Skip to content

Instantly share code, notes, and snippets.

@zeryx

zeryx/model_options.md

Created January 14, 2020 22:52
Show Gist options
  • Star 0 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
  • Save zeryx/3397b9988953e2dd3a7c35510d9ec4f2 to your computer and use it in GitHub Desktop.
Save zeryx/3397b9988953e2dd3a7c35510d9ec4f2 to your computer and use it in GitHub Desktop.
Download ZIP
Raw
model_options.md

Already done

  • ResNet 18 - Image Classification - ImageNet (not available for tensorflow)
  • ResNet 50 - Image Classification - ImageNet

Suggestions

  • MobileNet v2 - Image Classification
    • Why
      • Small architecture, and available on all platforms (including tensorflow)
      • Common vanilla model for edge devices (smart phones, etc)
    • Why Not
      • Not too different from ResNet 18
  • VGG 16 - Image Classification
    • Why
      • Classic deep learning archiecture
      • Available in all model zoos
      • Fewer layers than ResNet 18 or 50, but significant more parameters
    • Why Not
      • Not dramatically different from classic image processing
      • Doesn't expose us to more varied architectures
  • DenseNet 121 - Image Classification
    • Why
      • Modern deep learning architecture
      • Smaller footprint than ResNet 50
      • Available in all model zoos
    • Why Not
      • Not dramatically different from classic image processing
  • BERT - Machine Comprehension
    • Why
      • Interesting new architecture, lots of customer asks
      • Different architecture, uses attention
      • GPU utilization is wholy different from CNNs
    • Why Not
      • Not available in the vanilla model zoos in pytorch/tensorflow
      • May be somewhat tricky to make an apples/apples comparison as architecture requires custom layers
  • Faster RCNN - Image Object Detection
    • Why
      • Common image problem type
      • Common model architecture
      • Available on all model zoo's
    • Why Not
      • Not dramatically different from image classification in terms of compute flow
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment