Unofficial tensorflow.keras implementation of Lite-HRNet (Lite-HRNet: A Lightweight High-Resolution Network).
Lite-HRNet has an official implementation: official repo with mmpose configs. Lite-HRnet has also been merged into mmpose.
This implementation is based on the official one, you can find commented links to respective parts here and there.
The only dependency of this project is tensorflow. You can clone the repo and install the package locally:
git clone https://github.com/zimka/lite_hrnet_tfk
pip install lite_hrnet_tfk/
Or you can install it from github directly:
pip install git+https://github.com/zimka/lite_hrnet_tfk.git
Because of ambiguity of tensorflow nor tensorflow-gpu packages, none is installed by default.
You can either install appropriate tensorflow in advance, or specify it as extra dependency for pip during the installation:
pip install lite_hrnet_tfk[tensorflow-gpu]
If you want to run tests please install the package locally, then run pytest:
pytest lite_hrnet_tfk/tests
The easiest way to build a network is to use prepared configs.
from lite_hrnet_tfk.config import LiteHrnetConfig # config describes how the net should be built
from lite_hrnet_tfk.net import LiteHrnet # net uses config to compose separate modules into the network
config = LiteHrnetConfig.lite18(out_channels=42) # create prepared config with as many channels as necessary
print(config) # check config parameters, change whatever you want (or use other LiteHrnetConfig classmethod)
net = LiteHrnet(config)You can also build net from separate modules, check lite_hrnet_tfk.modules code for details.