---
license: apache-2.0
pipeline_tag: image-classification
library_name: coreai
tags:
  - core-ai
  - ios
  - imagenet
---

# PVT v2 b0 for Core AI

PVT v2 b0 (3.7M parameters, ImageNet-1k) from [timm](https://huggingface.co/docs/timm), exported to Apple Core AI with the [`apple/coreai-models`](https://github.com/apple/coreai-models/tree/main/models/pvt) recipe:

```sh
uv run models/pvt/export.py --dtype float16
```

| Folder       | Contents                                                                                  | Size   | Platforms  |
| ------------ | ----------------------------------------------------------------------------------------- | ------ | ---------- |
| `pvt-v2-b0/` | `pvt-v2-b0.aimodel`: `main(x) -> logits`, `x` is `[1, 3, 224, 224]` float16 RGB normalized with the ImageNet mean and std. `labels.json`: the 1000 ImageNet class names, by logit index | 7.1 MB | iOS, macOS |

## Use

With [`@react-native-ai/core-ai`](https://github.com/callstackincubator/ai/tree/main/packages/core-ai), download `pvt-v2-b0/` into app storage, then:

```ts
const { output } = await CoreAI.runTask(
  'classification',
  { id: 'pvt-v2-b0', sourceType: 'file', sourceUri: pathToFolder },
  { imageUri },
  { topK: 5 }
)
// output.classifications: [{ label: 'tabby, tabby cat', index: 281, confidence: 0.36 }, ...]
```

The Neural Engine compiler rejects this export on macOS 27.0, so load it on the GPU there (`computeUnits: 'gpu'`).

## License

PVT v2 and its timm weights are Apache-2.0 ([paper](https://arxiv.org/abs/2106.13797)). The labels come from timm's ImageNet metadata. This repository only converts the model to Core AI format.