Repository navigation
ncnn #6581
|
https://github.com/Tencent/ncnn/tree/master |
Replies: 5 comments 26 replies
You can set PIP_FIND_LINKS="path/to/folder/with/wheels" flet build apk |
|
Hi all, thanks for the great advice shared in this thread so far! I wanted to jump in and add a few technical points based on some similar challenges I've encountered recently. The PIP_FIND_LINKS approach suggested by @ndonkoHenri is the standard way to include those wheels, but be careful: since you are dealing with ncnn (which is C++), a simple wheel might not be enough if it isn't ABI-compatible with the specific NDK version Flet uses for Android builds. I had to implement AES from scratch in Python to get round-by-round visualization, which really forced me to understand how Flet handles native execution boundaries versus pure Python logic.
|
|
Nice project - thanks for sharing.
I’m noticing growing demand for ML/CV/OCR libraries mobile support. My plan, when I have some time, is to list the libraries we could support, build wheels for them, and put together usage examples to showcase what’s possible. If you want, we could create a Discord thread to discuss this further and brainstorm potential libraries to prioritize, usage examples to build, or any other stuff around this topic. |
|
ncnn 1.0.20260526 wheels have been published for Android and iOS. The whole engine ships statically inside a ~2MB wheel — no extra native libraries to worry about. This opens the path to run real computer-vision models (YOLO detection, classification, segmentation) inside your Flet app, fully offline, on both Android and iOS. InstallJust add it to your app's dependencies — [project]
dependencies = ["flet", "ncnn"]The workflow: export on desktop, run on deviceYou cannot run 1. On your desktop, once: pip install ultralytics
yolo export model=yolov8n.pt format=ncnn imgsz=640
# → yolov8n_ncnn_model/ containing model.ncnn.param + model.ncnn.bin(any PyTorch model works via pnnx; classic ONNX models via onnx2ncnn) 2. Ship the 3. On device: import cv2, ncnn, numpy as np
net = ncnn.Net()
net.opt.num_threads = 4
net.load_param(param_path) # text graph
net.load_model(bin_path) # weights
img = cv2.imread(image_path)
mat = ncnn.Mat.from_pixels_resize(
img, ncnn.Mat.PixelType.PIXEL_BGR2RGB,
img.shape[1], img.shape[0], 640, 640)
mat.substract_mean_normalize([], [1 / 255.0] * 3) # YOLO's /255
ex = net.create_extractor()
ex.input("in0", mat) # blob names come from the .param file
ret, out = ex.extract("out0")
pred = np.array(out) # → NMS / postprocessingFor YOLO postprocessing (boxes/NMS/annotation), don't hand-roll it: ncnn's own repo ships a complete yolov8 Python example, and the pure-python supervision package installs on mobile today and gives you ergonomic Run inference in If you have difficulties setting up a simple/running project, let me know, and i will provide you with a working sample you can run/try directly. Tips & gotchas
If you hit a model that exports but misbehaves on the mobile device, open an issue with the |
|
That's amazing! Your expertise in this field is truly outstanding. I'll try to get the whl file you built up and running. I'll reach out to you promptly if any issues arise. Thank you so much! |
|
I have an error when packaging an apk on Linux for Android. the dependenies is like this: dependencies = [ I also tried "ncnn>=1.0.20260526" or "ncnn>=1.0.0" . didn't work. |
|
Options to solve the issue on your end:
[tool.flet.android]
target_arch = ["arm64-v8a", "armeabi-v7a", "x86_64"]
# or in CLI
flet build apk --arch arm64-v8a armeabi-v7a x86_64 |
|
I just tried. Here is the error. ( I also tried not to include the ">=" in the dependencies. I have the same error. ) This is my pyproject: dependencies = [ [dependency-groups] [tool.pytest.ini_options] [tool.flet.app] [tool.flet.android] [tool.flet.android.permission] |
There are still issues. I suggest you build and package the project yourself to check. Another problem is that the packaging process iterates through all versions of certain packages, which takes a lot of time. |
|
I forked your repo and built on CI, and it works: https://github.com/ndonkoHenri/flet-ncnn-test/actions/runs/29093968876 Only thing I added is the CI workflow and the pyproject |
|
Now. Packaging apk works. I installed in an actual phone. I have this when launching it. Do you know what it is? Traceback (most recent call last): |
|
Try adding this to your [tool.flet.compile]
packages = false |
|
@yunguangli, concerning flet-camera, you need to pin it to the same version as flet itself. For example: dependencies = [
"flet>=0.86.0.dev2",
"flet-camera>=0.86.0.dev2",
]No need for dev_packages. |
|
Check this: #6677 (comment) |
|
I upgraded to flet 0.86.2 . everything works. thanks! |
How many frames per second can it run on your device? What are the hardware specifications? |
|
You can find some notes here: #6581 (comment) Give it a try and let us know if you have more findings to note. |
ncnn 1.0.20260526 wheels have been published for Android and iOS. The whole engine ships statically inside a ~2MB wheel — no extra native libraries to worry about.
This opens the path to run real computer-vision models (YOLO detection, classification, segmentation) inside your Flet app, fully offline, on both Android and iOS.
Install
Just add it to your app's dependencies —
flet buildresolves mobile wheels automatically:The workflow: export on desktop, run on device
You cannot run
ultralytics/torchon the phone (ultralyticsimports torch at module level, andtorchhas no mobile wheels yet). The good news: you don't need to. The split that the wh…