ShapeNetPart mIoU

[TeaWhiteBro]

Thank you for your wonderful work! After you released your config file, I ran the code using the config, and the final results ( Instance mIoU 85.62 ) were significantly lower than the results in the paper. At the same time, I used the pre-trained model you provided to run the test code, and the results ( Instance mIoU 86.10 ) were also slightly lower than those in the paper. I'm wondering if there's something wrong with my configuration? Here is my running log using pre-trained model you provided: launch mp with 1 GPUs, current rank: 0 [12/03 17:04:48 ShapeNetPartNormal]: dist_url: tcp://localhost:8888 dist_backend: nccl multiprocessing_distributed: False ngpus_per_node: 1 world_size: 1 launcher: mp local_rank: 0 use_gpu: True seed: 1500 epoch: 0 epochs: 150 ignore_index: None val_fn: validate deterministic: False sync_bn: False criterion_args: NAME: Poly1FocalLoss use_mask: False grad_norm_clip: 1 layer_decay: 0 step_per_update: 1 start_epoch: 1 sched_on_epoch: True wandb: use_wandb: False project: PointNext-ShapeNetPart tags: ['test'] name: ckpt_best.pth_20231203-170448-7Lbijd3qiHGNd5RtwULiKo use_amp: False use_voting: False val_freq: 1 resume: False test: False finetune: False mode: test logname: None load_path: None print_freq: 10 save_freq: -1 root_dir: log/shapenetpart pretrained_path: ckpt/ShapeNetPart/ckpt_best.pth datatransforms: train: ['PointsToTensor', 'PointCloudScaling', 'PointCloudCenterAndNormalize', 'PointCloudJitter', 'ChromaticDropGPU'] val: ['PointsToTensor', 'PointCloudCenterAndNormalize'] vote: ['PointCloudScaling'] kwargs: jitter_sigma: 0.001 jitter_clip: 0.005 scale: [0.8, 1.2] gravity_dim: 1 angle: [0, 1.0, 0] feature_keys: pos,x,heights dataset: common: NAME: ShapeNetPartNormal data_root: ../data/shapenetcore_partanno_segmentation_benchmark_v0_normal use_normal: True num_points: 2048 train: split: trainval val: split: test presample: True num_classes: 50 shape_classes: 16 num_points: 2048 normal_channel: True batch_size: 8 dataloader: num_workers: 6 num_votes: 10 refine: True lr: 0.001 min_lr: None optimizer: NAME: adamw weight_decay: 0.0001 sched: multistep decay_epochs: [90, 120] decay_rate: 0.5 warmup_epochs: 0 model: NAME: BasePartSeg encoder_args: NAME: SPoTrEncoder blocks: [1, 1, 1, 1, 1] strides: [1, 2, 2, 2, 2] width: 128 in_channels: 7 sa_layers: 3 sa_use_res: True num_layers: 3 expansion: 4 radius: 0.1 radius_scaling: 2.5 nsample: 32 gamma: 16 num_gp: 16 tau_delta: 0.1 aggr_args: feature_type: dp_df reduction: max group_args: NAME: ballquery normalize_dp: True conv_args: order: conv-norm-act act_args: act: relu norm_args: norm: bn decoder_args: NAME: SPoTrPartDecoder cls_args: NAME: SegHead globals: max,avg num_classes: 50 in_channels: None norm_args: norm: bn rank: 0 distributed: False mp: False task_name: shapenetpart cfg_basename: spotr is_training: False run_name: ckpt_best.pth_20231203-170448-7Lbijd3qiHGNd5RtwULiKo run_dir: log/shapenetpart/ckpt_best.pth_20231203-170448-7Lbijd3qiHGNd5RtwULiKo log_dir: log/shapenetpart/ckpt_best.pth_20231203-170448-7Lbijd3qiHGNd5RtwULiKo ckpt_dir: log/shapenetpart/ckpt_best.pth_20231203-170448-7Lbijd3qiHGNd5RtwULiKo/checkpoint code_dir: log/shapenetpart/ckpt_best.pth_20231203-170448-7Lbijd3qiHGNd5RtwULiKo/code log_path: log/shapenetpart/ckpt_best.pth_20231203-170448-7Lbijd3qiHGNd5RtwULiKo/ckpt_best.pth_20231203-170448-7Lbijd3qiHGNd5RtwULiKo.log cfg_path: log/shapenetpart/ckpt_best.pth_20231203-170448-7Lbijd3qiHGNd5RtwULiKo/cfg.yaml ../data/shapenetcore_partanno_segmentation_benchmark_v0_normal/processed/test_2048_fps.pkl load successfully

[BuLingBin]

I obtained the same result on the pre-trained model.

[PJin0]

Thank you for the interest to our paper.

We fixed the error of the code and you can get the correct result with the updated version.