cjnani16 / deeprfs_tf Goto Github PK

get model training working with TF version 1.13

• common data

• general path function

• reproducible virtual environments based on requirements.txt

  • Should be able to pip install -r requirements.txt to fully reproduce virtualenv

https://arxiv.org/pdf/1708.07689.pdf
https://github.com/sebastian-lapuschkin/understanding-age-gender-deep-learning-models

• reproduce figs 2 and 4 on gender classification on a conceptual level - how does removing negative values affect appearance?
• attempt gender classification on our VPNL images, look for general commonalities (e.g. side of jaw highlighting for males)

use the VPNL faces to iteratively look at relevance maps & removing diagnostic pixels.

• at each step, compute average relevance map for both male/female faces so that we're removing the same pixels from each face
• replace pixels that are relevant > a threshold with background greyscale value
• re-run relevance propagation

Following up on Issue #3 :

Motivation

So far, @cjnani16 has only been able to train on 1% of the data due to hardware constraints. One advantage of moving to Nyla is that training on the entire dataset (potentially many times over) should be do-able.

Implementation

• Verify that running training script using 1% of data works as expected on Nyla, i.e., run the exact same notebook that has been running on Amoxicillin to make sure the packages are all installed as expected.
• Write a notebook that trains using the entire dataset and run on Nyla. Save checkpoints to disk (and copy to Amoxicillin if desired).
• Work with @eshedmargalit to convert that notebook into a robust/reproducible python script that can be executed in isolation.