Closed LinfengUP closed 1 year ago
I try to train scannet v2 from scratch. I follow the setting in configs by copying trianset 4 times and set prepare_epoch to 20. After 128 epochs of training, the mAP is only about 30. Can anyone share training scannet from scratch config?
Here is my config
model: prepare_epoch: 20 channels: 32 num_blocks: 7 semantic_classes: 20 instance_classes: 18 sem2ins_classes: [] semantic_only: False ignore_label: -100 grouping_cfg: score_thr: 0.2 radius: 0.04 mean_active: 300 class_numpoint_mean: [-1., -1., 3917., 12056., 2303., 8331., 3948., 3166., 5629., 11719., 1003., 3317., 4912., 10221., 3889., 4136., 2120., 945., 3967., 2589.] npoint_thr: 0.05 # absolute if class_numpoint == -1, relative if class_numpoint != -1 ignore_classes: [0, 1] instance_voxel_cfg: scale: 50 spatial_shape: 20 train_cfg: max_proposal_num: 200 pos_iou_thr: 0.5 test_cfg: x4_split: False cls_score_thr: 0.001 mask_score_thr: -0.5 min_npoint: 100 eval_tasks: ['semantic', 'instance'] # fixed_modules: ['input_conv', 'unet', 'output_layer', 'semantic_linear', 'offset_linear'] data: train: type: 'scannetv2' data_root: 'dataset/scannetv2' prefix: 'train' suffix: '_inst_nostuff.pth' training: True repeat: 4 voxel_cfg: scale: 50 spatial_shape: [128, 512] max_npoint: 250000 min_npoint: 5000 test: type: 'scannetv2' data_root: 'dataset/scannetv2' prefix: 'val' suffix: '_inst_nostuff.pth' training: False with_label: True voxel_cfg: scale: 50 spatial_shape: [128, 512] max_npoint: 250000 min_npoint: 5000 dataloader: train: batch_size: 4 num_workers: 4 test: batch_size: 1 num_workers: 1 optimizer: type: 'Adam' lr: 0.004 fp16: False epochs: 128 step_epoch: 128 save_freq: 4 pretrain: '' work_dir: ''
I try to train scannet v2 from scratch. I follow the setting in configs by copying trianset 4 times and set prepare_epoch to 20. After 128 epochs of training, the mAP is only about 30. Can anyone share training scannet from scratch config?
Here is my config