---
license: mit
tags:
- cad
- template-fitting
- pytorch
---

# TempFit checkpoint

Model weights for template-based industrial CAD scene reconstruction.

[Code and installation](https://github.com/Daisy-1227/TempFit) · [Dataset](https://huggingface.co/datasets/Daisy-1227/TempFit)

The `tempfit-scratch` directory contains `model.safetensors` and its companion `model.json`.
The model weights occupy approximately 508.49 MB. The accompanying metadata
provides the configuration and file checksums used by TempFit.

## Usage

Install TempFit and place the dataset under its `data/` and `splits/` directories.
Download the weights from the repository root:

```bash
pip install -U huggingface_hub
hf download Daisy-1227/TempFit --include "tempfit-scratch/*" --local-dir checkpoints
```

Run autoregressive inference:

```bash
python inference.py --config configs/inference.yaml --checkpoint checkpoints/tempfit-scratch/model.safetensors --output_dir outputs/tempfit-scratch
```

Initialize a new training run from these weights:

```bash
accelerate launch --multi_gpu --num_processes 8 --mixed_precision bf16 train.py --config configs/train.yaml --pretrained checkpoints/tempfit-scratch/model.safetensors --output_dir outputs/tempfit-scratch-finetune
```

For a single GPU:

```bash
python train.py --config configs/train.yaml --pretrained checkpoints/tempfit-scratch/model.safetensors --output_dir outputs/tempfit-scratch-finetune
```

Gradient accumulation adjusts to preserve the default global batch size of 128.

Keep `model.json` alongside its weight file. The package supports inference and
pretrained initialization; optimizer and scheduler states are excluded.

## License

The weights are released under the MIT License. Dataset access and use are
governed separately by the dataset license.