Tensorflow implementation of our paper at ICCV 2017, which proposes a supervised approach to learn highly compressed image representations.
It requires a path_file with paths of all training images along with their label ids.
To train from a pre-trained base-CNN and fine-tuning only the new layers, as we do in the paper, run:
ipython train_subic --m 8 --k 256 --nclass 1000 --img_path path_file --pretrained trained_model --skip_last 2 --finetune 2
Or for a full training run:
ipython train_subic --m 8 --k 256 --nclass num_classes --img_path path_file
Check the arguments in train_subic.py to try different parameters and settings.
The path_file should look like,
path/to/imagenet/images/000001.jpg 0
path/to/imagenet/images/000002.jpg 4
Download a SuBiC model trained on ImageNet will be available soon (links will be set on this page). It has 8 layers, the first 7 layers are of VGG_M_128 with weights from this caffe model.
For a quick test we provide VGG_M_128 features for pascalvoc images with labels, download from [link available soon]. Then, add the paths of the downloaded files in test_subic.py and run:
ipython test_subic trained_model VGG_M_128 --m 8 --k 256 --testset pascalvoc_features
To test with settings we used in the paper, you need to put the path file(s) for images and labels in test_subic.py corresponding to their datasets.
Otherwise, directly run as:
ipython test_subic trained_model VGG_M_128 --m 8 --k 256 --testset any_dataset --nclass num_classes --split split_size --N dataset_size --img_path path_file
by downloading this program, you commit to comply with the license as stated in the LICENSE.md file.
author: Jain Himalaya email: see ICCV 2017 paper for email adress.