Arthur Pignet
Arthur Pignet
New contributivity measueament based on statistical distances between 2 distributions: - The partner-specific probability distribution of the label, wrt to the input; (estimated via maximun likelihood wrt to the partner's...
The current initialization of s-models is done with a matrix which shape is hardcoded to be 10-by-10. This shape should be computed from the model
With the current implementations of the various mpls, the computational time is not at all related to a true federated scenario, for two main reasons: - the partners cannot train...
The MNIST nn have 1,199,882 parameters, for an accuracy of 0.9893 +/- 0.01 (keras fit, epochs=12, batch_size=168) To compare with, the cifar model have 1,250,858 parameters, and reach 0.7888 +/-...
See above: One example, sometime there is #, sometime there isn't. 2021-01-14 17:54:30 | INFO | ### Splitting data among partners: 2021-01-14 17:54:30 | INFO | Train data split: 2021-01-14...
Currently we can set the validation and test sets as local or global. It could be interesting to had both usages
When using the ESC50 dataset, we preprocess the data at each instanciation of a dataset. However the preprocessing is quite long. It would be great to store the preprocessed inputs...
Currently most of the functions work with the assumptions that the labels are one-hot-encoded vectors. Besides the fact that it is not responsive, sometime we need to play with label...
Today, to simulate a corrupted partner we can set all the labels of its dataset to 1 or shuffle its labels. It could be really interesting if the whole dataset...