vote_features = self.pv_transformer(
src=torch.cat([key_features.permute(1, 0, 2), voxel_key_features.permute(1, 0, 2)], dim=0),
tgt=vote_candidate_features.unsqueeze(1).permute(1, 0, 2),
pos_res=torch.cat([key_pos_emb.permute(1, 0, 2).unsqueeze(0), voxel_key_pos_emb.permute(1, 0, 2).unsqueeze(0)], dim=1)
).squeeze(0)

About transformer module, why you choose vote_candidate_features as Q, and torch.cat([key_features.permute(1, 0, 2), voxel_key_features.permute(1, 0, 2)], dim=0) as K ?
@Nightmare-n

The environment in readme include the waymo-open-dataset-tf-2-2-0 and cuda 11.1 .But tensorflow 2.2.0 requires cuda 10.1 so that i can't extract point cloud data from tfrecord

I was trying to build the network for kitti, and I noticed that the yaml file only contains 'Car', so I changed it to the three classes, but it began to show that the target classes' dim is not 3. I did not find any other information about how to get the correct structure.

Hi there
Thanks for your work first.
My question is for velodyne_reduced, how can I generate it? (64 lines in kitti_dataset.py)
Or it's the same thing as Velodyne?

Hi! Thank you for sharing this interesting work.
I would like to know how you generated the visualization Figure 6, can you share the code?

Hi, very good work! Fast in inference and very light-weighted.

May I know how the training went on your machine? I had finished training, but according to training logs, the loss basically stays the same. Considering I made a few changes in the code, I am trying using completely original code to train. The loss starts to decrease, but very slowly.

I did not change much of the code, I made two major changes:

1. When processing the waymo dataset using your code, it raised an error, because there is no file_client._map_path. I changed the code to:

def process_single_sequence(sequence_file, save_path, sampled_interval, client, has_label=True, use_two_returns=True):
    sequence_name = os.path.splitext(os.path.basename(sequence_file))[0]

    # print('Load record (sampled_interval=%d): %s' % (sampled_interval, sequence_name))
    if not client.exists(sequence_file):
        print('NotFoundError: %s' % sequence_file)
        return []

    # dataset = tf.data.TFRecordDataset(client._map_path(sequence_file), compression_type='')
    dataset = tf.data.TFRecordDataset(str(sequence_file), compression_type='')
    cur_save_dir = save_path / sequence_name
    cur_save_dir.mkdir(parents=True, exist_ok=True)
........

2. When I did the first training, I changed the dist_train.sh to OpenPCDet format, which is:

#!/usr/bin/env bash
set -x
NGPUS=$1
PY_ARGS=${@:2}

echo "#######################################" $PY_ARGS

while true
do
    PORT=$(( ((RANDOM<<15)|RANDOM) % 49152 + 10000 ))
    status="$(nc -z 127.0.0.1 $PORT < /dev/null &>/dev/null; echo $?)"
    if [ "${status}" != "0" ]; then
        break;
    fi
done
echo $PORT

python3 -m torch.distributed.launch --nproc_per_node=${NGPUS} --master_port $PORT train.py --launcher pytorch ${PY_ARGS}

The first training log is attached as well.
train-waymo-pvt-ssd.log

Hi, I wonder whether you have any pre-trained models on the KITTI dataset? It can be a great help!

Besides, may I ask which implementation of PointPillars and SECOND you used? Is it OpenPCDet?

hello, congratulations! An excellent work, could you please share your code?

您好，我想请问一下，我已经使用V100-SXM2-32GB * 1卡成功在Kitti数据集中复现结果，
但输入python -m pcdet.datasets.waymo.waymo_dataset --func create_waymo_infos --cfg_file tools/cfgs/dataset_configs/waymo_dataset.yaml换用waymo数据集进行数据预处理时，报错如图所示，您知道什么原因吗

作者你好 工作非常好 请问一下有NuScenes数据集的配置文件吗

when I train waymo dataset, i meet the problem.
I use openpcdet to produce the waymo dataset.

KeyError: 'extrinsic'

Thanks for your work first.
My question is why your codes cannot be run on the latest SPCONV version(2.3.6).
After several days of debugging, I finally found that the spconv version you use here is weird.

For example,
If you use the .indices method to get the voxel index of the feature map backbone.
In the version you provide, the order of batch dim is increased. In this way, every batch is organized separately.
it looks like (0, x, y, z), (0, x, y, z) , (1, x, y, z), (1, x, y, z) , (2, x, y, z), (2, x, y, z)

However, from 2.2 version, the return value is out of order.
it looks like (0, x, y, z), (1, x, y, z) , (1, x, y, z), (0, x, y, z) , (2, x, y, z), (1, x, y, z)

Could you please check it?
The lastest version of spconv is much faster than 2.1 on Ampere Architecture card