Scoring accuracy is not at par with original CLIP model

[ua741]

For the image of the cat, we compared the scores from the ONNX-based model with those from clip.cpp.

It appears that the ONNX model doesn't provide accurate output. Although we understand that the scores may not match exactly, the relative order should ideally be the same.

clip.cpp output

clip-android-demo (ONNX) output #2

[shubham0204]

@ua741 Are you using the int8 quantized model from clip.cpp? This repo is using int8 quantized ONNX models.

[shubham0204]

@ua741 I have created an Android app that uses JNI bindings built over clip.cpp. Check the project here: https://github.com/shubham0204/CLIP-Android

Also, here's the PR I made to clip.cpp for adding the JNI bindings: monatis/clip.cpp#101