incorrect testing results
qiang-sean-liu opened this issue · comments
I ran testing on Apollosim dataset, but got results that did not make any sense. The classification scores were all around 0.5, so if the prob_th set to 0.7 as default, there were no lanes detected.
The model apollo_anchor3dlane.pth was used for testing, which was downloaded from the model zoo https://pan.baidu.com/s/1HPYxsNNSOO5CY7-RwAt9cw?pwd=bqvy.
There was a warning during testing:
'''
load checkpoint from local path: ./pretrained/apollo_anchor3dlane.pth
The model and loaded state dict do not match exactly
missing keys in source state_dict: cls_layer.1.layer.0.weight, cls_layer.1.layer.0.bias, cls_layer.1.layer.2.weight, cls_layer.1.layer.2.bias, cls_layer.1.layer.4.weight, cls_layer.1.layer.4.bias, reg_x_layer.1.layer.0.weight, reg_x_layer.1.layer.0.bias, reg_x_layer.1.layer.2.weight, reg_x_layer.1.layer.2.bias, reg_x_layer.1.layer.4.weight, reg_x_layer.1.layer.4.bias, reg_z_layer.1.layer.0.weight, reg_z_layer.1.layer.0.bias, reg_z_layer.1.layer.2.weight, reg_z_layer.1.layer.2.bias, reg_z_layer.1.layer.4.weight, reg_z_layer.1.layer.4.bias, reg_vis_layer.1.layer.0.weight, reg_vis_layer.1.layer.0.bias, reg_vis_layer.1.layer.2.weight, reg_vis_layer.1.layer.2.bias, reg_vis_layer.1.layer.4.weight, reg_vis_layer.1.layer.4.bias
'''
Is this the underlining problem? Please advise on how to fix this.
Thanks!
It seems that you have used an unmatched config file. The checkpoint you used corresponds to this file: configs/apollosim/anchor3dlane.py. If you'd like to use configs/apollosim/anchor3dlane_iter.py during testing, you'll have to download this checkpoint: https://pan.baidu.com/s/1f4Ssts_cUU7kGtXUyRulLA?pwd=pfe5