V2V-C Model Training: Training Loss Decreases but Validation Loss Stagnates

[lmy628]

Hello,

I'm encountering an issue while training the V2V-C model where the ego vehicle is camera agent and the camera/lidar ratio is set to 1. This is the first training stage as described in the paper.

Issue:

• The training loss is consistently decreasing as expected.
• However, the validation loss remains unchanged despite the training loss decreasing.

Training Details:

• I use train_camera.py and fix the lidar branch parameters
• The training set is opv2v train.
• The validation set is opv2v validate.

I'm looking forward to any advice or suggestions on how to address this issue. Thank you in advance for your help!

[lmy628]

May I ask if you use some checkpoints from model zoo of BEVFormer? Looking forward to your reply. Thank you !

[cope-code]

May I ask if you use some checkpoints from model zoo of BEVFormer? Looking forward to your reply. Thank you !

你好，请问你的问题解决了吗，我现在是不能训练，使用train_camera.py进行训练，并添加camera_to_lidar_ratio: 1
ego_mode: 'camera',报错 File "HMViT/opencood/models/cross_view_transformer.py", line 36, in forward
x = batch_dict['camera']
KeyError: 'camera'