No such file or directory: '/root/sda2/HEAL/data/OPV2V/train/2021_09_10_12_07_11/216/000309_depth0.png'

[Lesliewsq]

Hi，I have a problem when I train on OPV2V dataset. python opencood/tools/train.py -y opencood/hypes_yaml/opv2v/MoreModality/2_modality_end2end_training/lidar_camera_coalign.yaml

FileNotFoundError: [Errno 2] No such file or directory: '/root/sda2/HEAL/data/OPV2V/train/2021_09_10_12_07_11/216/000309_depth0.png'

[yifanlu0227]

Hi, you need to download the OPV2V-H dataset, which includes the depth file.

[Lesliewsq]

I have downloaded the original OPV2V dataset and created a dataset folder under HEAL as follow. Should I put OPV2V-H-depth.zip and OPV2V-H-LiDAR.zip under /HEAL/OPV2V/ and then unzip them or just put them under a new folder and then unzip them?

├── OPV2V │ ├── additional │ ├── test │ ├── train │ └── validate

[yifanlu0227]

Please unzip the OPV2V-H-depth.zip and OPV2V-H-LiDAR.zip outside of OPV2V folder. They will produce a parallel folder OPV2V-H, just as the README illustrated.

[Lesliewsq]

Still the same error.

[xhjy2020]

Still the same error.

+1, It seems that the default configuration file does not point to depth data.

[Lesliewsq]

Still the same error.

+1, It seems that the default configuration file does not point to depth data. Yeah，I found that the camera.png and depth.png data are read from the same path, but aren't they placed in two different datasets?

[yifanlu0227]

I'm very sorry about that. @xhjy2020 @Lesliewsq Please modify line 150 in opencood/data_utils/datasets/basedataset/opv2v_basedataset.py to

                    depth_files = self.find_camera_files(cav_path, 
                                                timestamp, sensor="depth")
                    depth_files = [depth_file.replace("OPV2V", "OPV2V_Hetero") for depth_file in depth_files]

There are other fixes for OPV2V-H dataset file reading. Please use the latest commit via

git fetch origin
git merge origin/main

[yifanlu0227]

This is because I convert the RGB image & depth image into a hdf5 file for fast loading. In this case, I do not load the depth image from OPV2V_Hetero folder and omit the error.

Please note that it's OPV2V_Hetero not OPV2V-H, and I just pushed another commit to fix my typo.

@xhjy2020 @Lesliewsq

[yifanlu0227]

By the way, you can also speed up the image loading by storing all images in an HDF5 format. They are larger but facilitate your training speed.

You can find the conversion script in opencood/utils/img2hdf5.py

[xhjy2020]

By the way, you can also speed up the image loading by storing all images in an HDF5 format. They are larger but facilitate your training speed.

You can find the conversion script in opencood/utils/img2hdf5.py

Nice! thank you very much for your patient explanation!!

[yifanlu0227]

I hope I didn't cause more trouble.

I just pushed the third commit and apologize for my typo again @xhjy2020

                    depth_files = self.find_camera_files(cav_path, 
                                                timestamp, sensor="depth")
                    depth_files = [depth_file.replace("OPV2V", "OPV2V_Hetero") for depth_file in depth_files]

[Lesliewsq]

I have tried this command python opencood/tools/train.py -y opencood/hypes_yaml/opv2v/MoreModality/2_modality_end2end_training/lidar_camera_coalign.yaml , and an error occurred. However，opencood/hypes_yaml/opv2v/MoreModality/2_modality_end2end_training/lidar_camera_attfuse.yaml works.

It seems lidar_camera_coalign.yaml config file has problem with it.

No model_train_init function Traceback (most recent call last): File "opencood/tools/train.py", line 189, in main() File "opencood/tools/train.py", line 118, in main ouput_dict = model(batch_data['ego']) File "/root/miniconda3/envs/CoAlign/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl return forward_call(*input, *kwargs) File "/root/sda2/WSQ/HEAL/opencood/models/heter_model_baseline_ms.py", line 212, in forward cls_preds = self.cls_head(fused_feature) File "/root/miniconda3/envs/CoAlign/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl return forward_call(input, **kwargs) File "/root/miniconda3/envs/CoAlign/lib/python3.7/site-packages/torch/nn/modules/conv.py", line 446, in forward return self._conv_forward(input, self.weight, self.bias) File "/root/miniconda3/envs/CoAlign/lib/python3.7/site-packages/torch/nn/modules/conv.py", line 443, in _conv_forward self.padding, self.dilation, self.groups) RuntimeError: Given groups=1, weight of size [2, 256, 1, 1], expected input[1, 384, 256, 256] to have 256 channels, but got 384 channels instead.

[yifanlu0227]

Yes, the shrink_header is missing. Please check the latest commit and now it should work. @Lesliewsq

[Lesliewsq]

Thanks for the reply. Nice work！ @yifanlu0227

[Lesliewsq]

Both 3_modality and 4_modality yaml files report "channel size mismatch" error. CUDA_VISIBLE_DEVICES=1 python opencood/tools/train.py -y opencood/hypes_yaml/opv2v/MoreModality/3_modality_end2end_training/m1m2m3_pyramid.yaml

Traceback (most recent call last): File "opencood/tools/train.py", line 189, in main() File "opencood/tools/train.py", line 118, in main ouput_dict = model(batch_data['ego']) File "/root/miniconda3/envs/CoAlign/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl return forward_call(*input, kwargs) File "/root/sda2/WSQ/HEAL/opencood/models/heter_pyramidcollab.py", line 145, in forward feature = eval(f"self.encoder{modality_name}")(data_dict, modality_name) File "/root/miniconda3/envs/CoAlign/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl return forward_call(*input, *kwargs) File "/root/sda2/WSQ/HEAL/opencood/models/heter_encoders.py", line 79, in forward batch_dict = self.spconv_block(batch_dict) File "/root/miniconda3/envs/CoAlign/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl return forward_call(input, kwargs) File "/root/sda2/WSQ/HEAL/opencood/models/sub_modules/sparse_backbone_3d.py", line 114, in forward x = self.conv_input(input_sp_tensor) File "/root/miniconda3/envs/CoAlign/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl return forward_call(*input, *kwargs) File "/root/miniconda3/envs/CoAlign/lib/python3.7/site-packages/spconv/pytorch/modules.py", line 137, in forward input = module(input) File "/root/miniconda3/envs/CoAlign/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl return forward_call(input, **kwargs) File "/root/miniconda3/envs/CoAlign/lib/python3.7/site-packages/spconv/pytorch/conv.py", line 187, in forward 1] == self.in_channels, "channel size mismatch" AssertionError: channel size mismatch

[xhjy2020]

same error when i train m3_no_fusion using opencood/hypes_yaml/opv2v/Single/m3_SECOND32_pretrain.yaml. Is this due to the spconv version?

[yifanlu0227]

Are you using spconv 2.x? Could you provide more error information?

[xhjy2020]

Are you using spconv 2.x? Could you provide more error information?

yes, the version is spconv-cu117, 2.3.6

[xhjy2020]

Are you using spconv 2.x? Could you provide more error information?

yes, the version is spconv-cu117, 2.3.6

It is indeed an issue with the spconv version. I replaced 2.3.6 with 1.2.1, then this issue won't occur anymore. Thanks!!