jmitnik / facialdebiasing Goto Github PK

Todo's for what we need for the minimal (flexible) API

• Docstrings
• Controlling main-structure from one config
• Clearly defined main functions (trainer / evaluator / plotter?): higher-level and lower-level functions so that one can train their own methods
• Adjustable layers
• Variable readable datasets (different sources for example)
• Alpha dependent on number images

Edit this message to include zeikpoints:

Things not mentioned:

• Image cropping /preprocessing
• C1-C3
• Latent space size
• Required fix for disappearing/exploding W probabilities
  • They took their own way which is weird
• Amount of bins
  • Amount of samples for creating these bins?
• use of mu or distribution for choosing final probability
• Didn't mention huge difference in quality PPB
• Amount of windows not mentioned on test set,
  • Also a big weakpoint of their research, really poor measure to use any() in this

❗ Important: 5 pages

Sections

• Introduction - Pieter
• Background - Pieter
• Methodology: Network, Loss, Debiasing - ?
• Experiments
  • Dataset - Jonathan
• Results - ?
• Discussion & Conclusion - Jonathan

Crop to square of 64 x 64

• README.MD write

• Paper

  • ❗️ Ensure debiasing is used as 'debiasing' everywhere
  • Conclusion
  • Abstract
  • Simon
    • use \left( and \right) and \log in Latex for nice formatting of the formulas (as a Mathematician this makes me happy ;) )
    • "A model is less biased if the difference between the recall of the four subsets is lower." isn't it measured in terms of variance of the recall across the PBB classes (skin color, etc.) ?
    • "This proof indeed suggests a lower variance in accuracy scores for a VAE with the debiased sampler, when compared to a standard VAE." you mean here when compared to standard non-debiased classifier, right?
    • Explicitly state what is recall and precision in the context of face recognition, that is number of correct classified faces divided by etc. The thing is that precision and recall are actually from the field of Information Retrieval and people get often confused what is what :)
    • add a footnote for the unofficial code repo the first time you mention it in section 3
    • "From this formula \hat Q_i (z |X) is created" This feels like it drops out of nothing. How is it created and what are the i's representing?
  • Workload description in Appendix
  • Scan all TODO's for comments
  • Go through the paper, line-by-line
  • Discussion
    • Add detail about variance of amount of models
  • Appendix
    • Sample pictures of datasets ((reference to this somewhere))
    • Sub-images (reference to this somewhere)

• Jupyter: The main demonstration of how this works

  • Add Dante's plots in Jupyter
  • Best and worst labels in title (Plots)

• Code API: Test and install on a different computer from scratch

  • Add docstrings
  • Custom layers
  • add License
  • Remove all files
  • Repo name
  • Change direcotry

• Presentation

  • Introduction + Original paper
  • Our piece

So there are a couple of things we could try to prevent nans from happening on the long run:

There are some useful infomation about why nan problem could happen:
1.the learning rate
2.sqrt(0)
3.ReLU->LeakyReLU

or gradient clipping

Here questions about the Paper are placed and answered.
Label every question so there won't be any confusion.