python/export_onnx.py

import torch
from transformers import AutoModel, AutoProcessor

class PoolerOutputWrapper(torch.nn.Module):
    def __init__(self, model, model_part):
        super(PoolerOutputWrapper, self).__init__()
        if model_part == 'vision':
            self.model = model.vision_model
        elif model_part == 'text':
            self.model = model.text_model
        else:
            raise ValueError("model_part must be either 'vision' or 'text'")
    
    def forward(self, x):
        outputs = self.model(x)
        return outputs.pooler_output
    
# load the model and processor
ckpt = "google/siglip2-base-patch16-224"
model = AutoModel.from_pretrained(ckpt, device_map="auto").eval().to("cpu")
processor = AutoProcessor.from_pretrained(ckpt)


dummy_img = torch.randn(1, 3, 224, 224)
dummy_ids = torch.randint(1, 1000, (1, 64))

# export image onnx
vision_wrapper = PoolerOutputWrapper(model, 'vision')
torch.onnx.export(vision_wrapper,
    dummy_img,
    f"./onnx/siglip2-base-patch16-224_vision.onnx",
    input_names=['image'],
    output_names=['pooler_output'],
    export_params=True,
    opset_version=14)

# export text onnx
text_wrapper = PoolerOutputWrapper(model, 'text')
torch.onnx.export(text_wrapper,
    dummy_ids,
    f"./onnx/siglip2-base-patch16-224_text.onnx",
    input_names=['text'],
    output_names=['pooler_output'],
    export_params=True,
    opset_version=14)