| --- |
| license: apache-2.0 |
| tags: |
| - autonomous-driving |
| - perception |
| - bevformer |
| - 3d-object-detection |
| - pytorch |
| - synthetic-data |
| datasets: |
| - Zgogo0/simData-Dataset |
| library_name: pytorch |
| pipeline_tag: object-detection |
| --- |
| |
| # BEVFormer-Tiny (Trained on SimData) |
|
|
| ## Model Description |
|
|
| This repository contains model weights for **BEVFormer-tiny**, a camera-only 3D object detection model. These weights have been trained or fine-tuned on the synthetic **SimData dataset** (a NuScenes-format synthetic dataset containing Highway, Urban, and Parking scenarios). |
|
|
| - **Architecture**: BEVFormer (Tiny variant) |
| - **Input Modality**: Multi-view Camera (6 cameras) |
| - **Training Data**: [SimData-Dataset](https://huggingface.co/datasets/Zgogo0/simData-Dataset) |
| - **Framework**: PyTorch / MMDetection3D |
|
|
| ## Model Zoo / Checkpoints |
|
|
| This repository provides three different variations of the trained model. Please select the weight file that best suits your testing needs. |
|
|
| | File Name | Training Strategy | Epochs | Description | |
| | :--- | :--- | :--- | :--- | |
| | **`epoch_24.pth`** | **Train from Scratch** | 24 | Trained directly on the SimData dataset from scratch for 24 epochs. | |
| | **`epoch_30.pth`** | **Train from Scratch** | 30 | Trained directly on the SimData dataset from scratch for 30 epochs (longer training schedule). | |
| | **`use_predcheckout_train_on_simdata.pth`** | **Fine-tuning (Transfer Learning)** | 24 | **Recommended**. Initialized from a model pre-trained on real-world NuScenes data (`bevformer_tiny_epoch_24.pth`), then fine-tuned on the SimData dataset for an additional 24 epochs. This model usually exhibits better convergence and domain adaptation capabilities. | |
|
|
| *(Note: If your file for the 30-epoch run is named differently, please check the file list in the 'Files' tab, but the logic remains as described above.)* |
|
|
| ## How to Use |
|
|
| Since BEVFormer is a custom architecture based on `mmdetection3d`, you cannot load it directly using `AutoModel`. You need to use the original BEVFormer codebase and load these weights manually. |
|
|
| ### 1. Download Weights |
| You can download the specific weight file using the `huggingface_hub` python library. |
|
|
| ```python |
| from huggingface_hub import hf_hub_download |
| |
| # Define the repository and the specific file you want |
| repo_id = "Zgogo0/Bevformer_Simdata" # Replace with your actual Repo ID |
| |
| # Example: Downloading the fine-tuned model |
| filename = "use_predcheckout_train_on_simdata.pth" |
| |
| checkpoint_path = hf_hub_download(repo_id=repo_id, filename=filename) |
| print(f"Model downloaded to: {checkpoint_path}") |
