def report_val(self, loss):
logging.info("[val] [it: {}] [{}: {}] [previous best {}: {}]".format(
self.it, self.criterion.evaluate_loss_name(), loss,
self.criterion.evaluate_loss_name(), self.best_val_loss))
writer.add_scalar("val_{}".format(self.criterion.evaluate_loss_name()),
loss,
global_step=self.it)
def report_test(self, loss):
logging.info("[test] [it: {}] [{}: {}]".format(
self.it, self.criterion.evaluate_loss_name(), loss))
writer.add_scalar("test_{}".format(
self.criterion.evaluate_loss_name()),
loss,
global_step=self.it)
def log_likelihood(self, it, **kwargs):
ais = config.get("others", "ais", default=False)
if ais:
self.model.update_log_partition()
ll = evaluate(self.te, self.runner, lambda v: self.model.log_likelihood(v, ais=ais),
batch_size=min(10000, len(self.te)))
logging.info("[log_likelihood] [it: {}] [log_likelihood: {}]".format(it, ll))
writer.add_scalar("log_likelihood", ll, global_step=it)
def report_train(self, loss, **kwargs):
logging.info("[train] [it: {}] [{}: {}]".format(
self.it, self.criterion.train_loss_name(), loss))
writer.add_scalar("train_{}".format(self.criterion.train_loss_name()),
loss,
global_step=self.it)
for k, v in kwargs.items():
writer.add_scalar("{}".format(k), v, global_step=self.it)
def classify(self, it, **kwargs):
tr_features, tr_ys = extract_labelled_feature(self.q, self.labelled_tr)
te_features, te_ys = extract_labelled_feature(self.q, self.labelled_te)
classifiers = kwargs.get("classifiers", "lsvm")
if not isinstance(classifiers, list):
assert isinstance(classifiers, str)
if classifiers == "all":
classifiers = ["kn", "svm", "lsvm", "logistic"]
else:
classifiers = [classifiers]
for classifier in classifiers:
acc = classify_features(tr_features, tr_ys, te_features, te_ys, classifier)[-1]
logging.info("[{} classify] [it: {}] [acc: {}]".format(classifier, it, acc))
writer.add_scalar("{}_classify".format(classifier), acc, global_step=it)
def fisher(self, it, **kwargs):
fisher_ = evaluate(self.te, self.runner, lambda v: criterions.ssm(v, self.model),
batch_size=min(10000, len(self.te)))
logging.info("[fisher] [it: {}] [fisher: {}]".format(it, fisher_))
writer.add_scalar("fisher", fisher_, global_step=it)