Android apps which demonstrate the use of OpenAI's CLIP model for zero-shot image classification and text-image retrieval using clip.cpp

• Fully on-device inference of the CLIP model (generate text and image embeddings)
• Uses JNI bindings over clip.cpp which itself is based on ggml for efficient inference

The project consists of two Gradle modules, app-zero-shot-classify and app-text-image-search which contain the sources files for the 'Zero Shot Image Classification' and `Text-Image-Search' apps respectively.

1. Clone the project and open the resulting directory in Android Studio. An automatic Gradle build should start, if not click on the Build menu and select Make Project.

git clone https://github.com/shubham0204/CLIP-Android --depth=1

2. Connect the test-device to the computer and make sure that the device is recognized by the computer.

3. Download one of the GGUF models from the HuggingFace repository. For instance, if we download the CLIP-ViT-B-32-laion2B-s34B-b79K_ggml-model-f16.gguf model, we need to push it to the test-device's file-system using adb push,

adb push CLIP-ViT-B-32-laion2B-s34B-b79K_ggml-model-f16.gguf /data/local/tmp/clip_model_fp16.gguf

[!INFO] It is not required to place the GGUF model in the /data/local/tmp directory. A path to the app's internal storage or any other directory which the app can accessible is also allowed.

4. For both modules, in MainActivityViewModel.kt, ensure that the MODEL_PATH variable points to the correct model path on the test-device. For instance, if the model is pushed to /data/local/tmp/clip_model_fp16.gguf, then the MODEL_PATH variable should be set to /data/local/tmp/clip_model_fp16.gguf. Moreover, you can configure NUM_THREADS and VERBOSITY variables as well.

private val MODEL_PATH = "/data/local/tmp/clip_model_fp16.gguf"
private val NUM_THREADS = 4
private val VERBOSITY = 1

5. Select one of the module in the Run / Debug Configuration dropdown in the top menu bar, and run the app on the test-device by clicking on the Run button (Shift + F10) in Android Studio.

1. Navigate to this fork of clip.cpp and clone the branch (add-android-sample) and open the resulting directory in Android Studio.

2. The project contains two modules, app and clip. The AAR of the clip module can be found in the app-text-image-search/libs and app-zero-shot-classify/libs directories of this project. Running gradlew clip:assemble should build the debug and release versions of clip as an AAR.

3. The AAR can be added to the libs directory of your project and added as a dependency in the build.gradle file of the app module.

dependencies {
    // ...
    implementation(files("libs/clip.aar"))
    // ...
}

• CLIP: Connecting Text and Images
• Learning Transferable Visual Models From Natural Language Supervision
• clip.cpp
• ggml
• shubham0204's PR that adds JNI bindings to clip.cpp

@article{DBLP:journals/corr/abs-2103-00020,
  author       = {Alec Radford and
                  Jong Wook Kim and
                  Chris Hallacy and
                  Aditya Ramesh and
                  Gabriel Goh and
                  Sandhini Agarwal and
                  Girish Sastry and
                  Amanda Askell and
                  Pamela Mishkin and
                  Jack Clark and
                  Gretchen Krueger and
                  Ilya Sutskever},
  title        = {Learning Transferable Visual Models From Natural Language Supervision},
  journal      = {CoRR},
  volume       = {abs/2103.00020},
  year         = {2021}
}