Open yimingzhou1 opened 1 year ago
Thanks for the reporting. We will update the ckpt.
How did you solve this problem
you can write a simple script to modify the keys of the state_dict and overwrite the .pth file. The only difference if you notice is the extra prefix 'detector'.
You can also check this for ready implementations: https://discuss.pytorch.org/t/how-to-use-the-helper-function-consume-prefix-in-state-dict-if-present/129505
Hi,
Thanks for open-sourcing the code of your great work! Looks like there are some bugs when running the current
tools/inference.py
.When using the
configs/bdd100k/qdtrack-frcnn_r50_fpn_12e_bdd100k.py
as config andqdtrack-frcnn_r50_fpn_12e_bdd100k-13328aed.pth
as the checkpoint (from the google drive link provided in the README file), the model and loaded state dict do not match exactly. Looks like you updated the name of layers but didn't update the definition of the pre-trained model. Manually changing the layer names in the.pth
file will work.