Few-Shot-Attribute-and-Multi-Label-Classification-Using-Deep-Neural-Network

Code for Few-Shot Attribute and Multi-Label Classification Using Deep Neural Network implemented in pytorch

Installation

• Install PyTorch by selecting your environment on the website and running the appropriate command.
• Clone this repository.
  • Note: We currently only support Python 3+.
• Install unzip and unrar
• Then download the dataset by following the instructions below.
• Follow the run instructions below for training and testing the model.

Datasets

• Go to the data folder
• Run "sh download_<dataset>.sh" to download and preprocess the datasets
• The prepocessing code for each dataset will be automatically called by the corresponding download_<dataset>.sh file
• Some of the datasets are downloaded directly from the main repository.
• Some of the datasets which need some modifications are uploaded to figshare after modification.
• Following datasets are used:
  • LFW - images https://ndownloader.figshare.com/files/14758619, attributes http://www.cs.columbia.edu/CAVE/databases/pubfig/download/lfw_attributes.txt
  • CelebA - images https://ndownloader.figshare.com/files/14759408, attributes https://ndownloader.figshare.com/files/14759465
  • yeast - https://excellmedia.dl.sourceforge.net/project/mulan/datasets/yeast.rar
  • emotions - https://sci2s.ugr.es/keel/dataset/data/multilabel/emotions.zip
  • birds - http://sourceforge.net/projects/mulan/files/datasets/birds.rar
  • enron - https://sci2s.ugr.es/keel/dataset/data/multilabel/enron.zip
  • medical - https://sci2s.ugr.es/keel/dataset/data/multilabel/medical.zip
  • llog - https://liquidtelecom.dl.sourceforge.net/project/meka/Datasets/LLOG-F.arff
  • slashdot - https://liquidtelecom.dl.sourceforge.net/project/meka/Datasets/SLASHDOT-F.arff
  • scene - https://sci2s.ugr.es/keel/dataset/data/multilabel/scene.zip

Experiments

• Run the <...>_train_test.py file to get the results for our proposed Multi-Label model
• Run the <...>_train_test_baseline.py file to get the results for the baseline
• Provide additional choices:
  • --gpu _ : gpu id on which the program will be run
  • --epochs _ : Number of iterations(default 20) for Phase 1 (train on first half of labels and test on remaining) and Phase 2 (train on second half of labels and test on remaining). Each iteration has 1000 episodes
  • --shot _ : N-shot - Number of support examples per label (default 20).
• Average Test F1 Score will be displayed after the Phase 1 and Phase 2 tests are completed