Supported Features:
- General instance and video classification, detection, tracking, and search
- Supervised and self-supervised support
- Distributed traing
- Face detection: retina face
This is an initial design for a tag system, including training and inference.
- Dataset: Providing dataloader for both image and video
- Backbone: Providing various pretrained model selection for both image and video support
- Tasks: Providing various task trainers and inferences, including detection, classification and search
- Loss: Providing loss function selection and parameter adjuster, supporting both image and video
- Optimizer: Providing optimizer function selection and parameter adjuster, supporting both image and video
- Metrics: Providing evaluation functions for different tasks and models
- Distributions: Providing CPU and GPU support versions, also extending to Persia in the future
python setup.py build
python setup.py install
Or
pip install .
export PYTHONPATH=YOURPATH/tag-framework-system/:$PYTHONPATH
Pretrained model Download Link
Pretrained model
-
P3D-199 trained on Kinetics dataset: Download Link
-
P3D-199 trianed on Kinetics Optical Flow (TVL1): Download Link
-
P3D-199 trained on Kinetics600, RGB, 224&299: Change the value of GAP kernel from 5 to 7 if 224, to 9 if 299 Download Link
-
Pretrained model: Network trained on the set of kuaishou commodity with visual and language infomation Download Link
-
Use the following command to fine tune the model with the latest ebiz data:
cd tag_framework/networks/ebiz_item_classify/
sh at012_auto_train.sh
Currently, we do not have pretrained C3D network.
You can use retina face now, just look at example/retinaface/test_face.py
Copyright (c) 2020, Kuaishou. All rights reserved.