Repository for Image Recognition Challenges
This implements training & test results of the most popular image classifying challenges, including cifar-10, cifar-100, imagenet(ILSVRC-2012).
| GPUs | numbers | nvidia-version | dev | memory |
|---|---|---|---|---|
| GTX TitanX | 1 | 367.57 | local | 12G |
| GTX TitanX | 4 | 375.20 | server | 12G |
There are two versions of implementations in this repository.
| Dataset | network | dropout | Top1 acc(%) |
|---|---|---|---|
| CIFAR-10 | wide-resnet 40x10 | 0.3 | 96.35 |
| CIFAR-100 | wide-resnet 34x10 | 0.3 | 81.88 |
| ILSVRC-2012 | will be updated | - | - |
| Cat vs Dog | wide-resnet 34x2 | 0.3 | 98.75 |
CIFAR-10 dataset is consisted with 50,000 training images and 10,000 testing images. Each image is consisted in an RGB format with the size of 32 x 32 pixels.
There are 10 corresponding labels to each image. The labels are descripted below.
CIFAR-100 dataset is consisted with 50,000 training images and 10,000 testing images. Each image is consisted in an RGB format with the size of 32 x 32 pixels.
There are 100 corresponding labels to each image. The labels are descripted below.
ILSVRC(Imagenet Large Scale Visual Recognition Challenge)
Cat vs Dog Challenge dataset is consisted with 25,000 training images of cats and dogs with various pixels. We will automatically split this to make a "skewed dataset".
Unlike the conventional challenge, the optimal goal to this repository is to maximize the accuracy whilst the skewness of the training data set.
We will consist the training data with 1,000 cats and 10,000 dogs, while consisting the test set with 2,500 cats and 2,500 dogs.
