Training TraffiCamNet using the Kitti Dataset

Please provide the following information when requesting support.

• Hardware: Desktop (RTX 3070) (T4/V100/Xavier/Nano/etc)
• Network Type: Detectnet_v2, TrafficCamNet
• TLT Version: toolkit_version: 3.21.11 (obtained from running tao info)
• Training spec file: detectnet_v2_train_resnet18_kitti_traffic.txt (5.6 KB)

  • Error: INFO:tensorflow:Graph was finalized.
    2022-02-20 01:20:28,903 [INFO] tensorflow: Graph was finalized.
    INFO:tensorflow:Restoring parameters from /tmp/tmpk8r9ylb2/model.ckpt-0
    2022-02-20 01:20:29,286 [INFO] tensorflow: Restoring parameters from /tmp/tmpk8r9ylb2/model.ckpt-0
    Traceback (most recent call last):
    File “/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/client/”, line 1365, in _do_call
    return fn(*args)
    File “/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/client/”, line 1350, in _run_fn
    target_list, run_metadata)
    File “/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/client/”, line 1443, in _call_tf_sessionrun
    tensorflow.python.framework.errors_impl.NotFoundError: 2 root error(s) found.
    (0) Not found: Key cost_sums/person-bbox not found in checkpoint
    [[{{node save/RestoreV2}}]]
    (1) Not found: Key cost_sums/person-bbox not found in checkpoint
    [[{{node save/RestoreV2}}]]
    0 successful operations.
    0 derived errors ignored.

Ipynb File: detectnet_v2_from_forum.ipynb (5.1 MB)

I’m attempting to modify the base detectnet_v2 example to work with TrafficCamNet. What I’ve done is to take the base version of the detectnet_v2 .ipynb file, obtained from downloading the cv_samples_v1.3.0 and modify things where necessary in order to get them to work with the TrafficCamNet model.

I was attempting to follow this forum post: Retraining Trafficcamnet with custom vehicle dataset

I tried to avoid the issue that this user had, where the other classes were pruned off.

Based on my training config file, I attempted to map the data from the kitti dataset labels into the preexisting classes utilized by TrafficCamNet.

Also from the forum post linked above, will TrafficCamNet lose it’s ability to classify road signs or people if I train it with the kitti dataset? My worry is that when the user trained to try to home in on cars and two-wheelers, the model’s performance degraded significantly and the ability to detect people seemed to be lost.

Would I be better off downloading the already pruned version of TrafficCamNet and performing inference using that? If so, how would I go about doing that?

I know that the baseline is to use “tao detectnet_v2”, but after that, I’m not sure how to have the model perform inference on data on my local machine, or how to run it through TensorRT for later use on a Jetson Nano.

Any help is appreciated and if any additional info is needed, please let me know!

If you train with public KITTI dataset which has not road signs class, the trained tlt model will not detect road sign.

Yes, you can directly use the pruned version of TrafficCamNet.
See TrafficCamNet — TAO Toolkit 3.21.11 documentation

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.