Skip to content

Latest commit

 

History

History
54 lines (40 loc) · 2.93 KB

pretrained.md

File metadata and controls

54 lines (40 loc) · 2.93 KB
Raw

Downloading the Pretrained Models

The pretrained models must be located in workspace/experiments/model_ref.

You can download all the models (13 GB uncompressed) at once, or download only the models you need.

If you download the full archive

If you put the compressed file in the <root_folder> and extract with

> cd <root_folder> # so that workspace/experiments/model_ref would be immediately visible.
> tar xzfv model_ref.tar.gz

it will automatically extract the files to <root_folder>/workspace/experiments/model_ref (the archive is made from inside the <root_folder>.)

If you want to download manually

By visiting http://www.cs.ubc.ca/~shafaei/dataset/odtest/ you get a directory listing of all the relevant files. You can navigate and download the models that you like. Make sure that you preserve the directory names when downloading the files.

The easy way to fetch a subset of the pretrained models is to use wget with the recursive option like this example.

> cd <root_folder> # so that workspace/experiments/model_ref would be immediately visible.
> wget -r -nH -np --cut-dirs=3 --reject="index.html*" -e robots=off \
    http://www.cs.ubc.ca/~shafaei/dataset/odtest/workspace/experiments/model_ref/MNIST_VGG.HClass/

You can run this with your own preferred subdirectory. All the files below the address would be downloaded and correctly put in the right directory of the project.

Pre-trained Reference Models

The classification performance of the shared models on the entire D1_train is as follows:

Dataset VGG Resnet
MNIST 99.89% (19 MB) 99.89% (70 MB)
FashionMNIST 98.82% (19 MB) 98.75% (70 MB)
CIFAR10 97.63% (159.8 MB) 97.75% (94.3 MB)
CIFAR100 91.40% (161.3 MB) 92.05% (95.1 MB)
TinyImagenet 69.71% (162.9 MB) 89.59% (95.9 MB)
STL10 93.62% (201.7 MB) 92.32% (94.3 MB)

The KWayLogistic classification performance of the shared models on the entire D1_train is as follows:

Dataset VGG Resnet
MNIST 99.91% (19 MB) 99.91% (70 MB)
FashionMNIST 98.36% (19 MB) 98.73% (70 MB)
CIFAR10 97.34% (159.8 MB) 97.51% (94.3 MB)
CIFAR100 91.87% (161.3 MB) 91.82% (95.1 MB)
TinyImagenet 72.23% (162.9 MB) 65.95% (95.9 MB)
STL10 95.18% (201.7 MB) 93.34% (94.3 MB)