edvardHua

Below is the SOCs (supported GPU) I have tested. Model | Snapdragon 845 | Snapdragon 660 | Hisilicon 960 | Exynos 7420 ---- | --- | --- | --- |...

You can use the coco dataset to training it, just need minor modification of code. But the training code i provide is a little bit mess, i will update it...

@jgcbrouns @tucan9389 Network architecture optimize and make model run faster on mobile device. Check [mv2_hrnet](https://github.com/edvardHua/PoseEstimationForMobile/blob/enhance/training/pefm/networks/mv2_hrnet.py) for newly added network, i will update the code gradually at this branch.

I clone my project and rebuild it again. Below is the environment of mine: - Operation System: `macOS 10.13.6` (mace not support build under windows now) - Android Studio: `3.0.1`...

@jestyf @MahajanTarun @Kanbo0409 Pls check the correct version of ndk and build it under unix-like operation system. One of my friend try to build and encounter some errors and fixed...

@jestyf @MahajanTarun @Kanbo0409 **Different environments may encounter different error when you build mace-demo ( I can't cover all situations). To avoid this, i suggest using docker.** > Pls clone the...

@MoonBunnyZZZ Oh, i mean the model is almost converge after 12h according to the loss value. I will correct it in `README.md`. Well, the plot on `README.md` show 250,000 step...

In GTX 1080Ti, the "volatile GPU-util" is 0%~80%. If you model already fixed, you can delete the validate code in training to reduce circumstance of 0% "volatile GPU-util".

It seem the data type mismatch. You need to debug it by yourself.

Pls check rep for using GPU on android. Now it can reach 50+ FPS on Snapdragon 845. 😁