cjnani16 / deeprfs_tf

moving primarily to tensorflow -- Facial Feature Relevance for CNNs
1 stars 0 forks source link

reproduce gender classification paper results #4

Open soniapolt opened 5 years ago

soniapolt commented 5 years ago

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

eshedmargalit commented 5 years ago

Main problem with reproduction of figures is that their values look [-1, 1], ours are [0, 1]. This is likely not a technical plotting issue and instead indicates that relevance is not being computed the same way in the paper and in our notebooks.

cjnani16 commented 4 years ago

Main problem with reproduction of figures is that their values look [-1, 1], ours are [0, 1]. This is likely not a technical plotting issue and instead indicates that relevance is not being computed the same way in the paper and in our notebooks.

For gender it seems like doing a difference could be a solid approach to getting a similar type of data, although it would be better to find confirmation that this is in line with the methods of the paper by finding the relevant part of the methods description within it. I will be focusing primarily on this issue for the near future