I’ve used a torchvision for a mobile image classification app but it is absolutely massive for what it does.
The apk itself is 90mb, but the download size is 30mb. What’s causing this massive disparity in size? Is there any way for me to reduce this?
For pytorch-android 1.5.0 size of the aar is 25mb,
https://bintray.com/pytorch/maven/org.pytorch%3Apytorch_android/1.5.0#files/org%2Fpytorch%2Fpytorch_android%2F1.5.0
Nightly builds are a bit bigger.
What is the size of your vision model that you pack in application?
(As an idea to have smaller model you may try to use quantized model or models with smaller size like mobileNetV2 or download model on the device at runtime)