42 lines
2.3 KiB
Markdown
42 lines
2.3 KiB
Markdown
|
# Übergabe Hailo
|
||
|
|
|
||
|
|
* Raphael Maehnle
|
||
|
|
* Thomas Hamböck
|
||
|
|
|
||
|
|
## Agenda
|
||
|
|
|
||
|
|
1. CVAT: Login, Data Import, Labeling, Data export
|
||
|
|
1. Create a new Project; Important: Labels eintragen, möglichst in konsistenter Reihenfolge (zu anderen Projekten)
|
||
|
|
1. Früher gab es nur Tasks => da war Labels konsistent halten aufwändiger; Anm.: Projects funktionieren zZ noch nicht stabil
|
||
|
|
2. Anm.: Früher gab es ein Speicherproblem; Bilder blieben im RAM => häufigeres Seite neu laden sinnvoll
|
||
|
|
2. Jobs: Aufteilung von Bildern auf annotierende Leute (User Assignment)
|
||
|
|
3. Bilder annotieren: Tastenkürzel ("n" > neue Bounding Box, ...): https://openvinotoolkit.github.io/cvat/docs/manual/advanced/shortcuts/
|
||
|
|
4. "..." > export as dataset > Yolo1.1
|
||
|
|
2. hailo setup for development (training)
|
||
|
|
1. prerequisite: nvidia-docker
|
||
|
|
2. clone hailo model zoo: https://github.com/hailo-ai/hailo_model_zoo
|
||
|
|
3. download & run docker container "Hailo Software Suite (for Ubuntu 20.04)"
|
||
|
|
1. note: volume mount is missing in the setup guide => will be required for importing data
|
||
|
|
3. Retraining Yolov5m
|
||
|
|
1. Tutorial for yolov5 training
|
||
|
|
https://github.com/hailo-ai/hailo_model_zoo/blob/master/docs/RETRAIN_ON_CUSTOM_DATASET.md
|
||
|
|
2. dataset meta setup:
|
||
|
|
http://192.168.0.63:8930/raphael.maenle/hailo_inference/-/blob/main/dataset/dataset/dataset.yaml
|
||
|
|
labels aus CVAT nochmal in die liste übertragen, nc == len(names)
|
||
|
|
3. from readme: `docker run -it --gpus all -ipc=host -v /path/to/dataset/:/dataset yolov5:v0`
|
||
|
|
based on https://github.com/hailo-ai/hailo_model_zoo/blob/master/training/yolov5/Dockerfile
|
||
|
|
4. https://github.com/hailo-ai/hailo_model_zoo/blob/master/docs/RETRAIN_ON_CUSTOM_DATASET.md?plain=1#L183
|
||
|
|
epochs 3 ist etwas wenig; ~100 wäre besser
|
||
|
|
`--data` => auf dataset
|
||
|
|
5. Export to ONNX: https://github.com/hailo-ai/hailo_model_zoo/blob/master/docs/RETRAIN_ON_CUSTOM_DATASET.md?plain=1#L192
|
||
|
|
For visualisation: ONNX viewer, e.g.: https://github.com/lutzroeder/Netron
|
||
|
|
4. Yolov5 hailo quantisation + compile
|
||
|
|
1. `hailo_sw_suite_docker_run.sh` => add `-v` as `DOCKER_ARGS` for volume mount in script
|
||
|
|
2. network config: `yolov5m.yaml`
|
||
|
|
3. do `compile` => includes `quantize`
|
||
|
|
`.hef` => network bitstream
|
||
|
|
5. hailo setup for deployment (inference)
|
||
|
|
1. `HailoRT` => setup für runtime
|
||
|
|
6. Inference
|
||
|
|
|