Well trained weights on nuscenes

[hehualin-tut]

Well trained weights on nuscenes

[edwardzhou130]

Hello, thanks for your interest in our work. Sorry, I currently do not have access to my previous pretrained weight. If you need to lower the memory cost, you can try changing the following parameters:

1. batch size -> 1 or 2

   centerformer/configs/nusc/nuscenes_centerformer_separate_detection_head.py

   Line 188 in 5a949b8

   samples_per_gpu=4,

2. channel size

   centerformer/configs/nusc/nuscenes_centerformer_separate_detection_head.py

   Lines 34 to 69 in 5a949b8

   	backbone=dict(
   	type="SpMiddleResNetFHD", num_input_features=5, ds_factor=8),
   	neck=dict(
   	type="RPN_transformer_multitask",
   	layer_nums=[5, 5, 1],
   	ds_num_filters=[256, 256, 128],
   	num_input_features=256,
   	tasks=tasks,
   	use_gt_training=True,
   	corner = True,
   	obj_num= 500,
   	assign_label_window_size=window_size,
   	transformer_config=dict(
   	depth = 3,
   	heads = 4,
   	dim_head = 64,
   	MLP_dim = 256,
   	DP_rate=0.3,
   	out_att = False,
   	cross_attention_kernel_size = [3,3,3]
   	),
   	logger=logging.getLogger("RPN"),
   	),
   	bbox_head=dict(
   	type="CenterHeadIoU_1d",
   	in_channels=256,
   	tasks=tasks,
   	dataset='nuscenes',
   	weight=0.25,
   	assign_label_window_size=window_size,
   	corner_loss=True,
   	iou_loss=False,
   	code_weights=[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.2, 0.2, 1.0, 1.0],
   	common_heads={'reg': (2, 2), 'height': (1, 2), 'dim':(3, 2), 'rot':(2, 2), 'vel': (2, 2)},
   	),
   	)

3. voxel size -> [0.1, 0.1, 0.2]

   centerformer/configs/nusc/nuscenes_centerformer_separate_detection_head.py

   Line 32 in 5a949b8

   voxel_size=[0.075, 0.075, 0.2],