Pre-trained model files missing

[Mat8e]

Hi,
Thanks for sharing the code of your method. I am trying to reproduce your results with your code but I am missing the pre-trained weight files you are using to initialize your networks: pre_weight/FPN18_retinanet_968.pth and pre_weight/first_stage/fusion_split/voxelnet-35210.tckpt.
Is there a way to download those files somewhere or pre-train the networks in a similar way ?

[DanielZ98]

Hi, did you find the FPN18_retinanet_968.pth?

[Mat8e]

Hi, no I commented the line in the end and started a new training from 0. Apparently, from what I remember from the paper, it is simply a retinanet pretrained on the Kitti dataset. If you managed to come to this stage of the debugging, and want to persevere to reach close results to the one advertised in the paper, you need to correct the mistake in second/pytorch/models/rpn.py line 410:

original: box_preds = self.conv_box(bev_feature)
correction: box_preds = self.conv_box(concat_feat)

After my investigations I still did not manage to find the "pointnet" used for the feature pooling and it is an annoying mystery. However with that correction you become very close to their results (~1-2%) without even using the second stage nor the pertained retinanet.
Good luck !

[DanielZ98]

Thanks a lot

[mc171819]

hi, i want to ask that did you have some mistake in 3D-CVF/second/core/box_np_ops.py when run from spconv.utils import rbbox_iou, rbbox_iou_loss. I can't import rbbox_iou_loss correctly. Did you meet this problem?

[mc171819]

Hi, no I commented the line in the end and started a new training from 0. Apparently, from what I remember from the paper, it is simply a retinanet pretrained on the Kitti dataset. If you managed to come to this stage of the debugging, and want to persevere to reach close results to the one advertised in the paper, you need to correct the mistake in second/pytorch/models/rpn.py line 410:

original: box_preds = self.conv_box(bev_feature) correction: box_preds = self.conv_box(concat_feat)

After my investigations I still did not manage to find the "pointnet" used for the feature pooling and it is an annoying mystery. However with that correction you become very close to their results (~1-2%) without even using the second stage nor the pertained retinanet. Good luck !

hi, when you run this code did you find many bugs? will you mind sending me the code you run successfully? thank you very much!

[FANGZHENGFA]

Hi, no I commented the line in the end and started a new training from 0. Apparently, from what I remember from the paper, it is simply a retinanet pretrained on the Kitti dataset. If you managed to come to this stage of the debugging, and want to persevere to reach close results to the one advertised in the paper, you need to correct the mistake in second/pytorch/models/rpn.py line 410:

original: box_preds = self.conv_box(bev_feature) correction: box_preds = self.conv_box(concat_feat)

After my investigations I still did not manage to find the "pointnet" used for the feature pooling and it is an annoying mystery. However with that correction you become very close to their results (~1-2%) without even using the second stage nor the pertained retinanet. Good luck !

Hi，
how to starte a new training from 0?
sincerely.

[guzuoxu]

Hi! Could you provide more details?

[FANGZHENGFA]

Hi! Could you provide more details?

I just commented the code of pre-train-model-load, and train from 0

[guzuoxu]

Hi! Could you provide more details?

I just commented the code of pre-train-model-load, and train from 0

Thank you! I understand what you mean. In addition, when you are training 3DCVF, is the pkl file generation method on the kitti dataset the same as the second generation method?

[FANGZHENGFA]

Hi! Could you provide more details?

I just commented the code of pre-train-model-load, and train from 0

Thank you! I understand what you mean. In addition, when you are training 3DCVF, is the pkl file generation method on the kitti dataset the same as the second generation method?

It depends on the version of SECOND.
3D-CVF's pkl files is the same as SECOND v1.5. You can follow the README file in SECONDv1.5 and generate the pkl files.
The pkl of SECOND-master(and v1.5.1) is different to 3D-CVF.

[guzuoxu]

I got it, thank you very much for your reply！

[jiangliangliang2020]

I got it, thank you very much for your reply！

Hi,I commented the code of pre-train-model-load,but I am missing this file: KITTI/object/gt_database/7392_Car_2.bin.Did you meet this problem?

[jiangliangliang2020]

I got it, thank you very much for your reply！

Hi,I commented the code of pre-train-model-load,but I am missing this file: KITTI/object/gt_database/7392_Car_2.bin.Did you meet this problem?

