Giters

dingmyu / HR-NAS

HR-NAS: Searching Efficient High-Resolution Neural Architectures with Lightweight Transformers (CVPR21 Oral)

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

pretrained models

seekingdeep opened this issue · comments

commented

@dingmyu hi,
were the pretrained models uploaded such as for the imagenet and cityscapes.
can it run inferencing using cpu-only
can hr-nas detect small objects

commented

Hi,

  • The pre-trained model is currently unavailable. Please stay tuned.
  • The searched model can run in CPU-only mode, but its efficiency (which cannot run in real-time) will be much slower than GPU.
  • Sample images can be found in the paper. We have not evaluated the performance of detecting small objects because HR-NAS aims to search for efficient models.

Thanks,
Mingyu

commented

last question here:

  • have you considered pruning the final efficient models, inorder to reduce the flops and parameter?
    i can send you some links to interesting spare pruning methods if you want.
  • have you considered exporting to onnx so it can run on arm devices such as raspberry pi?
commented

last question here:

  • have you considered pruning the final efficient models, inorder to reduce the flops and parameter?
    i can send you some links to interesting spare pruning methods if you want.
  • have you considered exporting to onnx so it can run on arm devices such as raspberry pi?

Hi,

  • The search algorithm of HR-NAS is based on pruning. And we tried to manually reduce some channels of the searched model, the accuracy drops a little.
  • We have not tried the model on chips or arms and do not have the plan of exporting to onnx, thanks. PR is welcomed.

Thanks.

commented

thanks