export_onnx.md

Guidelines for exporting clip onnx models

Export clip onnx model

1. After meeting the installation environment requirements in the original huggingface/openai/clip-vit-base-patch32 repository for exporting the onnx model, execute the following command to export the onnx model:

optimum-cli export onnx --model weights_path/ --task image-to-text --opset 18 path_to_output/

# weights_path refers to the original model file
# image-to-text refers to the type of the model
# path_to_output refers to the path where the decoder model will be saved
After execution, the model.onnx will be generated in the path_to_output.

2. Export clip text and images model

cd python
python truncated_onnx.py --model path_to_output/

After execution, the clip_text.onnx  and clip_images will be generated in the path_to_output.

Note

To remove the attention_mask input from the clip_text model, you can use the onnx-modify tool.