chilang

/*! (C) The Hyperaudio Project. MIT @license: en.wikipedia.org/wiki/MIT_License. */
/*! Version 2.1.4 */
'use strict';

var caption = function () {
  var cap = {};

  function formatSeconds(seconds) {
    if (typeof seconds == 'number') {
      //console.log("seconds = "+seconds);

chilang

ttest, pval = ttest_ind(local_mean_f1_scores, fed_avg_f1_scores)
print("statistically significant:", pval < 0.05)
print("p =", pval)

chilang

hyperparams = {
    'n_runners': 10,
    'sample_size': 5,
    'rounds': 100,
    'combine': 'weighted', # or 'mean'
    'partition_params': {
        'scheme': 'uniform' # stratified, non-iidness, noise injection etc.
    },
    'runner_hyperparams': {
        'epochs': 1,

chilang

skews = {0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.99}
scores = {}
for skew in skews:
    print("eval", skew)
    hyperparams = {
        'n_runners': 10,
        'sample_size': 10,
        'rounds': 1,
        'combine': 'weighted', # or 'mean'
        'partition_params': {

chilang

def compare_f1(hyperparams, rounds=1):
    local_mean_f1_scores = []
    local_max_f1_scores = []
    fed_avg_f1_scores = []
    for i in range(0, rounds):
        fed_avg = FedAvg(**hyperparams)
        fed_avg.fit(X_train, y_train)
        preds = fed_avg.predict(X_test)
        fed_avg_f1_scores.append(f1_score(y_test, preds, average='weighted'))
        tmp = []

chilang

def label_distribution_skew(x, y, partitions, skew=1):
    def runner_split(N_labels, N_runners):
        """number of labels to assign to n runners"""
        runner_labels = round(max(1, N_labels / N_runners))
        runner_split = round(max(1, N_runners / N_labels))
        return runner_labels, runner_split

    runner_data = []
    N_labels = np.unique(y).shape[0]
    n_labels, n_runners = runner_split(N_labels, partitions)

chilang

    for _ in range(self.rounds):
        r_intercepts, r_coefs, r_weights = self._collect_models(runners, N)
        self.intercept_ = np.average(r_intercepts, axis=0, weights=r_weights)
        self.coef_ = np.average(r_coefs, axis=0, weights=r_weights)

    self.global_model = set_weights(self.intercept_, self.coef_, np.unique(y))
    return self

chilang

def _collect_models(self, runners, N):
    r_intercepts, r_coefs, r_weights = [], [], []
    self.models = []
    for runner in random.sample(runners, k=self.sample_size):
        r_model = runner.optimise(self.intercept_, self.coef_, self.runner_hyperparams)
        self.models.append(r_model)
        r_intercepts.append(r_model.intercept_)
        r_coefs.append(r_model.coef_)
        r_weights.append(runner.dataset_size()/N if self.combine == 'weighted' else 1/self.sample_size)
    return r_intercepts, r_coefs, r_weights

chilang

def fit(self, X, y):
    if self.intercept_ is None or self.coef_ is None:
        features = X.shape[1]
        self.intercept_ = np.zeros(1)
        self.coef_ = np.zeros((1, features))

chilang

1. Backup iPad/iPhone to Mac (without encryption)

2. Locate backup folder and run https://code.google.com/p/iphone-dataprotection/:

    python python_scripts/backup_tool.py ~/Library/Application\ Support/MobileSync/Backup/XXXXXXX outdir

3. Open sqlite db under using SquirrelSQL:

    outdir/HomeDomain/Library/Notes/notes.sqlite