Created by: shippingwang

CI should detect/test useful statistics, such as

• models can RUN or not,
  • random choosing 10 architectures
  • run fake dataset, only need 8 or 16 samples to save time.
  • support some post-processing
    • Actually, some models cannot run eval.py, inference script, etc. by indicating the stored (params/models) directly! one way to solve this problem is to add post-processing in code, it will insert more 'specified model' codes in scripts. Integrate all processing to solve CI issues will confuse the users, the other is to skip CI. So it produces a new demand ->
  • Add a white list to skip CI, There are some historical problems cannot solve together.
• models do RUN the correct results
  • compare the layers result to make sure finally result is stable.