Missing model topology/Architecture while load in tfjs. Even loaded the wights and config file but each frame processing time depends upon system/hardware .

[vasanthhr]

Hello, Appreciate your work and nice librarries.

• Question1:

I am using tfjs(tensorflowjs) to load your weight file, but its giving error saying "load topology or architecture before inference", looks like you provided only weight file as json.

Is there any way you can guide me , how to extract weight file with topology/architecture of face detection model. Or can you export and post it in your github.?

• Question2:

I loaded your weight file face_detection_model-weights_manifest.json using your tiny-yolov2.js , but the prediction time taking is more, and also it depends upon system/hardware where I am executing, If systems has good graphics card then it predicts and gives the output faster, otherwise its taking much time.

Is there any way, I can make changes so that your weight file and architecture works seamlessly, faster and consistent execution time in all system's browser ?

• Question3:

I preferred to train tiny-yolo-2 using darknet, and then planning to convert into H5 model and then to model,json. But it gives 20MB file size(including topology and weights). I can reduce to 5 MB based on quantization method in tfjs.

Can you please guide me how to reduce the model size in KB, as you already did it( your model size in face-api repo are in KB). Please guide me here.

I can take your paid consultation , Look forward to hearing from you.