ArseniuML

KeyError 'model_state' while loading checkpoint

1

I am trying to start inference on Argoverse2 dataset via the next command: python test.py --cfg_file cfgs/argo2_models/cbgs_voxel01_voxelnext.yam --ckpt ../../VoxelNeXt_Argo2.pth --set DATA_CONFIG.DATA_SPLIT.test val However, there is an error while loading checkpoint:...

I am trying to launch VoxelNeXt, but it is silently dying at VoxelGenerator creation: ``` class VoxelGeneratorWrapper(): def __init__(self, vsize_xyz, coors_range_xyz, num_point_features, max_num_points_per_voxel, max_num_voxels): try: from spconv.utils import VoxelGeneratorV2 as...

LIDAR data seems not to be consistent

1

I tried to visualize vertical angle distribution of LIDAR points and expect this histogram to be discrete and have 64 peaks, corresponding to 64 LIDAR beams. However, this histogram seems...

Incremental/additional training of FSDv2

4

Is there a way to apply additional training on our data to pre-trained checkpoint (say on Argoverse2)? If there is a simple CNN, I can fix conv layers weights and...

Correctness of FSDv2 performance on Argoverse2

1

I run dist_test.sh with Argo2 config and obtained the same metrics that were reported in article (AP and others). However, it seems that test.py loads files from train split: ```...

As I can understand, LIDAR frames are located in segment/dynamic/raw-revolutions folders, and 3d annotations are located in segment/dynamic/box/3d_body folders. But it seems that number of files in that folders are...

Classes in train and val splits

1

I tried to investigate, object of what types are annotated in train and val splits: ``` train_dataset = AiMotiveDataset(data_root, 'train') train_classes = set() for anno in train_dataset.dataset_index: with open(anno, 'r')...

In file fsdv2_head.py: ``` # Add a dummy background class to the front when using sigmoid padding = scores.new_zeros(scores.shape[0], 1) scores = torch.cat([scores, padding], dim=1) score_thr = cfg.get('score_thr', 0) results...