IceVision

---

Quick Fact :

• https://airctic.com/0.12.0/install/
• It wraps up fastai, pytorch lightning

• Its unified API allow to choose libraries → their models → their backbones

  # e.g. lib = mmdet, model = retinanet, backbone = resnet50_fpn_1x
  model_type = models.mmdet.retinanet
  backbone = model_type.backbones.resnet50_fpn_1x(pretrained=True)
  

Prerequisite Knowledge :

Basic knowledge of

• Fastai / Pytorch Lightning
• YOLO

---

Frequent Issue :

• Issue - Icevision is suck on dependencies, the functional env so far :

  # packages
  python==3.7.13
  icevision[all]==0.11.0
  yolov5-icevision==6.0.0
  "opencv-python-headless<4.3"
  setuptools==61.2.0
  

• issue - how to do inference with the exported model ?

  • i.e. how to load trained checkpoints

  • option 1 - official inference method model_from_checkpoint

    • weird, need to install mmcv-full even if you use yolo5:

      pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cpu/torch1.10.0/index.html
      #mmcv-full-1.5.3.tar.gz
      

  • option 2 - use fastai / torch to load them manually

• issue - training or data loader hanged locally (CPU)

  • because CPU env lead to deadlock
  • https://github.com/pytorch/pytorch/issues/1355
  • You need to set num of workers to 0

• issue - images / plot not showing

  • you need to put %matplotlib inline or others in the top of notebook
  • https://github.com/matplotlib/matplotlib/issues/14534

• issue - about “image truncated”, below code can solve it

  from PIL import ImageFile
  ImageFile.LOAD_TRUNCATED_IMAGES = True
  

Comments