[guzuoxu]

I got it, thank you very much for your reply！

Hi,I commented the code of pre-train-model-load,but I am missing this file: KITTI/object/gt_database/7392_Car_2.bin.Did you meet this problem?

No, I did not encounter this problem. If there is a missing file, how about adding this file to the corresponding folder?

[jiangliangliang2020]

Thank you very much for your reply. How did you get the car.bin file?

…

---Original--- From: ***@***.***> Date: Fri, Dec 3, 2021 09:47 AM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [rasd3/3D-CVF] Pre-trained model files missing (#8) I got it, thank you very much for your reply！ Hi,I commented the code of pre-train-model-load,but I am missing this file: KITTI/object/gt_database/7392_Car_2.bin.Did you meet this problem? No, I did not encounter this problem. If there is a missing file, how about adding this file to the corresponding folder? — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

[guzuoxu]

Did you follow create_data.py in sencond1.5 to generate data?

[jiangliangliang2020]

I git it，Thank you very much！

…

---Original--- From: ***@***.***> Date: Fri, Dec 3, 2021 10:36 AM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [rasd3/3D-CVF] Pre-trained model files missing (#8) Did you follow create_data.py in sencond1.5 to generate data? — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

[qrzeffort]

Hi! Could you provide more details?

I just commented the code of pre-train-model-load, and train from 0

Thank you! I understand what you mean. In addition, when you are training 3DCVF, is the pkl file generation method on the kitti dataset the same as the second generation method?

Hi，I modified code according to the steps in issue, but it didn't succeed.Now,I wonder what the "train from 0" mean?

[qrzeffort]

And I occurred error：

Traceback (most recent call last):
File "pytorch/train.py", line 691, in
fire.Fire()
File "/home/nn/anaconda3/envs/3dcvf/lib/python3.6/site-packages/fire/core.py", line 141, in Fire
component_trace = _Fire(component, args, parsed_flag_args, context, name)
File "/home/nn/anaconda3/envs/3dcvf/lib/python3.6/site-packages/fire/core.py", line 471, in _Fire
target=component.name)
File "/home/nn/anaconda3/envs/3dcvf/lib/python3.6/site-packages/fire/core.py", line 681, in _CallAndUpdateTrace
component = fn(*varargs, **kwargs)
File "pytorch/train.py", line 416, in train
raise e
File "pytorch/train.py", line 250, in train
example = next(data_iter)
File "/home/nn/anaconda3/envs/3dcvf/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 819, in next
return self._process_data(data)
File "/home/nn/anaconda3/envs/3dcvf/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 846, in _process_data
data.reraise()
File "/home/nn/anaconda3/envs/3dcvf/lib/python3.6/site-packages/torch/_utils.py", line 385, in reraise
raise self.exc_type(msg)
ValueError: Caught ValueError in DataLoader worker process 0.
Original Traceback (most recent call last):
File "/home/nn/anaconda3/envs/3dcvf/lib/python3.6/site-packages/torch/utils/data/_utils/worker.py", line 178, in _worker_loop
data = fetcher.fetch(index)
File "/home/nn/anaconda3/envs/3dcvf/lib/python3.6/site-packages/torch/utils/data/_utils/fetch.py", line 44, in fetch
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/home/nn/anaconda3/envs/3dcvf/lib/python3.6/site-packages/torch/utils/data/_utils/fetch.py", line 44, in
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/home/nn/3D-CVF/second/pytorch/builder/input_reader_builder.py", line 18, in getitem
return self._dataset[idx]
File "/home/nn/3D-CVF/second/data/dataset.py", line 70, in getitem
prep_func=self._prep_func)
File "/home/nn/3D-CVF/second/data/preprocess.py", line 344, in _read_and_prep_v9
example = prep_func(input_dict=input_dict)
File "/home/nn/3D-CVF/second/data/preprocess.py", line 204, in prep_pointcloud
gt_boxes, points = prep.random_flip(gt_boxes, points)
ValueError: too many values to unpack (expected 2)

I tried to modify the config and data files，but it didn't succeed. Can someone help me？

[zijianh4]

Did you follow create_data.py in sencond1.5 to generate data?

Hi, do you solve this problem? I use the create_data.py from SECOND-1.5 but it seems not correct. Thanks!