cur_head[0]
self.feature_index = cur_head[1]
self.op_type = cur_head[2]
self.mlp_list = cur_head[3]
self.bn = cur_head[4]
self.layer_type = cur_head[5]
self.scope = cur_head[6]
if head_idx == 0: # stage 1
self.head_cfg = cfg.MODEL.FIRST_STAGE
elif head_idx == 1: # stage 2
self.head_cfg = cfg.MODEL.SECOND_STAGE
else: raise Exception('Not Implementation Error!!!') # stage 3
# determine channel number
if self.head_cfg.CLS_ACTIVATION == 'Sigmoid':
self.pred_cls_channel = self.anchor_num
elif self.head_cfg.CLS_ACTIVATION == 'Softmax':
self.pred_cls_channel = self.anchor_num + 1
if self.layer_type == 'IoU':
self.pred_cls_channel = self.anchor_num
self.reg_method = self.head_cfg.REGRESSION_METHOD.TYPE
anchor_type = self.reg_method.split('-')[-1] # Anchor & free
pred_reg_base_num = {
'Anchor': self.anchor_num,
'free': 1,
}
self.pred_reg_base_num = pred_reg_base_num[anchor_type]
pred_reg_channel_num = {
'Dist-Anchor': 6,
'Log-Anchor': 6,
'Dist-Anchor-free': 6,
# bin_x/res_x/bin_z/res_z/res_y/res_size
'Bin-Anchor': self.head_cfg.REGRESSION_METHOD.BIN_CLASS_NUM * 4 + 4,
}
self.pred_reg_channel_num = pred_reg_channel_num[self.reg_method]
self.pre_channel = cfg.MODEL.NETWORK.FIRST_STAGE.ARCHITECTURE[self.feature_index[0]-1][15]
layer_modules = []
pre_channel = self.pre_channel
for i in range(len(self.mlp_list)):
layer_modules.append(pt_utils.Conv1d(pre_channel, self.mlp_list[i], bn=self.bn))
pre_channel = self.mlp_list[i]
self.layer_modules = nn.Sequential(*layer_modules)
self.head_predictor = BoxRegressionHead(self.pred_cls_channel, self.pred_reg_base_num, self.pred_reg_channel_num,
self.is_training, self.head_cfg.PREDICT_ATTRIBUTE_AND_VELOCITY,
self.bn, pre_channel)
def __init__(self, mlp_list, bn, is_training, pre_channel):
super().__init__()
self.mlp_list = mlp_list
self.bn = bn
self.is_training = is_training
mlp_modules = []
for i in range(len(self.mlp_list)):
mlp_modules.append(
pt_utils.Conv1d(pre_channel, self.mlp_list[i], bn=self.bn))
pre_channel = self.mlp_list[i]
self.mlp_modules = nn.Sequential(*mlp_modules)
self.ctr_reg = pt_utils.Conv1d(pre_channel,
3,
activation=None,
bn=False)
self.min_offset = torch.tensor(
cfg.MODEL.MAX_TRANSLATE_RANGE).float().view(1, 1, 3)
def __init__(self,
radius_list,
nsample_list,
mlp_list,
is_training,
bn_decay,
bn,
fps_sample_range_list,
fps_method_list,
npoint_list,
use_attention,
scope,
dilated_group,
aggregation_channel=None,
pre_channel=0,
debugging=False,
epsilon=1e-5):
super().__init__()
self.radius_list = radius_list
self.nsample_list = nsample_list
self.mlp_list = mlp_list
self.is_training = is_training
self.bn_decay = bn_decay
self.bn = bn
self.fps_sample_range_list = fps_sample_range_list
self.fps_method_list = fps_method_list
self.npoint_list = npoint_list
self.use_attention = use_attention
self.scope = scope
self.dilated_group = dilated_group
self.aggregation_channel = aggregation_channel
self.pre_channel = pre_channel
mlp_modules = []
for i in range(len(self.radius_list)):
mlp_spec = [self.pre_channel + 3] + self.mlp_list[i]
mlp_modules.append(pt_utils.SharedMLP(mlp_spec, bn=self.bn))
self.mlp_modules = nn.Sequential(*mlp_modules)
if cfg.MODEL.NETWORK.AGGREGATION_SA_FEATURE and (len(self.mlp_list) !=
0):
input_channel = 0
for mlp_tmp in self.mlp_list:
input_channel += mlp_tmp[-1]
self.aggregation_layer = pt_utils.Conv1d(input_channel,
aggregation_channel,
bn=self.bn)