---
license: mit
tags:
  - point-cloud
  - 3d
  - self-supervised
  - point-transformer
  - semantic-segmentation
---

# Sonata Model Weights

Model checkpoints for [Sonata](https://github.com/cad-results/sonata) — self-supervised pre-trained Point Transformer V3 for 3D point cloud downstream tasks.

## Files

| Path | Description |
|------|-------------|
| `pretrained/sonata.pth` | Sonata base pretrained backbone |
| `pretrained/sonata_small.pth` | Sonata small pretrained backbone |
| `pretrained/sonata_linear_prob_head_sc.pth` | Linear probe head |
| `big/best_big_geometry_model.pth` | Best big geometry finetuned model |
| `medium/best_medium_geometry_model.pth` | Best medium geometry finetuned model |
| `mixed/best_mixed_geometry_model.pth` | Best mixed geometry finetuned model |
| `binary/best_binary_model.pth` | Best binary classification model |
| `small/best_small_geometry_model.pth` | Best small geometry finetuned model |

## Source Code

[https://github.com/cad-results/sonata](https://github.com/cad-results/sonata)