ravi03071991 / nips-global-paper-implementation-challenge Goto Github PK

This is implementation of NIPS 2017 Paper Selective Classification For Deep Neural Networks as a part of NIPS Global Paper Implementation Challenge

1. Python 3.6
2. Keras
3. Tensoflow

• ImageNet validation dataset can be downloaded from here.
• ILSVRC2012_validation_ground_truth.txt contains ground truth labels for ImageNet validation dataset.
• imagenet-classes-dict.dat is a pickle dictionary, if you input a class you get a number from 1 to 1000 corresponding to the ground truth in the ILSVRC2012_validation_ground_truth.txt file.
• The weights of the model trained as suggested in paper on CIFAR-10 and CIFAR-100 datasets can be downloaded from CIFAR-10 WEIGHTS(93.67% accuracy)and CIFAR-100 WEIGHTS(70.52% accuracy).

• Evaluating on CIFAR-10 dataset.

python eval/cifar10_vgg16.py

• Evaluating on CIFAR-100 dataset.

python eval/cifar100_vgg16.py

• Evaluating on ImageNet validation dataset using VGG16 top1.

python eval/vgg16_top1.py

• Evaluating on ImageNet validation dataset using VGG16 top5.

python eval/vgg16_top5.py

• Evaluating on ImageNet validation dataset using ResNet50 top1.

python eval/resnet50_top1.py

• Evaluating on ImageNet validation dataset using ResNet50 top5.

python eval/resnet50_top5.py

• On CIFAR-10 using VGG16.

• On CIFAR-100 dataset using VGG16.

• On ImageNet Validation dataset using VGG16 Top1.

• On ImageNet Validation dataset using VGG16 Top5.

• On ImageNet Validation dataset using ResNet50 Top1.

• On ImageNet Validation dataset using ResNet50 Top5.

• Achieved 60% test coverage guaranteed with 99.9% probability at 3% error rate top-5 ImageNet classification.