"""
Custom Dataloaders for each of the considered datasets
"""
import os
from torchvision import datasets
from mixmo.augmentations.standard_augmentations import get_default_composed_augmentations
from mixmo.loaders import cifar_dataset, abstract_loader
from mixmo.utils.logger import get_logger
LOGGER = get_logger(__name__, level="DEBUG")
class CIFAR10Loader(abstract_loader.AbstractDataLoader):
"""
Loader for the CIFAR10 dataset that inherits the abstract_loader.AbstractDataLoader dataloading API
and defines the proper augmentations and datasets
"""
def _init_dataaugmentations(self):
(self.augmentations_train,
self.augmentations_test) = get_default_composed_augmentations(
dataset_name="cifar", )
def _init_dataset(self, corruptions=False):
self.train_dataset = cifar_dataset.CustomCIFAR10(
root=self.data_dir,
train=True,
download=True,
transform=self.augmentations_train)
"""
Wrapper functions for metric tracking
Mostly taken from https://github.com/bayesgroup/pytorch-ensembles/blob/master/metrics.py
"""
import numpy as np
from sklearn.metrics import roc_auc_score
from mixmo.utils import visualize
from mixmo.utils.logger import get_logger
from mixmo.core import metrics_ensemble
LOGGER = get_logger(__name__, level="INFO")
def merge_scores(scores_test, scores_val):
"""
Aggregate scores
"""
scores_valtest = {}
for key in scores_test:
key_valtest = "final/" + key.split("/")[1]
if key.startswith("test/"):
keyval = "val/" + key.split("/")[1]
value = 0.5 * (scores_test[key]["value"] + scores_val[keyval]["value"])
if scores_test[key]["string"].endswith("%"):
value_str = f"{value:05.2%}"
else:
value_str = f"{value:.6}"
stats = {"value": value, "string": value_str}
