Removed tod3cap_fusion

[chreisinger]

Hello,

Is there a particular reason why you deleted tod3cap_fusion in commit 37bbee7?

Best regards,
Christian

[ucaszyp]

Hi，to make testing and usage more convenient, we have integrated the code for both fusion and camera-only mode into a unified codebase. Different training and testing modes can be switched easily through configuration files.

[ucaszyp]

To achieve better training results, it is necessary to first train the detector (BEVFormer or our implemented BEVFusion) until convergence (you can refer to bevfusion_tiny_stage1.py or bevformer_tiny_stage1.py). Then, load the trained LLAMA-Adapter model and train the entire network end-to-end (refer to bevfusion_tiny_stage3.py or bevformer_tiny_stage3.py).

[ucaszyp]

If computational resources are abundant, after training the detector, you can also freeze the detector parameters and train the LLaMA-Adapter separately, before finally training the entire network end-to-end. When using 8 GPUs for training, it's recommended to use the weights from epochs 18-20 for the detector. Subsequent training of the LLaMA-Adapter alone or end-to-end training with all components typically only requires 5-10 epochs each.

[ucaszyp]

Due to the large size of the pre-trained model （including the parameters of LLaMA）, we are currently working on uploading our pre-trained model.

[ucaszyp]

Hello,

Is there a particular reason why you deleted tod3cap_fusion in commit 37bbee7?

Best regards, Christian

If you have any questions, please feel free to contact us. We will continuously update issues related to the dataset and code.