def _fit_(self, X, Y, epochs, n_instances, tf_session, verbose):
try:
for epoch in range(epochs):
for start in range(0, n_instances, self.batch_size):
end = np.min([start + self.batch_size, n_instances])
tf_session.run(
self.optimizer,
feed_dict={
self.X: X[start:end],
self.Y: Y[start:end]
},
)
if verbose == 1:
progress_bar(end, n_instances, status="Fitting")
if verbose == 1:
c = tf_session.run(self.loss,
feed_dict={
self.X: X,
self.Y: Y
})
print("Epoch {}: cost {} ".format((epoch + 1), np.mean(c)))
if (epoch + 1) % 100 == 0:
c = tf_session.run(self.loss,
feed_dict={
self.X: X,
self.Y: Y
})
self.logger.info("Epoch {}: cost {} ".format((epoch + 1),
np.mean(c)))
self.step_decay(epoch)
except KeyboardInterrupt:
self.logger.info("Interrupted")
c = tf_session.run(self.loss, feed_dict={self.X: X, self.Y: Y})
self.logger.info("Epoch {}: cost {} ".format((epoch + 1),
np.mean(c)))
def _tune_threshold(self, X_val, Y_val, thin_thresholds=1):
scores = self.predict_scores(X_val)
probabilities = np.unique(scores)[::thin_thresholds]
threshold = 0.0
best = f1_measure(Y_val, scores > threshold)
try:
for i, p in enumerate(probabilities):
pred = scores > p
f1 = f1_measure(Y_val, pred)
if f1 > best:
threshold = p
best = f1
progress_bar(i, len(probabilities), status='Tuning threshold')
except KeyboardInterrupt:
self.logger.info("Keyboard interrupted")
self.logger.info('Tuned threshold, obtained {:.2f} which achieved'
' a micro F1-measure of {:.2f}'.format(
threshold, best))
return threshold
