import onnxruntime as ort
from huggingface_hub import hf_hub_download

for f in ("cn_clip_vision.onnx", "cn_clip_text.onnx"):
    hf_hub_download("felixdu/chinese-clip-vit-base-patch16-onnx", f, local_dir="cn-clip-b16-onnx")

vis_session = ort.InferenceSession("cn-clip-b16-onnx/cn_clip_vision.onnx")
txt_session = ort.InferenceSession("cn-clip-b16-onnx/cn_clip_text.onnx")

Details

Base model: OFA-Sys/chinese-clip-vit-base-patch16
ONNX opset: 18
Image input: 224x224 RGB, normalized with CLIP mean/std
Text tokenizer: BertTokenizerFast from OFA-Sys/chinese-clip-vit-base-patch16 (max_length=52)
Output dim: 512 (L2-normalized)
Verified: PyTorch vs ONNX cosine similarity = 1.000000 for both vision and text