This repository is designed to do performance analyses of classifiers against the objectnet image set, using Tensorflow.
Setting this up is a pain in the ass, long-winded, and prone to breaking. Don't do it unless run-time is getting prohibitive, and you'll probably need a £500+ graphics card to speed things up considerably.
The repo does not contain the models, you must add these yourself (and do not add them to the repo, they're too big). All models you want to test are downloaded to downloads. Here's an example for ResNet50:
cd downloads
wget https://storage.googleapis.com/tensorflow/keras-applications/resnet/resnet50_weights_tf_dim_ordering_tf_kernels.h5
cp resnet50_weights_tf_dim_ordering_tf_kernels.h5 ../model
Only one model should be in the models directory at any one time.
The repo does not contain images, you must add these yourself (and do not add them to the repo, they're too big). On the willslab-ply server, the easiest solution to gain full access to objectnet is to make a symbolic link to the dataset drive:
ln -s /mnt/datasets/objectnet/images/ input/images
On my home desktop
ln -s ~/objectnet/images input/images