mallman/bioclip-2-coreml

BioCLIP 2 ExecuTorch (CoreML)

ExecuTorch .pte exports of the BioCLIP 2 visual encoder for on-device inference on Apple devices (iOS 18+ / macOS 15+).

Source code & export scripts: github.com/mallman/CoreMLCLIP

Files

File	Precision	Backend	Compute Units	Size
bioclip2_visual_fp16_all.pte	fp16	CoreML + XNNPACK fallback	CPU + GPU + ANE	~581 MB
bioclip2_visual_fp32_cpu.pte	fp32	XNNPACK	CPU only	~1.1 GB

The fp16 CoreML variant is recommended for deployment — it leverages the Apple Neural Engine.

Model Details

• Source model: imageomics/bioclip-2 (MIT license)
• Architecture: ViT-L/14 visual encoder (~302M params)
• Input: [1, 3, 224, 224] float tensor (RGB, normalized)
• Output: 768-dim L2-normalized embedding vector
• ExecuTorch version: 1.1.0
• Minimum deployment target: iOS 18 / macOS 15

Usage

The exported model takes a preprocessed image tensor and returns an L2-normalized embedding. Classification is a dot product against precomputed text embeddings.

Preprocessing

Parameter	Value
Input size	224 x 224
Resize	Bicubic, shortest edge to 224
Crop	Center crop
Color space	RGB, [0, 1] range
Normalization mean	[0.48145466, 0.4578275, 0.40821073]
Normalization std	[0.26862954, 0.26130258, 0.27577711]

Text Embeddings

Precomputed text embeddings for 867K species are available from the TreeOfLife-200M dataset:

• txt_emb_species.npy
• txt_emb_species.json

Verification

Both variants were verified against the original PyTorch model using a CC-BY licensed monarch butterfly photo from iNaturalist:

Variant	Cosine Similarity	Max Abs Diff	Top-1	Top-5
fp16 CoreML	0.999999	0.000183	Danaus plexippus	5/5 match
fp32 XNNPACK	1.000000	0.000000	Danaus plexippus	5/5 match

How to Reproduce

git clone https://github.com/mallman/CoreMLCLIP.git
cd CoreMLCLIP
pip install -r requirements.txt
python export_bioclip2.py

See the GitHub repo for full instructions.