nuscenes-c generate failed about robobev HOT 13 CLOSED

I can run it successfully on my side. When running bash tools/generate_dataset.sh, you are expected in the ./corruptions folder. However, according to the error information, it seems to look for images within the ./corruptions folder. Please make sure the data_root in the config is correct.

If the above is correct, this might be caused by nuscenes_infos_val.pkl. nuScenes dataset path cannot be changed after you generate it, because it will save the image path the time you generate it. You can load it to check if the path is correct.

from robobev.

Yes, when generating the nuScenes-C dataset, it will use original loading, not custom loading. The latter is only used when benchmarking models.

from robobev.

when I remove all the exciting pkl file in my dataroot, it raised that ' No such file or directory: nuscenes_infos_temporal_val.pkl', when I link 'nuscenes_infos_temporal_val.pkl' again , it comes that img file does not exist: ./data/nuscenes/samples/CAM_FRONT/n015-2018-07-11-11-54-16+0800__CAM_FRONT__1531281439762460.jpg . which is not the dataroot I wrote in config, how can I fix it ?

from robobev.

How do you generate nuscenes_infos_temporal_val.pkl? I think the problem might be in this file.

from robobev.

I used bevdet data generation, in which I want to test it profermance.
So what is the right anno data in anno_root?

from robobev.

It depends on the model you want to test. If you want to test BEVDet, just need to follow it's repo to generate the .pkl file. You can use the following script to debug and check whether the image path in it is wrong.

import pickle
import os
import mmcv

# the annotation root, change it to your own
anno_path = '/nvme/konglingdong/models/RoboDet/data/nuscenes_infos_val_4d_interval3_max60.pkl'
data_root = '/nvme/share/data/sets/nuScenes/samples'
# save path of the correct anno file
save_path = '/nvme/konglingdong/models/RoboDet/data/nuscenes_infos_val_4d_interval3_max60_custom.pkl'

def main():

    with open(anno_path, 'rb') as f:
        anno = pickle.load(f)
    
    prog_bar = mmcv.ProgressBar(len(anno['infos']))
    for i in range(len(anno['infos'])):
        for sensor in anno['infos'][i]['cams'].keys():
           # check weather filename here is correct
            filename = os.path.split(anno['infos'][i]['cams'][sensor]['data_path'])[1]
            anno['infos'][i]['cams'][sensor]['data_path'] = os.path.join(data_root, sensor, filename)
        prog_bar.update()
    
    with open(save_path, 'wb') as f:
        pickle.dump(anno, f)

if __name__=='__main__':
    main()

from robobev.

Additionally, please check if you can run BEVDet on clean nuScenes dataset using tools/dist_test.sh. If so, the problem might not caused by the nuscenes_c.py config

from robobev.

it shows good
[>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 6019/6019, 5026.3 task/s, elapsed: 1s, ETA: 0s
so maybe I need to modify anno_file from nuscenes_infos_temporal_val.pkl to nuscenes_infos_val_4d_interval3_max60.pkl?

from robobev.

Actually, the name doesn't matter. If you set the data_root in the above script correctly, you can try using the newly generated anno file. Also, please check this #2 (comment).

from robobev.

I use this test and it shows good, so why I can not generate nuscenes-c? I am confused now .. I use the same dataroot

from robobev.

Please set a breakpoint here:

A simple work around: change the filepath manually here to the correct one.

from robobev.

Thanks for your kindness response, I just manually modified here and it works now.

from robobev.

I'm glad to hear that. we are sorry this repo is working in process and some parts haven't been well-organized. we will organize it later.

from robobev.