Authors: Chen Xing, Sercan O. Arik, Zizhao Zhang, Tomas Pfister
Link: https://arxiv.org/abs/1912.01730
We propose distance-based learning from errors (DBLE) to improve confidence calibration of DNNs. DBLE bases its confidence estimation on distances in the representation space. We first adapt prototypical learning for training of a classification model for DBLE. It yields a representation space where the distance from a test sample to its ground-truth class center can calibrate the model performance. At inference, however, these distances are not available due to the lack of ground-truth labels. To circumvent this by approximately inferring the distance for every test sample, we propose to train a confidence model jointly with the classification model by merely learning from mis-classified training samples, which we show to be highly beneficial for effective learning. On multiple datasets and architectures, we demonstrate that DBLE outperforms alternative single-modal confidence calibration approaches.
DBLE is model agnostic and can be applied to classification task on any datasets. In this directory, we provide example implemetations for 3 widely-used neural network architectures:
- MLP
- VGG
- ResNet
Detailed model architectures can be specified by modifying hyper-parameters of every model architecture. For example, in dble.py, we initialize ResNet-50 for CIFAR dataset as,
resnet_model = resnet.Model( wd=flags.weight_decay,
resnet_size=50, bottleneck=True, num_classes=flags.num_classes_train,
num_filters=16, kernel_size=3, conv_stride=1, first_pool_size=None,
first_pool_stride=None, block_sizes=[8, 8, 8], block_strides=[1, 2, 2],
data_format='channels_last')
If other model architectures except MLP, VGG or ResNet are required, the class implementation of the architectures should be put into './new_arch.py'. 'new_arch' can be replaced by the name of the new architecture.
In class new_arch, two functions encoder(self, inputs, training) and confidence_model(self, mu, training) should be implemented.
encoder(self, inputs, training) takes the raw samples as input and output their representations. The encoder architecture should have the appropriate inductive bias for the input data type, e.g. ResNet for images.
confidence_model(self, mu, training) takes the sample representations as input and output their variances for confidence calibration. The confidence_model architecture can be a simple low capacity model, such as few MLP layers.
DBLE can be applied to classification tasks for confidence calibration on any datasets. In this directory, we demonstrate DBLE on 4 image classification data-sets,
- MNIST
- CIFAR-10
- CIFAR-100
- Tiny-ImageNet
The pre-processing of data-sets is in 'data_loader.py'. For example, the loading and pre-processing of CIFAR-10 is,
def _load_cifar10():
(x_train, y_train), (x_test, y_test) = cifar10.load_data() x_train =
x_train.astype('float32') x_test = x_test.astype('float32') fields = x_train,
np.squeeze(y_train) fields_test = x_test, np.squeeze(y_test)
return fields, fields_test ` ` def augment_cifar(batch_data, is_training=False):
image = batch_data if is_training: image =
tf.image.resize_image_with_crop_or_pad(batch_data, 32 + 8, 32 + 8) i =
image.get_shape().as_list()[0] image = tf.random_crop(image, [i, 32, 32, 3])
image = tf.image.random_flip_left_right(image) image =
tf.image.per_image_standardization(image)
return image
If other data-sets are required to test DBLE, 2 functions _load_new_data() and augment_new_data(batch_data, is_training=False) should be implemented. 'new_data' can be replaced by the name of the new data-set.
_load_new_data() reads the raw samples and labels of the training and the test set and returns two tuples of arrays, fields and fields_test. Each tuple contains the array of samples and the corresponding labels. The first dimensions of the two arrays should be equal.
augment_cifar(batch_data, is_training=False) takes the tensor of a batch of samples as input and output the processed batch of samples.
All hyper-parameters required for the training and evaluation of DBLE are listed in 'main_dble.py'. Here we list some main hyper-parameters that should be adjusted with respect to the data sets and model architectures used.
parser.add_argument('--model_name', type=str, default='vgg')
parser.add_argument('--dataset', type=str, default='cifar10')
parser.add_argument( '--num_cases_train', type=int, default=50000)
parser.add_argument( '--num_cases_test', type=int, default=10000)
parser.add_argument( '--num_samples_per_class', type=int, default=5000, help='')
parser.add_argument( '--num_classes_total', type=int, default=10, help='Number
of classes in total of the data set.' ) parser.add_argument(
'--num_classes_test', type=int, default=10, help='Number of classes in the test
phase. ') parser.add_argument( '--num_classes_train', type=int, default=10,
help='Number of classes in a protoypical episode.' ) parser.add_argument(
'--num_shots_train', type=int, default=10, help='Number of shots (supports) in a
prototypical episode.') parser.add_argument( '--train_batch_size', type=int,
default=100) parser.add_argument('--weight_decay', type=float, default=0.0005)
Here is the command example for the experiment that classifies CIFAR-10 with VGG-11.
python3 -m main_dble
--num_classes_train 10 --num_classes_test 10 --num_shots_train 10
--train_batch_size 100 --number_of_steps 80000 --num_tasks_per_batch 2
--num_classes_total 10 --num_samples_per_class 5000 --dataset cifar10
--model_name vgg --num_cases_train 50000 --num_cases_test 10000 --weight_decay
0.0005 --log_dir ./vgg11_cifar10/